-------------------------------------------------------------------------------- title: "AI Detector: Free AI Checker for ChatGPT, Claude & Gemini | Pangram" description: "AI detection platform developed by Stanford, Tesla and Google researchers for identifying content from ChatGPT, Claude, Gemini and other AI models" last_updated: "August 2026" source: "https://www.pangram.com/" -------------------------------------------------------------------------------- # Core Product Information Pangram is the AI Detector for ChatGPT, Gemini & Claude — remarkable accuracy. Pangram offers AI detection capabilities with near-zero false positive rates for identifying content generated by ChatGPT, Claude, Gemini, and other AI models. The technology was developed by AI researchers from Stanford, Tesla, and Google. The University of Maryland has reviewed and validated Pangram as the most reliable and accurate AI detection tool in the market. # Product Offerings ## Core Solutions - Dashboard - Chrome Extension - API access - Integration capabilities - Plagiarism detection - Multilingual support - AI Image Detection - Team plans with shared billing and admin controls # Target Markets & Use Cases ## Education Sector - Classroom AI transparency tools - AI content and plagiarism detection for teachers - Academic integrity preservation tools for universities - Integration with Canvas and Google Classroom ## Publishing Industry - AI-generated content identification - Copyright compliance support (following US copyright office ruling on AI content) - Unbiased detection system for non-native speakers ## Trust & Safety - AI content moderation capabilities - Automated workflow support - Cost-effective scaling for trust and safety teams - Performance exceeding human expert detection # Key Partnerships & Integrations - Canvas - Google Classroom - Quora - Tremau - The Transparency Company - NewsGuard # Educational Resources - The State of AI Detection in 2025 - Most common AI phrases - Technical report available on arXiv (2402.14873) # Platform Features - One-click AI detection - Text analysis capabilities - Example text testing options - Data privacy protection - Multiple integration options - Support for various content types (essays, reviews, blog posts) Note: The platform requires cookie collection for website traffic and performance analysis. -------------------------------------------------------------------------------- title: "Pricing | Pangram" description: "Comprehensive AI detection pricing plans for individuals, teams, educational institutions, developers and enterprise customers with varying scan limits and features" last_updated: "August 2026" source: "https://www.pangram.com/pricing" -------------------------------------------------------------------------------- # Pangram Labs Pricing Structure ## Individual Plans ### Free Plan - $0, no payment method needed - Scan up to 2,000 words per day (twenty free checks daily) - AI assistance detection - 3 image detection scans daily - Interpretability features to understand what parts of the text are AI - File upload and OCR for scanned documents - Detection in over 20 languages - Chrome extension and Google Docs integration ### Individual Plan - $20/month, or $15/month billed annually ($180/year) - Scan up to 300,000 words per month - All Free plan features - Plagiarism detection available on every scan - 100 image detection scans monthly - Automatically scan entire feeds on social media, Substack, and more ### Professional Plan - $65/month, or $45/month billed annually ($540/year) - Scan up to 1,500,000 words per month - All Free plan features - Plagiarism detection available on every scan - 500 image detection scans monthly - $200 in monthly API usage included ### Team Plan - $20 per seat/month, or $15 per seat/month billed annually ($180 per seat/year) - Starts at 2 seats; add more anytime - 300,000 words per month per seat - Everything in the Individual plan - Admin controls and member management - Unified billing ## Educational Institutions - Institutional license (contact Pangram for pricing) - Features include: - Unlimited AI checks and plagiarism checks integrated directly into the LMS - Automatic plagiarism detection including copy-paste plagiarism and re-used submissions from past years - Visibility into usage, trends, and statistics on AI use and plagiarism - Full regulatory compliance and data controls; Pangram does not train on student data - Integrations with Canvas, Brightspace, Moodle, Google Classroom, Google Docs, and more ## Developer Credit - Pay-as-you-go prepaid credits: add funds from $5 to $2,000 - API key included - Realtime checks: $0.05 per 100 words - Bulk document scans: 20% discount vs realtime - Auto-refill available - Free access available for academic researchers ## Enterprise Plan - Custom pricing based on volume - Features include: - Progressive volume discounting beyond developer tier - SOC 2 compliance - Enterprise-level security - API access - Dedicated development and support channel Additional Information: - SOC2 TYPE2 verification by AssuranceLab - Integrations available for Canvas and Google Classroom - Academic research support program available - All plans include dashboard features and interpretability tools -------------------------------------------------------------------------------- title: "AI Detector API | Pangram" description: "Technical documentation and access details for Pangram's AI content detection API with Python SDK and REST endpoints" last_updated: "2024" source: "https://www.pangram.com/solutions/api" -------------------------------------------------------------------------------- # Pangram AI Content Detector API ## API Access Methods ### Python SDK - Package name: pangram-sdk - Installation: `pip install pangram-sdk` - Basic usage: ```python from pangram import Pangram pangram_client = Pangram() text = "The quick brown fox jumps over the lazy dog." result = pangram_client.predict_extended(text) ``` - Full documentation available at pangram.readthedocs.io ### REST API - Endpoint: https://text.api.pangramlabs.com - Authentication: Requires x-api-key header - Content-Type: application/json - Sample curl request: ```bash curl 'https://text-extended.api.pangramlabs.com' \ -X POST \ -H 'Content-Type: application/json' \ -H 'x-api-key: my-api-key' \ -d '{ "text": "The quick brown fox jumps over the lazy dog." }' ``` ## API Key Access 1. Create account or login to dashboard 2. Navigate to API section 3. Locate API key in left sidebar ## Pricing Tiers ### Developer Credit - Pay-as-you-go prepaid credits: add funds from $5 to $2,000 - API key included - Realtime checks: $0.05 per 100 words - Bulk document scans: 20% discount vs realtime - Auto-refill available - Free access available for academic researchers ### Enterprise Plan - Custom pricing based on volume - Features: - Progressive volume discounting - SOC 2 compliance - Enterprise security features - Developer API access - Dedicated support channel ### High Volume Support - Custom solutions available for queries exceeding Pro plan limits - Research API credits available for non-commercial projects - Direct support for large-scale AI detection implementations ## Additional Features - Word-based billing: usage is metered by word count ($0.05 per 100 words realtime; bulk jobs 20% cheaper) - Interpretability analysis showing AI-generated sections - File upload capabilities with OCR for scanned documents - Integration support for educational platforms including Canvas and Google Classroom - SOC 2 Type 2 verification by AssuranceLab ## Data Privacy & Compliance - Full regulatory compliance for educational institutions - Enterprise-level security protocols - Detailed documentation available for data privacy requirements Here's the LLMS.txt document for the Pangram careers page: -------------------------------------------------------------------------------- title: "Careers | Pangram" description: "Current job openings and career opportunities at Pangram Labs, an AI detection technology company based in Brooklyn, NY" last_updated: "2024" source: "https://jobs.ashbyhq.com/pangramlabs" -------------------------------------------------------------------------------- ## Company Overview Pangram Labs develops AI detection technology with multiple product offerings: Products: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities Market Focus: - Education sector (teachers) - Publishing industry - Content moderation/Trust and safety Technical Resources: - AI detection model - Technical documentation - Published research paper (arxiv.org/abs/2402.14873) Company Certifications: - SOC2 TYPE2 verified by AssuranceLab Contact Information: - Email: info@pangram.com - Location: Brooklyn, NY - Community: Discord community available - Social Media: Present on Instagram, Twitter, and LinkedIn -------------------------------------------------------------------------------- title: "About Us | Pangram" description: "Pangram Labs builds AI detection tools to mitigate issues from generative AI proliferation, founded by Stanford alumni Max Spero and Bradley Emi" last_updated: "2024" source: "https://www.pangram.com/about-us" -------------------------------------------------------------------------------- # Company Overview Pangram Labs focuses on mitigating issues caused by generative AI proliferation, aiming to ensure powerful language models have a net positive impact on society. # Founders ## Max Spero - CEO, Co-founder - Machine learning engineer with experience at Nuro (led active learning), Google, Two Sigma, and Yelp - B.S. in theoretical computer science and M.S. in artificial intelligence from Stanford University - Expertise in deploying machine learning products - Active member of Magic: the Gathering cube community ## Bradley Emi - CTO, Co-founder - AI researcher specializing in deep learning products - Led deep learning research group at Absci (generative AI drug discovery) - Former core computer vision team member at Tesla Autopilot - Published multiple deep learning research papers with Stanford Vision Lab - B.S. in physics and M.S. in artificial intelligence from Stanford University - Interests include education, philosophy, and golf # Origin Max Spero and Bradley Emi met as freshman year dormmates at Stanford University. Both shared enthusiasm for AI's potential to transform society before collaborating to establish Pangram Labs. # Product Categories ## Solutions - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities ## Use Cases - Teachers - Publishers - Content Moderation # Resources ## Technical Documentation - AI detection model details - Technical implementation guide - Published technical report (arxiv.org/abs/2402.14873) - Data privacy documentation ## Educational Resources - The State of AI Detection in 2025 - Most common AI phrases # Compliance - SOC2 TYPE2 certified (verified by AssuranceLab) # Contact Information - General inquiries: info@pangram.com - Career opportunities: careers@pangram.com - Community: Discord community available - Social presence: Instagram, Twitter, LinkedIn # Additional Information - Actively hiring and seeking mission-aligned individuals - Data privacy FAQ and status monitoring available - Terms of Service and Privacy Policy documentation accessible Here's my LLMS.txt extraction for the Press page: -------------------------------------------------------------------------------- title: "Press | Pangram" description: "Press coverage and media appearances featuring Pangram Labs' AI detection technology across major publications" last_updated: "August 2026" source: "https://www.pangram.com/press" -------------------------------------------------------------------------------- # Press Coverage Overview Pangram Labs provides AI detection technology and has received coverage in major media outlets including the New York Times, TechCrunch, Nature, The Atlantic, Wired, The Washington Post, The Guardian, Bloomberg, Science, and the Financial Times. Press inquiries can be directed to press@pangram.com. Funding milestones covered by press: $4 million raised (Reuters, June 2025) and a $9 million round (TechCrunch, July 2026). ## Recent Media Coverage (Chronological, newest first) 1. New York Times (July 29, 2026) — "Sick of A.I.-Generated Content? The 'Slop Janitor' Is Here to Help." by Niko Gallogly 2. TechCrunch (July 29, 2026) — "As AI content floods the internet, Pangram raises $9M to detect it" by Rebecca Bellan 3. Snopes (July 29, 2026) — "A note to Snopes readers about 13 articles from 2024" by Jessica Lee 4. Nature (July 6, 2026) — "Universities are relying on AI-detection software to catch cheating. How well do the programs work?" by Anna McKie 5. Substack (June 30, 2026) — "5 Rules of AI Writing" by Tom Rachman 6. Springer (June 29, 2026) — "Who wrote this? Evaluating the reliability of AI detection tools in higher education" by Van Vlasselaer, Van Droogenbroeck & Spruyt 7. Substack (June 23, 2026) — "Q&A: Pangram CEO" by Tom Rachman 8. The Atlantic (May 21, 2026) — "This Literary AI Scandal Changes Everything" by Vauhini Vara 9. The Washington Post (May 20, 2026) — "These 5 charts show how ChatGPT is flooding our lives" by Kevin Schaul 10. Futurism (May 20, 2026) — "Top Literary Magazine Offers Bizarre Response to Accusations that it Published an AI Generated Short Story" by Frank Landymore 11. Wired (May 19, 2026) — "Literary Prize winners are facing AI allegations. It feels like the new normal" by Miles Klee 12. The Guardian (May 19, 2026) — "Obvious markers of AI: doubts raised over winner of short story prize" by Aisha Down and Ella Creamer 13. X (May 18, 2026) — "@ashebytes podcast interview" (podcast) 14. Substack (May 18, 2026) — "the internet is full of people who never say anything" by Ochuko (As Seen On By Ochuko) 15. The Atlantic (May 1, 2026) — "Did a Human Write This?" by Charlie Warzel 16. 404 Media (April 27, 2026) — "Study Finds A Third of New Websites are AI-Generated" by Matthew Gault 17. 404 Media (April 27, 2026) — "People Using AI to Represent Themselves in Court Are Clogging the System" by Emanuel Maiberg 18. User Mag (April 27, 2026) — "How Much of Substack Is Actually AI" by Taylor Lorenz 19. The Leverage (April 24, 2026) — "What's the Bet: Pangram" by Evan Armstrong 20. Wired (April 22, 2026) — "The Pope's Warnings About AI Were AI-Generated, a Detection Tool Claims" by Miles Klee 21. Wired (April 15, 2026) — "AI Slop Is Making the Internet Fake-Happy" by Kate Knibbs 22. Imperial College London, Internet Archive, Stanford University (April 14, 2026) — "The Impact of AI-Generated Text on the Internet" by Jonas Dolezal, Sawood Alam, Mark Graham, Maty Bohacek 23. Bloomberg Odd Lots (April 2, 2026) — "This Is How To Tell if Writing Was Made by AI" with Joe Weisenthal & Tracy Alloway (podcast) 24. The Atlantic (March 25, 2026) — "How AI Is Creeping Into the New York Times" by Vauhini Vara 25. New York Times (March 19, 2026) — "Horror Novel 'Shy Girl' Canceled Over Suspected A.I. Use" by Alexander Alter 26. Financial Times (February 1, 2026) — "Artificial intelligence researchers hit by flood of slop" by Melissa Heikkilä 27. The Atlantic (January 22, 2026) — "Science is Drowning in AI Slop" by Ross Andersen 28. The Wellesley News (December 9, 2025) — "Wellesley pilots AI checker Pangram amid faculty concerns about AI" by Jessica Chen and Noufeesa Yahyaoui 29. The Chronicle of Higher Education (December 8, 2025) — "The Conference Where ChatGPT Wrote One in Five Reviews (Maybe)" by Stephanie M. Lee 30. NBC Bay Area (December 1, 2025) — "AI is reshaping the newsroom" by Scott McGrew 31. Nature (November 27, 2025) — "Major AI conference flooded with peer reviews written fully by AI" by Miryam Naddaf 32. Inside Higher Ed (November 19, 2025) — "AI Likely Driving Surge in Letters to the Editor" by Kathryn Palmer 33. Press Gazette (November 12, 2025) — "Google promises Discover 'fix' as more fake AI stories top rankings" by Rob Waugh 34. WKRN (November 12, 2025) — "Inspera and Pangram in Partnership – Bringing Accuracy and Transparency to Educators" (EIN Presswire) 35. Science (November 3, 2025) — "Letters to scientific journals surge as 'prolific debutante' authors likely use AI" by Jeffrey Brainard 36. The Decoder (November 2, 2025) — "Pangram achieves near-perfect results in AI text detection tests, study reveals" by Jonathan Kemper 37. Press Gazette (October 29, 2025) — "Study claims 9% of US newspaper articles at least partly AI generated" by Charlotte Tobitt 38. The Innovator (September 19, 2025) — "Generative Watermarking: A Building Block to Digital Trust" by Jennifer L. Schenker 39. Tom's Guide (September 17, 2025) — "I tested dozens of AI detectors — this one (claims 99% accuracy) beat the rest" by Amanda Caswell 40. Science (September 12, 2025) — "Far more authors use AI to write science papers than admit it, publisher reports" by Jeffrey Brainard 41. Nature (September 11, 2025) — "AI tool detects LLM-generated text in research papers and peer reviews" by Miryam Naddaf 42. Press Gazette (September 3, 2025) — "AI hoaxer, a fake X account and the case of Margaux Blanchard" by Charlotte Tobitt 43. TechRadar (September 2, 2025) — "I'm an AI expert and here's why fake AI reviews are going to be a massive problem, very soon" by Max Spero 44. KQED Close All Tabs (August 27, 2025) — "Teachers Strike Back Against AI Cheating" by Morgan Sung 45. Delco Today (August 23, 2025) — "DCCC English Professor Uses AI to Increase Students' Engagement and Prepare Them for the Future" by Helen Harris 46. Edtech Insiders (August 22, 2025) — "Week in EdTech 8/13/25: Feat. Max Spero of Pangram Labs" by Alex Sarlin 47. Education Writers Association (August 22, 2025) — "New Partnership Between Pangram and Qwoted Gives Journalists a Reliable Way to Spot AI-Text" 48. The Cheat Sheet (August 12, 2025) — "390: Pangram on AI Detection Accuracy, Transparency" with Max Spero 49. Inc. (August 11, 2025) — "Fake AI Reviews Are Spreading Fast. Here's What Businesses Can Do About It" by Chris Morris 50. The AI Journal (August 5, 2025) — "AI Detection as a Transparency Tool Can Help Students Think for Themselves" by Max Spero 51. FE News (August 5, 2025) — "My Computer Tells Me My Teachers Were Right All Along" by Bradley Emi 52. EdSurge (July 21, 2025) — "I Embraced AI in My Community College English Class — and My Students Loved It" by Susan E. Ray 53. School for Startups Radio (July 16, 2025) — "Max Spero on School for Startups Radio" with James Beach 54. eLearn Magazine (July 10, 2025) — "AI By Eye: Improving Your AI Writing Detection Skills—And Where Detection Tools Fit" by Max Spero 55. Reuters (June 24, 2025) — "Former Tesla, Google engineers raise $4 million for AI-text detection startup Pangram" 56. PlagiarismToday (June 4, 2025) — "The Problem with AI Polishing" by Jonathan Bailey 57. The Atlantic (April 29, 2025) — "The Great Language Flattening" by Victoria Turk 58. NewsGuard Reality Check (January 23, 2025) — "Russian Propagandist Turns His Sights to German Election" by Leonie Pfaller, Roberta Schmid, and McKenzie Sadeghi 59. Associated Press (December 23, 2024) — "The Internet Is Rife With Fake Reviews. Will AI Make It Worse?" by Haleluya Hadero 60. Rest of World (November 12, 2024) — "Phony X accounts are meddling in Ghana's election" by Caroline Haskins 61. Fortune (October 29, 2024) — "Researchers disagree about the speed of gen AI adoption. But one thing's clear: The tech is increasingly everywhere" by Sharon Goldman 62. Wired (October 28, 2024) — "AI Slop Is Flooding Medium" by Kate Knibbs ## Product Solutions - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual Capabilities ## Resources Available Technical Documentation: - AI detection model details - Technical implementation guide - Published technical report (arxiv.org/abs/2402.14873) - Data privacy documentation Educational Resources: - The State of AI Detection in 2025 - Most common AI phrases -------------------------------------------------------------------------------- title: "Sign up | Pangram" description: "Pangram provides AI writing detection tools through a dashboard and browser extension with daily free checks." last_updated: "August 08, 2025" source: "https://www.pangram.com/signup" -------------------------------------------------------------------------------- # Pangram AI Detection Platform Pangram offers AI writing detection capabilities through a comprehensive dashboard system. Users receive twenty free AI checks per day (up to 2,000 words daily). ## Core Features - Dashboard interface for AI writing detection - Chrome extension compatible with any Learning Management System (LMS) - Single-click AI detection functionality in Google Docs - Full-document analysis with high accuracy rates ## Account Access - Sign up available through Google authentication - Alternative email and password registration option - Free tier includes twenty daily AI checks (up to 2,000 words per day) ## Additional Information - Product and service email updates available with opt-out option - Existing users can access the platform through the login portal at pangram.com/login ## Platform Benefits - Streamlined AI detection workflow - LMS integration capabilities - Google Docs compatibility - Comprehensive document analysis - User-friendly dashboard interface Authentication options include Google single sign-on or traditional email/password registration. The platform sends optional product and service updates to registered users, with the ability to opt out of communications at any time. -------------------------------------------------------------------------------- title: "Log in | Pangram" description: "AI detection platform offering dashboard analytics and LMS integration for identifying AI-generated content" last_updated: "August 08, 2025" source: "https://www.pangram.com/login" -------------------------------------------------------------------------------- Pangram provides AI detection capabilities through multiple access points and features: Core Features: - Dashboard interface for AI writing detection - Twenty daily complimentary AI content checks (up to 2,000 words per day) - Chrome browser extension compatible with Learning Management Systems - Single-click AI detection functionality within Google Docs - Comprehensive full-document analysis with high accuracy rates Authentication Options: - Google SSO (Single Sign-On) integration - Email and password credentials - Password recovery functionality - New account creation during Google authentication Account Creation: - Automatic account generation available through Google authentication - Standard signup process for email-based accounts - Optional product and service email updates with opt-out capability Access Points: - Direct login at pangram.com/login - Alternative signup path at pangram.com/signup - Password reset available at pangram.com/forgot-password Note: New users receive automatic account creation when authenticating through Google SSO. Marketing communications are optional with user control over email preferences. -------------------------------------------------------------------------------- title: "AI Detection Dashboard | Pangram" description: "AI text detection platform offering analysis dashboard with LLM origin detection and segment-by-segment AI likelihood scoring" last_updated: "2024" source: "https://www.pangram.com/solutions/dashboard" -------------------------------------------------------------------------------- # Core Dashboard Functionality Pangram's AI Detection Dashboard analyzes text for AI authorship through both direct text input and file uploads. The system employs a sliding window approach to analyze individual text segments, providing granular AI likelihood scores for each segment. Maximum AI likelihood scores indicate the model's highest confidence level for AI-generated content within the text. # LLM Origin Detection The platform identifies the specific Large Language Model (LLM) that generated text by analyzing stylistic fingerprints. Supported LLM detection includes: - GPT-3.5 - GPT-4 - Gemini - Mistral - Claude - Llama # Dashboard Features ## Document Management - Chronological history tracking - Result filtering capabilities - Export functionality - Result sharing options - File upload support for .docx and .rtf formats - OCR capabilities for scanned documents # Pricing Plans ## Individual Plans - Free: $0 - Scan up to 2,000 words per day - AI assistance detection, interpretability features - File upload and OCR support - 3 image detection scans daily - Individual: $20/month, or $15/month billed annually ($180/year) - Scan up to 300,000 words per month - Plagiarism detection on every scan - 100 image detection scans monthly - Full dashboard access, interpretability features, file upload and OCR - Professional: $65/month, or $45/month billed annually ($540/year) - Scan up to 1,500,000 words per month - Plagiarism detection on every scan - 500 image detection scans monthly - $200 in monthly API usage included - Team: $20 per seat/month, or $15 per seat/month billed annually - 300,000 words per month per seat; starts at 2 seats - Admin controls, member management, unified billing ## Educational Institution License - Institutional license (contact Pangram for pricing) - Unlimited AI checks and plagiarism checks integrated directly into the LMS - Full regulatory data compliance; no training on student data - Integrations with Canvas, Brightspace, Moodle, Google Classroom, Google Docs, and more - Visibility into usage, trends, and statistics on AI use and plagiarism ## Developer Credit - Pay-as-you-go prepaid credits: add funds from $5 to $2,000 - API key included - Realtime checks: $0.05 per 100 words - Bulk document scans: 20% discount vs realtime - Auto-refill available - Free access available for academic researchers ## Enterprise Plan - Custom pricing - Progressive volume discounting - SOC 2 compliance - Enterprise security features - API access - Dedicated development and support channel # Technical Compliance - SOC2 Type 2 verification by AssuranceLab - Data privacy protections - Regulatory compliance for educational institutions -------------------------------------------------------------------------------- title: "Technical Report on the Pangram AI-Generated Text Classifier" paper_id: 2402.14873 authors: Bradley Emi, Max Spero submission_date: February 21, 2024 latest_version: v3 (July 29, 2024) category: Computer Science > Computation and Language (cs.CL) -------------------------------------------------------------------------------- PAPER CLASSIFICATION: - Primary: Computation and Language (cs.CL) - Secondary: Artificial Intelligence (cs.AI) - MSC Class: 68T50 - ACM Class: I.2.7 ABSTRACT: Pangram Text is a transformer-based neural network designed to differentiate between AI-generated and human-written text. The system achieves 38x lower error rates compared to existing commercial AI detection tools and zero-shot methods like DetectGPT. Testing covered 10 text domains: student writing, creative writing, scientific writing, books, encyclopedias, news, email, scientific papers, and short-form Q&A, across 8 open and closed-source language models. The system implements "hard negative mining with synthetic mirrors" training algorithm, delivering significantly reduced false positive rates for high-data domains like reviews. Testing confirms no bias against non-native English speakers and successful generalization to unseen domains and models. VERSION HISTORY: - v1: Feb 21, 2024 17:13:41 UTC (416 KB) - v2: Feb 26, 2024 05:28:41 UTC (418 KB) - v3: Jul 29, 2024 08:27:34 UTC (309 KB) AVAILABLE FORMATS: - PDF - HTML (experimental) - TeX Source - Other formats CITATION RESOURCES: - NASA ADS - Google Scholar - Semantic Scholar BIBLIOGRAPHIC TOOLS: - Bibliographic Explorer - Connected Papers - Litmaps - scite Smart Citations CODE/DATA TOOLS: - alphaXiv - CatalyzeX Code Finder - DagsHub - GotitPub - Hugging Face - Papers with Code - ScienceCast DEMO PLATFORMS: - Replicate - Hugging Face Spaces - TXYZ.AI LICENSE: Non-exclusive distribution license 1.0 -------------------------------------------------------------------------------- title: "Multilingual AI Detector (Spanish, French, Arabic & More…) | Pangram" description: "AI detection tool that identifies AI-generated content across 20+ languages with 99%+ accuracy" last_updated: "2025" source: "https://www.pangram.com/solutions/multilingual" -------------------------------------------------------------------------------- # Core Capabilities Pangram's multilingual AI detector identifies AI-generated and human-written content across 20+ languages with over 99% accuracy. The system provides AI Likelihood scores and maintains near-zero false positives. # Supported Languages Full language support includes: English, Arabic, Chinese, Czech, Dutch, French, German, Greek, Hindi, Hungarian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Turkish, Ukrainian, Urdu, and Vietnamese. # Accuracy Metrics by Language Language-specific accuracy rates: - Arabic: 99.95% accuracy, 0.10% false positive, 0.00% false negative - German: 99.85% accuracy, 0.00% false positive, 0.32% false negative - Spanish: 100.00% accuracy, 0.00% false positive, 0.00% false negative - Persian: 100.00% accuracy, 0.00% false positive, 0.00% false negative - French: 100.00% accuracy, 0.00% false positive, 0.00% false negative - Hindi: 99.79% accuracy, 0.00% false positive, 0.42% false negative - Chinese: 99.95% accuracy, 0.00% false positive, 0.11% false negative # Technical Implementation The system employs specialized tokenization for non-English languages and uses active learning to improve accuracy. The detector can identify both direct AI-generated content and text processed through translators ("double translated" content) with 99.99% accuracy. # Primary Use Cases - Education: Verification of student work and ESL content - Publishing: Content authenticity verification - Content moderation: AI-generated content detection - Marketing: Original content verification - Government: Document authenticity checking # Pricing Plans Individual Plans: - Free: $0, scan up to 2,000 words per day - Individual: $20/month ($15/month billed annually), 300,000 words per month, plagiarism detection included - Professional: $65/month ($45/month billed annually), 1,500,000 words per month, plagiarism detection and $200 monthly API usage included - Team: $20 per seat/month ($15 per seat/month billed annually), 300,000 words per seat per month Educational Institution: - Institutional license (contact Pangram for pricing) - Unlimited AI and plagiarism checks integrated directly into the LMS - Full regulatory compliance - Integrations with Canvas, Brightspace, Moodle, Google Classroom, and Google Docs Developer: - Pay-as-you-go prepaid credits (add funds from $5 to $2,000) - API key included - Realtime checks $0.05 per 100 words; bulk jobs 20% cheaper # Development Background Created by AI researchers from Stanford, Tesla, and Google. Endorsed by the University of Maryland as the "most reliable and most accurate tool on the market." -------------------------------------------------------------------------------- title: "Plagiarism Detection Extension | Pangram" description: "Pangram's plagiarism detection system scans text against billions of sources while simultaneously checking for AI-generated content." last_updated: "August 08, 2025" source: "https://www.pangram.com/solutions/plagiarism" -------------------------------------------------------------------------------- # Core Functionality Pangram's plagiarism detection system scans submitted text against billions of webpages, academic papers, news articles, books, and online content. The system simultaneously checks for both traditional plagiarism and AI-generated content. # Detection Process 1. Users upload text through the Pangram Dashboard with plagiarism detection enabled 2. System scans content across multiple source types: - Academic journals - Research papers - News articles - Books - Web content 3. Direct matches are flagged for review 4. AI-generated content and common AI phrases are identified 5. Detailed report generated showing content origins, shareable and downloadable # Pricing Tiers ## Individual Plans - Free: $0, scan up to 2,000 words per day - Individual: $20/month ($15/month billed annually), 300,000 words per month - Plagiarism detection available on every scan - Full dashboard access, interpretability features, file upload and OCR - Professional: $65/month ($45/month billed annually), 1,500,000 words per month - Plagiarism detection, 500 image detection scans, and $200 monthly API usage included - Team: $20 per seat/month ($15 per seat/month billed annually), 300,000 words per seat per month ## Educational Institutions - Institutional license (contact Pangram for pricing) - Unlimited AI checks and plagiarism checks integrated directly into the LMS - Automatic plagiarism detection including copy-paste plagiarism and re-used submissions from past years - Full regulatory compliance; no training on student data - Integrations with Canvas, Brightspace, Moodle, Google Classroom, and Google Docs ## Developer Credit - Pay-as-you-go prepaid credits: add funds from $5 to $2,000 - API key included - Realtime checks: $0.05 per 100 words - Bulk document scans: 20% discount vs realtime - Auto-refill available - Free access available for academic researchers ## Enterprise - Custom pricing based on volume - SOC 2 compliance - Enterprise security features - API access - Dedicated development and support channel # Sample Detection Example The system demonstrated plagiarism detection using a Great Gatsby essay example, identifying two sources: 1. New York Times: Content about the novel's century-long legacy 2. SparkNotes: Analysis of American progress symbols as corrupting forces # Integration Options - Dashboard access - Chrome extension - API integration - Canvas LMS - Google Classroom - Developer API access Note: Plan usage is metered by word count; each plan includes a monthly (or daily, for the Free plan) word allowance. -------------------------------------------------------------------------------- title: "AI Detection Integrations | Pangram" description: "Pangram's AI detection technology integrates across platforms through browser extensions, LMS systems, and custom solutions" last_updated: "2024" source: "https://www.pangram.com/solutions/integrations" -------------------------------------------------------------------------------- # Core Integrations Pangram offers three primary integration methods for AI detection: 1. Chrome Extension - Provides AI text detection capabilities on any website - One-click activation for instant analysis - Matches dashboard accuracy and detection capabilities - Available through Chrome Web Store 2. LMS Integrations (Canvas, Brightspace, Moodle, Google Classroom) - Automatic background AI detection processing on assignment submissions - Direct integration with Canvas SpeedGrader - AI and plagiarism scores surfaced within the LMS interface - Google Docs integration for in-document checks - Designed specifically for educational institutions 3. Custom Integration Solutions - Available for specialized implementation needs - Custom development and integration support - Direct consultation available through education team # Technical Infrastructure Pangram maintains SOC2 Type 2 certification, verified by AssuranceLab, ensuring security and reliability across all integrations. # Additional Platform Features Pangram's broader platform capabilities include: - Multilingual detection support - Plagiarism detection - API access - Dashboard interface - Content moderation tools - Educational resources and technical documentation # Support & Implementation Custom integration requests are handled through direct consultation with the Pangram team. Educational institutions can contact the dedicated education team for Canvas integration support and implementation guidance. # Documentation & Resources Technical documentation includes: - AI detection model documentation - Implementation guides - Technical report (available on arXiv: 2402.14873) - Data privacy documentation - Integration-specific guides Contact: info@pangram.com -------------------------------------------------------------------------------- title: "AI Image Detector - Check Images for AI Generation | Pangram" description: "Detect AI-generated images with Pangram's AI Image Detector. Upload an image, paste a URL, or verify files in batches." last_updated: "August 2026" source: "https://www.pangram.com/image-detector" -------------------------------------------------------------------------------- # AI Image Detector Pangram's AI Image Detector identifies AI-generated images. It is available on all plans, including the Free plan. ## How It Works - Upload an image, paste an image URL, or verify files in batches - The detector returns a likelihood that the image was AI-generated - Complements Pangram's text detection for full-content verification workflows ## Plan Quotas - Free plan: 3 image detection scans daily - Individual plan: 100 image detection scans monthly - Professional plan: 500 image detection scans monthly ## Primary Use Cases - Verifying authenticity of images in publishing and journalism - Moderating AI-generated imagery on user-generated content platforms - Checking submitted media in education and research contexts -------------------------------------------------------------------------------- title: "Contact Us | Pangram" description: "Contact information and support channels for Pangram's AI content detection solutions across education, enterprise, and research sectors" last_updated: "2024" source: "https://www.pangram.com/contact-us" -------------------------------------------------------------------------------- # Contact Options Pangram provides specialized contact channels for different user segments requiring AI-generated content detection solutions: ## Education Support Educational institutions seeking authorship tracking, academic integrity protection, and student outcome improvement can access LMS integrations and standalone products. Direct education inquiries through dedicated education contact portal. ## Enterprise Solutions Organizations requiring production-ready AI content detection receive customized solutions based on specific needs assessment and scoping. Enterprise inquiries handled through dedicated enterprise contact channel. ## Research Collaboration Non-commercial research studies focused on AI-generated content can receive support from Pangram. Research collaboration requests evaluated through dedicated research contact portal. ## Media Relations Press and media inquiries handled at press@pangram.com ## General Feedback Bug reports, feature requests, and general feedback accepted through dedicated feedback contact form. # Additional Support Channels ## Community Support - Discord Community: Available for immediate support at discord.gg/f7jDAPzWH3 - Social Media Presence: - Twitter: @pangramlabs - LinkedIn: /company/pangramlabs - Instagram: @pangramlabs ## General Contact Email: info@pangram.com # Product Solutions Pangram offers multiple AI detection tools: - Dashboard - Chrome Extension - API - Platform Integrations - Plagiarism Detection - Multilingual Capabilities # Compliance and Security - SOC2 TYPE2 certified (verified by AssuranceLab) - Data privacy documentation available - Technical specifications detailed in arxiv.org/abs/2402.14873 Note: All rights reserved © Pangram. -------------------------------------------------------------------------------- title: "Feedback | Pangram" description: "Pangram Labs' feedback submission system for bug reports, feature requests, user experience feedback, and subscription inquiries" last_updated: "2024" source: "https://www.pangram.com/contact-us/feedback" -------------------------------------------------------------------------------- # Feedback System Overview Pangram Labs maintains a dedicated feedback system for users to submit inquiries across multiple categories. Users can provide feedback through a structured form or join the Discord community for immediate support. ## Feedback Categories - Bug Reports - Feature Requests - User Experience Feedback - Subscription and Membership Questions - Other General Inquiries ## Support Channels - Web form submission with required email field - Discord Community: https://discord.gg/f7jDAPzWH3 # Company Solutions ## Core Products - Dashboard - Chrome Extension - API Access - Third-party Integrations - Plagiarism Detection - Multilingual Capabilities ## Industry Use Cases - Education (Teachers) - Publishing Industry - Content Moderation/Trust & Safety ## Technical Resources - AI Detection Model Documentation - Technical Implementation Guide - Research Paper: arxiv.org/abs/2402.14873 - Data Privacy Framework ## Educational Resources - State of AI Detection in 2025 Report - Common AI Phrases Analysis - Assessment Strategies for AI Era - Academic Integrity Framework # Company Information - Email: info@pangram.com - SOC2 Type 2 Certified (Verified by AssuranceLab) - Founded/Active: Operating as of 2024 - Social Presence: Instagram, Twitter, LinkedIn - Community Platform: Discord # Legal Framework - Terms of Service - Privacy Policy - Data Privacy FAQ - System Status Monitoring: status.pangram.com Note: All rights reserved by Pangram Labs © 2024 -------------------------------------------------------------------------------- title: "AI Detector for Teachers | Pangram" description: "AI detection tool for educators to verify student work authenticity and detect content generated by LLMs like ChatGPT" last_updated: "August 08, 2025" source: "https://www.pangram.com/use-cases/teachers" -------------------------------------------------------------------------------- # Core Capabilities Pangram's AI detector identifies content produced by all major LLM-powered tools, including ChatGPT, with verified reliability. The University of Maryland has reviewed and confirmed Pangram as the most reliable and accurate AI detection tool available. # Educational Implementation ## Current Usage Areas - Universities/Professors: Detection of LLM-generated content in essays and assignments - Teachers: Verification of student work originality - Personal Tutors: Progress tracking and plagiarism prevention - School Boards: Academic integrity policy enforcement - Students: Self-verification of work originality - Researchers: Analysis of AI usage trends and societal impact ## Technical Specifications - 99.99% accuracy in detecting AI content - False positive rate of 1 in 10,000 - Capability to identify specific sections of AI-generated text - Detection support for all major LLMs including ChatGPT, Claude, Quillbot, Grammarly, and AI humanizers ## Integration Capabilities Seamless integration with educational platforms: - Google Classroom - Google Docs - Canvas - Moodle - All web-based Learning Management Systems via Chrome extension # Usage Statistics and Impact ## Student AI Usage - 91% of students acknowledge using AI for academic work - Tool identifies specific LLM sources from Claude to GPT-4 ## Teacher Feedback - 95% of teachers rate Pangram's detection as most accurate - 93% endorse Pangram's transparency features - 86% report improved student-teacher discussions about AI use, leading to enhanced AI literacy # Partnerships and Trust Verified partnerships with: - Canvas - Google Classroom - Quora - Tremau - The Transparency Company - Newsguard # Primary Benefits 1. Reliable AI detection for fair grading 2. Minimal false positives ensuring trustworthy results 3. Detailed authorship analysis 4. Comprehensive LLM detection coverage 5. Integration with existing educational workflows The platform serves as both a detection tool and educational resource, facilitating discussions about appropriate AI use in academic settings while maintaining academic integrity standards. -------------------------------------------------------------------------------- title: "Log in | Pangram" description: "AI detection platform offering dashboard access and integration tools for identifying AI-generated content" last_updated: "August 08, 2025" source: "https://www.pangram.com/dashboard" -------------------------------------------------------------------------------- Pangram Labs provides AI detection capabilities through multiple access points and features: Core Features: - Dashboard interface for AI content checking - Twenty daily complimentary AI detection scans (up to 2,000 words per day) - Chrome extension compatible with Learning Management Systems - Google Docs integration with single-click AI detection - Comprehensive full-document analysis with high accuracy rates Authentication Options: 1. Google SSO (Single Sign-On) integration 2. Email and password credentials 3. New account creation through Google authentication Account Creation: - Automatic account generation available through Google authentication - Optional email signup pathway - Password recovery functionality available - Product and service updates provided with opt-out capability Access Points: - Direct dashboard login - Account signup for new users - Password reset functionality - Google-based authentication User Communications: - Product updates delivered via email - Service notifications available - Opt-out options provided for all communications Platform Integration: - LMS compatibility through Chrome extension - Google Docs native integration - Full document analysis capabilities Note: Platform provides both authentication and registration pathways with flexible options for account creation and management. Email communications include product updates with user control over notification preferences. Here's the LLMS.txt document for the Instagram page: -------------------------------------------------------------------------------- title: "Pangram Labs (@pangramlabs) • Instagram photos and videos" description: "Pangram Labs offers an AI writing detection tool and Chrome extension developed by former Stanford researchers" last_updated: "August 08, 2025" source: "https://www.instagram.com/pangramlabs/" -------------------------------------------------------------------------------- # Pangram Labs (@pangramlabs) • Instagram photos and videos Pangram Labs develops AI writing detection tools verified by third-party testing. The company was founded by former Stanford researchers and maintains an Instagram presence with 7 followers, 1 post, and 0 accounts followed. ## Product Offerings Pangram's main product is an AI writing detection tool with high accuracy claims. A Chrome extension version provides easy-to-use AI text detection capabilities directly in the browser. ## Social Media Presence Instagram Activity: - Total Posts: 1 - Followers: 7 - Following: 0 - Latest Post Date: January 29, 2025 - Content Focus: Product demonstrations and AI detection capabilities ## Company Overview Pangram Labs specializes in artificial intelligence detection technology, focusing on identifying AI-generated writing. The technology has undergone third-party verification to validate its accuracy claims. Note: This document reflects the Instagram profile content as of the last update date. Additional company information and product details may be available through other channels. -------------------------------------------------------------------------------- title: "The State of AI Detection in 2025 | Pangram" description: "A webinar presentation by Pangram Labs' CTO on the current state and future of AI detection technology" last_updated: "2024" source: "https://www.pangram.com/resources/the-state-of-ai-detection-in-2025" -------------------------------------------------------------------------------- # The State of AI Detection in 2025 ## Webinar Overview Bradley Emi, CTO of Pangram Labs, delivered a comprehensive session on AI Detection at the ICAI conference. The webinar covers three main areas: 1. Current academic research on AI detection 2. Linguistic traits for identifying AI-generated writing 3. Ethical implementation of AI detectors in academic settings ## Key Topics - Analysis of academic research developments in AI detection - Identification of key linguistic markers in AI-generated content - Methods for detecting AI-written content manually - Ethical guidelines for implementing AI detection tools - Integration of AI detection within broader academic integrity frameworks ## Resources Available - Presentation slides - Complete transcript - YouTube recording of the session ## Company Details Pangram Labs offers multiple AI detection solutions: ### Products - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities ### Use Cases - Educational institutions - Publishers - Content moderation teams ### Technical Resources - AI detection model documentation - Technical implementation guides - Published technical report (arxiv.org/abs/2402.14873) - Data privacy documentation ## Compliance & Security - SOC2 TYPE2 certified (verified by AssuranceLab) - Comprehensive data privacy protocols - Terms of service and privacy policies available ## Contact Information - Email: info@pangram.com - Community: Discord server available - Social Media: Present on Instagram, Twitter, and LinkedIn Note: This webinar is part of Pangram's educational resource series, which includes additional materials on AI phrases, assessment strategies, and academic integrity approaches in the AI era. -------------------------------------------------------------------------------- title: "What is a humanizer? | Pangram" description: "A detailed explanation of AI text humanizers, their techniques, and detection capabilities" last_updated: "January 27, 2025" source: "https://www.pangram.com/blog/what-is-a-humanizer" -------------------------------------------------------------------------------- # What is a humanizer? Humanizers are tools designed to modify AI-generated text to evade AI detection systems. They target students and others seeking to hide AI-authored content from detection tools like Turnitin. ## Core Functionality and Techniques Pangram researchers analyzed 19 publicly available humanizers and identified several common modification techniques: ### Synonym Replacement Humanizers substitute words with synonyms, often degrading text quality and clarity. Example: - Original: "I need to get my car fixed because the engine is making a strange noise." - Modified: "I require to obtain my vehicle repaired because the motor is creating a peculiar sound." ### Nonsensical Phrase Insertion Humanizers add random, meaningless text segments to confuse AI detectors. Example: "...as a result, will put more effort into their studies. CGSizeMake pp 18-23. Last but not least…" ### Text Quality Degradation Humanizers deliberately lower writing quality through grammar and spelling errors. Example: - Original: "In an era dominated by technology, the convenience of cell phones has transformed the way we communicate and access information. However, this advancement comes with grave consequences, particularly when cell phones are used while operating a vehicle." - Degraded: "Now a days technology let us convenience to communicate and recieves info more easly by cellphones, but this like to say the darkside of the technology for use this devices while is driving, because cellphones are a little from a great danger for everybody…" ## Detection Capabilities Pangram's AI detection model can identify over 90% of humanized content, even when specifically designed to evade particular detection systems. Research shows AI detectors can be trained to detect humanized text through subtle adjustments to their models. ## Risks of Using Humanizers - Text quality degradation - Introduction of nonsensical content - High detection risk despite evasion attempts - Potential academic integrity violations Contact: info@pangram.com -------------------------------------------------------------------------------- title: "Third-Party Research Study Shows Pangram is the Most Robust AI Detector | Pangram" description: "Research from multiple universities demonstrates Pangram's superior effectiveness at detecting AI-generated text, even after translation attacks" last_updated: "October 30, 2024" source: "https://www.pangram.com/blog/esperanto" -------------------------------------------------------------------------------- Researchers from the University of Houston, UC Berkeley, UC Irvine and Esperanto AI conducted a comprehensive study demonstrating Pangram's superior effectiveness as an AI text detector compared to both commercial and open-source alternatives. Research Study Details: - Title: "Esperanto: Evaluating Synthesized Phrases to Enhance Robustness in AI Detection for Text Origination" - Focus: Effects of language translation on AI detection capabilities - Test scope: 720,000 documents including news articles, scientific abstracts, Reddit posts, and product reviews - AI models tested against: GPT-3.5-Turbo, LLaMA 3, Mistral, Phi3, and Yi Double Translation Attack Method: - Process involves translating AI-generated text to another language (e.g., Japanese) and back to English - Creates variations in phrasing similar to paraphrasing tools - Many competing detectors fail this test; example showed competitor dropping from high confidence to 15% AI probability - Pangram maintains 99.99% AI detection confidence even after double translation - Can identify specific AI model source (e.g., GPT-4) after translation Performance Metrics: - Evaluation metric: True Positive Rate (TPR) at 1% False Positive Rate (FPR) - Pangram achieved >96% recall across all domains at 1% FPR - Achieved 85% accuracy on challenging short-form content (40-50 word reviews) - Competitor Performance Degradation: - GPTZero: Dropped from 97% to 42% (news domain) - GPTZero: Dropped from 65% to 9% (reviews domain) - ZeroGPT: Unable to achieve 1% false positive rate in most domains - Academic detectors (RADAR, LLMDet): Less than 50% accurate Key Findings: - Most open-source and commercial detectors proved ineffective even before translation attacks - Many competitors cannot maintain consistent false positive rates across domains - Pangram demonstrated superior robustness against translation-based evasion techniques - Particularly effective performance on longer-form content - Only detector showing consistent reliability for academic and commercial applications The research validates Pangram's effectiveness as an AI detection tool that maintains accuracy even when faced with common evasion techniques like translation attacks, setting it apart from current market alternatives. -------------------------------------------------------------------------------- title: "How to spot AI reviews | Pangram" description: "Guide to identifying AI-generated reviews by analyzing their distinctive writing patterns and characteristics" last_updated: "December 5, 2023" source: "https://www.pangram.com/blog/how-to-spot-ai-reviews" -------------------------------------------------------------------------------- ChatGPT and other AI text generators leave distinctive patterns that make their writing recognizable, similar to how authors have recognizable styles. Five key indicators help identify AI-generated reviews: 1. Length: AI reviews tend to be significantly longer than typical human reviews 2. Empty Descriptors: AI uses generic positive adjectives and adverbs without specific content (examples: fantastic, exceptionally, beautiful, perfect, thoughtful) 3. Clichéd Phrases: Common use of stock phrases like "I can't speak highly enough" and "I couldn't be more pleased" 4. Prompt Repetition: AI often repeats the full name/subject in the first sentence, while humans assume context 5. Structured Format: AI follows rigid paragraph structures with clear intro, body paragraphs (each covering one aspect), and conclusion Case Study 1 - Google Maps Review: A review for Pouch Camporee Field demonstrates these AI indicators through: - Excessive length with multiple detailed paragraphs - Generic positive descriptors without specific details - Formal structured format with intro/body/conclusion - Full venue name repeated in opening - Clichéd phrases throughout Case Study 2 - Amazon Water Bottle Review Analysis: Review scored 1.5/5 on AI indicators: - Length: Long (matches AI pattern) - Empty descriptors: Only one instance of "really" (unlike AI) - Clichés: None present (unlike AI) - Prompt repetition: Uses simple "water bottle" instead of full product name (unlike AI) - Structure: Partially structured but includes irregular elements like out-of-place cons paragraph (unlike strict AI formatting) The article emphasizes that spotting AI text requires analyzing multiple indicators together rather than relying on any single characteristic. -------------------------------------------------------------------------------- title: "A trusted proofreading tool now uses AI to write for students | Pangram" description: "Grammarly has evolved from a basic proofreading tool into an AI-powered writing assistant that can generate content for students" last_updated: "March 6, 2025" source: "https://www.pangram.com/blog/grammarly-uses-ai-to-write-for-students" -------------------------------------------------------------------------------- Grammarly's Evolution to AI Writing Tool Grammarly has transformed from a 2009 spelling and grammar checker into an AI-powered writing platform with over 40 million users. The company now incorporates generative AI capabilities that can write content on behalf of students. Key Development Timeline: - 2019: Introduced tone detector using rules and machine learning - 2020: Made first external investment in Docugami, an AI document processing company - 2023: Launched GrammarlyGo using OpenAI's language models for text generation - Current: Offers complete AI writing capabilities including paragraph rewriting and essay generation Academic Impact and Institutional Response: - February 2024: Georgia college student placed on academic probation for AI-flagged paper despite claiming only basic Grammarly usage - August 2024: Notre Dame updated policies to classify Grammarly under generative AI tool restrictions - Multiple institutions now require clarification on acceptable Grammarly usage in academic work New Grammarly Features: - Authorship tracking that records writing process - Distinction between typed text, copied content, and AI-generated language - Process reports available for students to demonstrate appropriate tool usage - Basic proofreading features separate from generative AI capabilities Pangram's Role: - Detects when Grammarly usage extends beyond basic editing - Identifies AI-generated content from Grammarly's advanced features - Supports academic integrity monitoring - Provides guidance on appropriate tool usage policies Recommendations for Teachers: - Establish clear policies on acceptable Grammarly usage - Distinguish between basic proofreading and AI generation features - Implement verification processes for student submissions - Contact info@pangram.com for questions about Grammarly-edited text evaluation Current Grammarly Marketing: - Promotes "Responsible AI that ensures your writing and reputation shine" - Markets itself as "an AI writing partner that helps you find the words you need" - Offers both basic proofreading and advanced AI writing capabilities - Provides documentation features for academic integrity verification -------------------------------------------------------------------------------- title: "Walking Through AI's Most Overused Phrases | Pangram" description: Analysis of common phrases and patterns that indicate AI-generated text, based on Pangram's dataset research last_updated: "February 21, 2025" source: "https://www.pangram.com/blog/walking-through-ai-phrases" -------------------------------------------------------------------------------- Pangram has developed an "AI phrases" tool that highlights commonly overused phrases in AI-generated text. Their research identified several categories of AI-specific language patterns: AI Self-Referential Phrases (with frequency compared to human text): - "As an AI language model..." (294,000x more common) - "I do not have personal" (67,000x more common) - "Unfortunately, I do not have enough" (54,000x more common) - "language model, I can not" (53,000x more common) Common Overused Phrases (with frequency compared to human text): - "as a poignant" (49,000x more common) - "As a powerful reminder" (43,000x more common) - "reminder of the enduring" (31,000x more common) - "faced numerous challenges" (30,000x more common) - "Our results provide new insights into" (22,000x more common) - "into the complex interplay" (21,000x more common) - "serves as a powerful" (10,000x more common) - "providing valuable insights into" (5,000x more common) - "serves as a testament" (4,000x more common) - "newfound sense of purpose" (4,000x more common) - "even in the face of unimaginable" (3,000x more common) - "reminder of the potential" (3,000x more common) Causes of AI Phrase Overuse: 1. Mode collapse: AI outputs become repetitive by relying on high-probability word sequences 2. RLHF (Reinforcement Learning from Human Feedback): Human annotators' ratings encourage certain patterns and phrases 3. Model-specific characteristics: Different models develop distinct phrase preferences based on training data and optimization Pangram Team's Favorite AI Phrases: - Max (CEO): "In the ever-evolving" (11,000x more common) - Bradley (CTO): "important to note" (3,000x more common) - Lu (Founding Engineer): "intricate nature" (6,000x more common) - Elyas (Founding Engineer): "vibrant tapestry" (17,000x more common) Research from the University of Maryland (Jenna Russell, Marzena Karpinska, and Mohit Iyyer) demonstrates that different AI models develop distinct phrase preferences based on their training and optimization methods. -------------------------------------------------------------------------------- title: "Understanding the EU's New AI Law | Pangram" description: "Analysis of the EU's Artificial Intelligence Act requirements, implications, and compliance guidelines for online platforms and AI systems" last_updated: "March 23, 2024" source: "https://www.pangram.com/blog/ai-act" -------------------------------------------------------------------------------- The European Parliament passed the Artificial Intelligence Act (AI Act), establishing a comprehensive regulatory framework for AI products and services in EU member states. Key Risk Categories: - Unacceptable Risk: Systems using subliminal/manipulative techniques or threatening fundamental rights are banned - High Risk: Critical infrastructure, education, healthcare, law enforcement, border management, and election systems face strict requirements - Limited Risk: General Purpose AI systems, including generative AI, have lighter transparency obligations Enforcement: - Applies to both AI producers and users - Non-compliance penalties: Up to €35 million or 7% of global annual revenue - Affects any business with EU end-users utilizing AI systems Platform Requirements: 1. Content Disclosure: Must disclose AI-generated content 2. Moderation Controls: Must reject illegal AI-generated submissions 3. Training Data Transparency: Must publish summaries of copyrighted training data 4. Content Marking: AI-generated audio, images, video, and text must be marked in machine-readable format Implementation Timeline: - Expected to become law by May 2024 - Provisions implemented in stages - Most transparency requirements effective one year after enactment Recommended Company Actions: 1. Develop clear guidelines and governance for user-generated content 2. Implement efficient content moderation workflows for AI detection 3. Create automations to prevent AI-generated content spread Current Industry Response: - YouTube and Instagram implementing self-reporting systems for "realistic" AI-generated content - Unclear if self-reporting will satisfy Act requirements - Major platforms actively developing compliance strategies The Act requires both producers (e.g., OpenAI/Google) and users to pass accuracy and transparency tests for consumer awareness of AI interactions. Online platforms seeing increased AI-generated content must implement transparency and moderation controls, particularly for user-generated content platforms, publishers, marketplaces, and social media sites. Contact Information: Email: info@pangram.com -------------------------------------------------------------------------------- title: "Blog | Pangram" description: "Collection of articles about AI detection technology, product updates, and educational resources from Pangram Labs" last_updated: "August 7, 2025" source: "https://www.pangram.com/blog" -------------------------------------------------------------------------------- # Blog Articles ## Latest Articles ### Product Updates - "Do AI Detectors Work Against GPT-5?" by Bradley Emi (August 7, 2025) - "How well can Pangram detect reasoning models?" by Bradley Emi (July 16, 2025) - Analysis of detection capabilities for OpenAI's o3 series, Claude 4, and Gemini 2 - "Does Pangram detect Meta's Llama 4?" by Bradley Emi (April 6, 2025) ### Education Resources - "Introducing Pangram's Plagiarism Detection" by Elyas Masrour (August 1st, 2025) - "Mirror, Mirror On The Wall, Who's The Realest Of Them All?" by Jason Nicholson (July 25, 2025) - Discussion on learning processes and technology - "How accurate is Pangram AI Detection on ESL?" by Bradley Emi (April 23, 2025) - "How can teachers tell if something is AI-generated?" by Max Spero (April 9, 2025) - "Comprehensive Guide to Spotting AI Writing Patterns" by Bradley Emi (April 2, 2025) - "All About False Positives in AI Detectors" - Focus on Pangram's accuracy rates ### Case Studies - "Three percent of front-page Amazon reviews are now AI-generated" by Max Spero (July 25, 2025) - Study reveals 93% of AI-generated reviews have 'Verified Purchase' badge - "30 Best AI Detectors Reviewed in 2025 (Free & Paid)" by Max Spero (June 10, 2025) - Comparative analysis of AI detection tools ### News - "Why does Pangram have a minimum word count?" by Elyas Masrour (May 23, 2025) - "What happens when an AI detector makes a mistake?" by Max Spero (May 15, 2025) - "Which is better: Pangram or Turnitin?" by Max Spero (May 13, 2025) ## Product Features Pangram offers multiple solutions: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities ## Use Cases - Educational institutions and teachers - Publishers - Content moderation and trust & safety ## Technical Resources - AI detection model documentation - Technical report available on arXiv (2402.14873) - Data privacy documentation - Implementation guides ## Educational Resources - The State of AI Detection in 2025 - Most common AI phrases -------------------------------------------------------------------------------- title: "AI Detector Chrome Extension | Pangram" description: "Browser extension for detecting AI-generated text across websites with single-click functionality" last_updated: "2024" source: "https://www.pangram.com/solutions/chrome-extension" -------------------------------------------------------------------------------- # Pangram AI Detector Chrome Extension ## Core Functionality Pangram's Chrome extension enables AI-generated text detection across any website through a browser-based interface. The tool provides instant AI content analysis with a single click, matching the accuracy of Pangram's dashboard solution. ## Usage Instructions 1. Text Selection: Highlight target text and right-click to select "Check for AI content" 2. Floating Icon: Adjustable visibility settings for the detection icon 3. Results Display: Analysis results appear in bottom right corner of webpage ## Pricing Tiers ### Free Plan ($0) - Scan up to 2,000 words per day - AI assistance detection - Interpretability features - File upload and OCR for scanned documents - Chrome extension and Google Docs integration ### Individual Plan ($20/month, or $15/month billed annually) - Scan up to 300,000 words per month - Plagiarism detection on every scan - Full dashboard access - Interpretability features - File upload and OCR ### Professional Plan ($65/month, or $45/month billed annually) - Scan up to 1,500,000 words per month - Plagiarism detection on every scan - $200 in monthly API usage included - Full dashboard features ### Educational Institution (contact Pangram for pricing) - Unlimited AI checks and plagiarism checks integrated directly into the LMS - Full regulatory compliance - Integrations with Canvas, Brightspace, Moodle, Google Classroom, and Google Docs ### Developer Credit - Pay-as-you-go prepaid credits: add funds from $5 to $2,000 - API key included - Realtime checks: $0.05 per 100 words - Bulk document scans: 20% discount vs realtime - Auto-refill available - Free access available for academic researchers ### Enterprise Plan (Variable Pricing) - Progressive volume discounting - SOC 2 compliance - Enterprise security features - API access - Dedicated development support channel ## Technical Details - Usage is metered by word count against each plan's daily or monthly allowance - Supports OCR processing for scanned documents - Includes interpretability features for identifying specific AI-generated sections - Integrates with learning management systems (Canvas, Google Classroom) - SOC 2 Type 2 verified by AssuranceLab ## Additional Features - Dashboard integration - File upload capabilities - Plagiarism detection - Multilingual support - Content moderation tools - Academic integrity features -------------------------------------------------------------------------------- title: "Introducing Pangram's AI Phrases | Pangram" description: "Pangram launches new AI phrase detection capabilities to identify common patterns in AI-generated text and improve detection explainability" last_updated: "February 14, 2025" source: "https://www.pangram.com/blog/pangram-ai-phrases" -------------------------------------------------------------------------------- # AI Phrase Detection Overview Pangram has developed a new AI Phrases tool that identifies and highlights common phrases overused in AI-generated text. The system analyzes text patterns that appear more frequently in AI writing compared to human writing. ## Common AI Phrase Examples - "Complex tapestry" - "A testament to" - "Delve" (identified by Paul Graham) ## Technical Implementation Pangram's detection process works in two stages: 1. Initial AI detection scan of the document 2. Secondary scan specifically for common AI phrases The training methodology involves: - Analysis of tens of millions of human and AI-generated documents - N-Gram analysis to compare word sequence frequencies - Statistical comparison between human and AI writing patterns - Identification of phrases used significantly more often in AI text ## Purpose and Benefits The AI phrase detection system serves multiple purposes: - Preserves authentic human voices - Provides explainable evidence for AI detection results - Offers quantifiable data showing phrase frequency differences between AI and human writing - Enhances overall detection interpretability - Helps users understand detection methodology ## Integration The AI Phrases tool is integrated directly into the Pangram dashboard, highlighting overused AI phrases within scanned documents. When phrases appear hundreds or thousands of times more frequently in AI text versus human text, they are flagged for user review. ## Company Information Contact: info@pangram.com Technical documentation: Available via technical report at arxiv.org/abs/2402.14873 ## Additional Features Pangram's complete solution includes: - Dashboard - Chrome Extension - API access - Integration capabilities - Plagiarism detection - Multilingual support - Content moderation tools - Solutions for teachers, publishers, and content moderators The platform maintains SOC2 Type 2 certification, verified by AssuranceLab. -------------------------------------------------------------------------------- title: "Do AI Detectors Work Against GPT-5? | Pangram" description: "Pangram's AI detector successfully identifies GPT-5 generated content while competing detectors fail to do so." last_updated: "August 08, 2025" source: "https://www.pangram.com/blog/gpt-5" -------------------------------------------------------------------------------- Pangram Labs conducted testing of their AI detection capabilities against OpenAI's newly released GPT-5 model. Previous successful detection tests by Pangram included: - Reasoning models (OpenAI's o3, Gemini 2.5 Pro, Anthropic's Claude 4) - Meta's Llama 4 - GPT 4.5 - Claude Writing Styles - GPT 4o, Llama 3, and Claude 3 Initial testing used a creative writing prompt asking GPT-5 to compose an argument for why shrimp is superior to lobster. The resulting text was a detailed comparison covering taste, texture, cooking methods, versatility, cost, and dining experience. Testing Results: - Pangram's detector identified the GPT-5 generated text with high confidence - Competing detectors failed to identify the AI-generated content: - GPTZero - ZeroGPT - Grammarly - UndetectableAI - Originality.AI All incorrectly classified the text as human-written The test document was the first attempt at detection, demonstrating Pangram's immediate effectiveness without specific training for GPT-5. Testing continued to evaluate GPT-5's improved writing capabilities through prompts designed to test both reasoning and creativity. -------------------------------------------------------------------------------- title: "Useful AI policies for educators | Pangram" description: "Comprehensive AI usage policies for educational settings, including both blanket prohibition and tiered permission frameworks" last_updated: "January 11, 2025" source: "https://www.pangram.com/blog/useful-ai-policies" -------------------------------------------------------------------------------- # Core AI Policy Frameworks for Education Two recommended AI policy frameworks address student AI usage in educational settings: ## Blanket Prohibition Policy This policy completely prohibits AI use for assignments while allowing AI for concept exploration. Prohibited Activities: - Using AI to complete assignments faster - Using AI as an answer reference - Generating any assignment content with AI - Creating outlines or section headers with AI - Generating assignment drafts - Using AI for editing or rewriting - Using generative features in Grammarly or Google Docs Permitted Activities: - Exploring course material conversationally with AI - Seeking concept clarification through AI interaction ## Tiered Permission System A flexible framework allowing different levels of AI usage across assignments. Tiers are cumulative, with each including previous tier permissions: Tier 0 - Zero Assistance - Completely original work without AI tools - Used for raw ability assessment - Examples: pen-and-paper work, textbook homework, in-class exams - No AI detection triggers Tier 1 - Basic Tool Use - Basic grammar/spell checking and calculators - No AI models permitted - Examples: typed essays, basic Google Docs spell-check - No AI detection triggers Tier 2 - AI Learning Tool - AI permitted for material engagement but not final work - No AI-generated outlines or content - Examples: Perplexity/Google research, concept clarification via ChatGPT - No AI detection triggers Tier 3 - AI Editing Tool - Limited AI use for brainstorming, outlining, rephrasing - Student work must remain majority original - Examples: AI outline assistance, draft critique, Grammarly AI tools - May trigger AI detection Tier 4 - AI Collaboration - Substantial AI assistance for advanced projects - Not for basic learning tasks - Examples: AI paper summarization, analysis code generation, research discussions - May trigger AI detection Implementation Notes: - Tier 2 typically serves as default assignment level - Specific projects may allow higher tiers - Tiers 3-4 may trigger AI detection with heavy AI usage - System adapted from AI Assessment Scale (Perkins et al.) The policy frameworks aim to maintain academic integrity while preparing students for real-world AI integration and understanding AI limitations. -------------------------------------------------------------------------------- title: "Mirror, Mirror On The Wall, Who's The Realest Of Them All? | Pangram" description: "An educator's perspective on AI's impact on learning and the importance of maintaining authentic human voices in education" last_updated: "July 25, 2025" source: "https://www.pangram.com/blog/mirror-mirror-on-the-wall-who-s-the-realest-of-them-all" -------------------------------------------------------------------------------- AUTHOR: Jason Nicholson PUBLICATION DATE: July 25, 2025 ARTICLE CONTENT: Teaching Background: - Over 20 years teaching experience across Humanities fields in multiple school types - Career motivated by facilitating student "aha moments" - instances of deep, personal understanding - Transitioned from student to teacher to preserve and share learning experiences AI Impact on Education: ChatGPT introduction two years ago created immediate student adoption for: - Homework answer generation - Assignment writing - General academic assistance Key Educational Concerns: 1. AI's role as a knowledge source 2. AI's impact on learning processes Teaching Philosophy: - Learning remains fundamentally unchanged despite technological advances - Learning defined as an "experience of a concept" activity - Human brain function maintains consistent patterns despite technological progress Four Critical Questions Facing Education: 1. Mirror Metaphor and AI Detection: - Traditional teaching reflects teacher-student knowledge exchange - AI's accuracy makes distinguishing human vs AI work challenging - AI detection tools necessary but not sufficient alone - Detection tools should be viewed as transparency aids rather than pure detection mechanisms 2. Learning Verification Challenge: - Risk of students producing correct answers without underlying knowledge - Concern about writing output without actual writing ability development 3. Epistemic Fragility: - Authentic thinking crucial for genuine learning moments - Teacher inability to distinguish AI from student work compromises development - Classroom purpose centers on fostering genuine understanding 4. Human Voice Preservation: - Fundamental need for authentic human expression - AI cannot replicate genuine personal voice - Essential to maintain student writing skill development for self-expression and communication Core Conclusion: Authentic education requires genuine human voice and real breakthrough moments of understanding. PRODUCT OFFERINGS: Solutions: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual Capabilities Use Cases: - Teachers - Publishers - Content Moderation Resources: - AI Detection Model Documentation - Technical Report (arxiv.org/abs/2402.14873) - Data Privacy Guidelines - Educational Resources including: * State of AI Detection in 2025 * Most Common AI Phrases -------------------------------------------------------------------------------- title: "AI Conference Papers are Increasingly Being Written by AI: up 370% since 2023" description: Analysis of AI-generated content in academic conference papers showing dramatic increase since 2023 last_updated: "September 30, 2024" source: "https://www.pangram.com/blog/academic-papers" -------------------------------------------------------------------------------- In February 2024, a Frontiers in Cell and Developmental Biology article contained obviously AI-generated figures, including one showing a rat with abnormally large testicles and nonsensical text, highlighting growing concerns about AI-generated academic content. A Nature news report revealed that when Claude 3.5-generated research papers were presented to scientific reviewers, they received higher ratings for novelty, excitement, feasibility and expected effectiveness compared to human-written papers. However, analysis of 4,000 AI-generated papers showed only 5% (200 papers) contained original ideas, with most content being recombinations from training data. AI-generated research creates several problems: - Adds noise to peer review process - Wastes reviewer time and effort - Produces convincing but potentially erroneous content - Contains hallucinations and logical inconsistencies difficult for experts to detect ICML (International Conference on Machine Learning) Policy: - Prohibits text generated entirely by large language models (LLMs) - Allows LLM use for editing/polishing author-written text - Policy aims to prevent plagiarism and other LLM-related issues Research Methodology: - Analyzed conference submissions from 2018-2024 - Focused on ICLR and NeurIPS conferences - Used OpenReview API to extract submissions - Applied Pangram's AI Detector to analyze abstracts Key Findings: - Significant increase in AI-generated content since 2023 - Zero false positives for pre-2022 submissions (validation period before LLMs existed) - 370% increase in AI-generated abstracts since 2023 - Model shows high accuracy for scientific abstract analysis The study demonstrates widespread policy violations within the AI research community, with researchers increasingly using AI to generate conference paper content despite explicit prohibitions. [Note: The source text appears to cut off mid-sentence at the end, so the extraction ends at the last complete point] -------------------------------------------------------------------------------- title: "Can AI detection catch Claude writing styles? | Pangram" description: "Analysis of Anthropic Claude's writing style presets and custom styles, testing their detectability using Pangram's AI detection tools" last_updated: "December 6, 2024" source: "https://www.pangram.com/blog/claude-writing-styles" -------------------------------------------------------------------------------- Anthropic released an update to Claude.ai in November 2024 adding selectable writing style presets and custom style options. The preset styles include: - Concise: Responds primarily in bullet points and lists - Explanatory: Produces longer, more detailed responses - Formal: Omits Claude's typical conversational elements and maintains professional tone Testing conducted using 250-word essays on the fall of Rome revealed all preset styles maintained Claude's characteristic writing patterns. The outputs remained detectably AI-generated regardless of style selection. Custom Writing Styles: - Interface allows users to input sample writing - AI generates writing style instructions (approximately 3 sentences) - System creates AI-generated example texts - Two test styles were created: 1. "Tech Storyteller" (based on author's blog posts) 2. "Scholarly Skeptic" (based on Slate Star Codex posts) Detection Results: - All custom style outputs remained clearly identifiable as AI-written - Pangram's detection model successfully identified AI authorship across all variations - Manual editing of prompts to include source examples did not significantly improve human-like qualities - AI frequently misinterpreted style instructions, focusing on superficial elements like increased usage of specific words (e.g. "complex") Common AI Writing Markers Identified: - Phrase "complex and multifaceted" appears 700x more often in AI vs human writing - "Intricate interplay" appears 100x more often in AI writing - "Played a crucial role" appears 70x more often in AI writing Conclusion: Claude's writing style feature functions primarily as a convenience tool for output formatting but does not enable AI text to evade detection. Pangram's detection model successfully identifies AI-generated content across all of Claude's writing style variations through holistic document analysis. -------------------------------------------------------------------------------- title: "Terms of Service | Pangram" description: "Legal terms and conditions governing the use of Pangram Labs' AI content detection services and platform" last_updated: "February 19, 2025" source: "https://www.pangram.com/terms-of-service.html" -------------------------------------------------------------------------------- # Pangram Labs Terms of Service Pangram Labs provides technology solutions for detecting AI-generated content at scale, enabling authentication of human-generated content and prevention of unwanted AI-generated content. ## Agreement Binding Users agree to terms by clicking "I Accept" or by downloading, installing, accessing, or using the service. The agreement includes the Privacy Policy (https://www.pangram.com/privacy-policy.html). ## Arbitration Notice Disputes require binding individual arbitration; users waive rights to jury trials, class actions, or representative proceedings, with exceptions detailed in Section 16. ## Eligibility Requirements - Minimum age: 18 years - No previous suspension/removal from service - Compliance with applicable laws and regulations - For organizations: Accepting individual must have binding authority ## Account Registration - Required information: name, email address, contact details - Users must maintain accurate, complete, non-misleading information - Password security responsibility lies with users - Account security concerns should be reported to info@pangram.com ## Payment Terms - Fees in U.S. Dollars - Non-refundable except where legally required - Pricing subject to change with advance notice - Promotional offers may vary between customers - Payment methods require pre-authorization for credit cards ## Subscription Details - Automatic renewal based on initial subscription period - Billing occurs on Subscription Billing Date - Cancellation required before renewal date to avoid charges - Cancellation available through account settings or info@pangram.com - Subscription fees viewable at https://pangram.com/pricing ## Account Delinquency Policies - Service access suspension for unpaid amounts - Additional charges for chargebacks and collections - Account deletion possible for invalid payment methods at renewal - Collection fees apply to delinquent accounts ## License Terms - Limited, non-exclusive, non-transferable, non-sublicensable, revocable license - Internal business use only - Restrictions: No reproduction, distribution, public display, derivative works - Modifications prohibited - Security/access control circumvention prohibited -------------------------------------------------------------------------------- title: "30 Best AI Detectors Reviewed in 2025 (Free & Paid) | Pangram" description: "Comprehensive review and accuracy testing of 30 AI detection tools for identifying AI-generated content from platforms like ChatGPT, Claude, and Gemini." last_updated: "2025" source: "https://www.pangram.com/blog/best-ai-detector-tools" -------------------------------------------------------------------------------- AI detection tools are becoming increasingly important as AI writing tools like ChatGPT become more prevalent. This review examines the top AI detectors based on accuracy testing and features. Detailed Reviews of Top AI Detectors: 1. Pangram Labs - Free tier: twenty free checks per day with signup (up to 2,000 words daily) - Paid pricing: from $15/month (billed annually) for 300,000 words per month - Test results: - AI detection accuracy: 9/9 (100%) - Human text accuracy: 3/3 (100%) - Features: Supports 20+ languages, Chrome and Canvas extensions - Development: Created by AI researchers from Stanford, Tesla and Google - Updates: Maintained current with new AI model releases - False positive rate: Near-zero 2. Quillbot - Free tier: 1,200 words - Paid tier: Unlimited words at $19.95/month - Test results: - AI detection accuracy: 4/9 (44%) - Human text accuracy: 3/3 (100%) - Language support: English, French, German, Spanish, Dutch - Scoring categories: AI-generated, AI-generated & AI-refined, Human-written & AI-refined, Human-written - Performance notes: Successfully detected Gemini content but failed on Claude tests and 2/3 ChatGPT tests 3. Scribbr - Free tier: Unlimited checks with no signup required - Paid tier: None - Test results: - AI detection accuracy: 4/9 (44%) - Human text accuracy: 3/3 (100%) - Language support: English, French, German, Spanish, Dutch - Performance notes: Successful with Gemini content, 1/3 success with ChatGPT, 0/3 with Claude 4. GPTZero - Free tier: 10k words/month with basic scans + 5 advanced scans - Paid tier: $14.99/month for 150k words - Test results: - AI detection accuracy: 7/9 (78%) - Human text accuracy: 3/3 (100%) - Launch date: 2023 - Features: Basic scoring system, detailed feedback for paid users - Input methods: Text paste, file drag-and-drop, local file upload 5. ZeroGPT - Free tier: 15k characters per detection - Paid tier: $9.99/month for 100k characters and 50 batch files - Test results: - AI detection accuracy: 6/9 (67%) - Human text accuracy: 1/3 (33%) - Performance breakdown: - Gemini: 3/3 success - Claude/ChatGPT: 3/6 success - Human text: 1/3 success -------------------------------------------------------------------------------- title: "Our Model | Pangram" description: "Pangram Labs' AI text detection model achieves industry-leading accuracy in identifying AI-generated content across multiple domains and language models" last_updated: "August 08, 2025" source: "https://www.pangram.com/our-model/ai-detection" -------------------------------------------------------------------------------- # Pangram Labs AI Detection Technology ## Core Technology Pangram's text classifier analyzes documents to determine whether content is human-written or AI-generated, providing likelihood predictions for AI generation. The model was developed by researchers with experience building AI systems at Tesla and Google. ## Performance Metrics - 9x lower error rate compared to other leading AI text detectors - 99%+ accuracy across all text categories - Less than 1% false positive rate - 99%+ recall rate when evaluating outputs from leading language models - 10x fewer documents misclassified compared to competitors ## Capabilities - Detects output from multiple LLMs including ChatGPT, GPT-4, LLaMA, Bard, Claude, and Mistral - Identifies AI-generated text modified by paraphrasing tools (Quillbot, Undetectable AI) - Maintains accuracy with co-written text, minor edits, and artificial typos - Supports diverse content types including reviews, news, emails, and essays - Rapid adaptation to new LLMs as they become publicly available ## Domain Coverage The classifier is trained on and evaluated across multiple content categories: - Reviews - News articles - Emails - Essays - Blog posts - Other professional and academic writing ## Technical Documentation Detailed performance data and methodology available in the Technical White Paper published February 2024. ## Compliance and Security - SOC2 Type 2 certified (verified by AssuranceLab) - Cookie collection for website traffic and performance analysis ## Product Integration Options - Dashboard interface - Chrome extension - API access - Third-party integrations - Plagiarism detection capabilities - Multilingual support ## Use Case Support - Educational institutions - Publishers - Content moderation teams - Trust and safety operations Note: This document reflects the public content available on the specified page as of the last update date. For current performance metrics and capabilities, refer to Pangram's latest technical documentation. -------------------------------------------------------------------------------- title: "How accurate is Pangram AI Detection on ESL? | Pangram" description: "Analysis of Pangram's AI detection accuracy for ESL writing, comparing false positive rates across multiple datasets and competing solutions" last_updated: "April 23, 2025" source: "https://www.pangram.com/blog/how-accurate-is-pangram-ai-detection-on-esl" -------------------------------------------------------------------------------- Previous AI detectors based on perplexity and burstiness exhibit bias against ESL (English as Second Language) writers due to their limited vocabulary and complex sentence construction patterns, resulting in high false positive rates when detecting AI-generated content. Key Research Studies: - Stanford study (July 2023): Tested 7 AI detectors on 91 TOEFL essays; showed >60% false positive rate on human ESL writing - ETS study (August 2024): Analyzed 2,000 GRE writing samples from nonnative speakers; found reduced bias when ESL data was well-represented in training sets Pangram Performance Metrics: Dataset performance on ESL writing (all datasets held out during training): - ELLIPSE: 0.00% false positive rate (3,907 samples) - ICNALE: 0.018% false positive rate (5,600 samples) - PELIC: 0.045% false positive rate (15,423 samples) - Liang TOEFL: 0% false positive rate (91 samples) Overall false positive rate: 0.032% across 25,021 samples Comparative Analysis: TurnItIn comparison (300+ word documents): - L2 English (ESL): Pangram 0.02% vs TurnItIn 1.4% false positive rate - L1 English (Native): Pangram 0.00% vs TurnItIn 1.3% false positive rate GPTZero comparison on Liang TOEFL dataset: - GPTZero: 1.1% false positive rate + 6.6% "Possible AI content" rate - Pangram: 0% false positive rate with high confidence Pangram's ESL Accuracy Methodology: 1. Data Diversity: - Includes nonnative English text in training data - Incorporates casual/conversational English - Uses diverse sources: social media, reviews, general internet text - Includes foreign domain English content 2. Training Approach: - Broader spectrum beyond academic essays - Representation of informal writing patterns - Inclusion of imperfect writing samples - Focus on multilingual capabilities [Note: The content appears to be cut off mid-section in the source. This extraction includes all available content from the provided webpage.] -------------------------------------------------------------------------------- title: "Research Collaborations | Pangram" description: "Research collaboration program for non-commercial studies focused on AI-generated content detection and analysis" last_updated: "2024" source: "https://www.pangram.com/contact-us/research" -------------------------------------------------------------------------------- # Research Collaborations Pangram Labs offers research support for non-commercial studies focused on AI-generated content. The program is open to academic institutions, universities, and research organizations conducting studies related to AI content detection and analysis. ## Research Support Program Pangram Labs provides research collaboration opportunities for: - Academic institutions - Universities - Research organizations - Non-commercial research projects - Studies focused on AI-generated content ## Application Process Interested researchers must submit an inquiry form including: - Email address - Full name - Institutional affiliation - Detailed research inquiry description - Demo meeting preference (optional) ## Product Offerings Pangram's core solutions include: - Dashboard - Chrome Extension - API access - Third-party integrations - Plagiarism detection capabilities - Multilingual content analysis ## Technical Resources Available documentation and research materials: - AI detection model technical specifications - Detailed methodology documentation - Technical report (arxiv.org/abs/2402.14873) - Data privacy protocols - Implementation guides ## Educational Resources Published research and analysis: - The State of AI Detection in 2025 - Most Common AI Phrases Analysis ## Company Information Contact: info@pangram.com Certifications: SOC2 TYPE2 (Verified by AssuranceLab) Social: LinkedIn, Twitter, Instagram Community: Discord community available at discord.gg/f7jDAPzWH3 -------------------------------------------------------------------------------- title: "Statement on Biden's AI Safety Executive Order | Pangram" description: "Pangram Labs' official response supporting Biden's executive order on AI safety while advocating for specific approaches to AI detection and regulation" last_updated: "October 31, 2023" source: "https://www.pangram.com/blog/ai-safety-executive-order" -------------------------------------------------------------------------------- # Statement on Biden's AI Safety Executive Order Authors: Max Spero and Bradley Emi Published: October 31, 2023 Company Update: Checkfor.ai has rebranded to Pangram Labs. The Biden administration released new AI safety and security standards including directives for AI content detection. The Department of Commerce will develop guidance for content authentication and watermarking to label AI-generated content, with federal agencies implementing these tools to authenticate government communications. Pangram Labs (formerly Checkfor.ai) fully supports the administration's commitment to safe LLM deployment while maintaining specific positions on AI content authentication: Key Position Statements: 1. Current detection solutions are insufficient; more investment needed specifically in AI detection 2. Watermarking is inadequate due to: - Creates false security - Can be circumvented when model weights are accessible - Cannot address unwatermarked open-source models - Local fine-tuning enables detection evasion 3. Regulation Requirements: - Must support open source ecosystem - Open source provides: - Checks and balances on tech companies - Consumer transparency - Public evaluation capabilities - Model benchmarking opportunities 4. Government Role Recommendations: - Focus on funding academic AI detection research - Support industry projects in detection - Establish detection benchmarks - Create evaluation criteria - Enable consumer awareness of detector limitations 5. Industry Standards: - Need established benchmarks - Require evaluation criteria - Essential for safe deployment of powerful language models Pangram Labs commits to developing reliable AI detection systems and industry-wide standards to enable safe deployment of next-generation language models. The company expresses support for the Biden administration's focus on distinguishing human and AI-generated content and offers to collaborate with researchers and policymakers on standards development. Company Information: - Name: Pangram Labs (formerly Checkfor.ai) - Focus: AI detection systems and content authentication - Mission: Protecting internet from spam and malicious AI-generated content - Commitment: Building reliable AI detection systems for safe LLM deployment -------------------------------------------------------------------------------- title: "How well can Pangram detect reasoning models? | Pangram" description: "Analysis of Pangram's AI detection capabilities for reasoning-based language models, including performance metrics and adaptation strategies" last_updated: "July 16, 2025" source: "https://www.pangram.com/blog/how-well-can-pangram-detect-reasoning-models" -------------------------------------------------------------------------------- Reasoning models represent a significant breakthrough in large language model technology in 2025, characterized by their ability to produce thinking tokens before generating output. REASONING MODEL FUNDAMENTALS: - Reasoning models combine normal LLM capabilities with thinking/reasoning tokens - Models excel at problem-solving, particularly in math and coding - Performance exceeds standard benchmarks through pre-response reasoning OPERATIONAL MECHANISM: - Implements "chain of thought" processing before generating responses - Example: Deepseek-R1 publicly exposes its thinking process - Models analyze user intent before token generation for improved logical organization MAJOR REASONING MODELS: OpenAI O-Series: - Available models: o1, o1-mini, o3, o3-pro, o4-mini - o3-pro identified as most capable in series Anthropic Models: - Claude 4 Opus and Claude 4 Sonnet feature "extended thinking" mode - Both incorporate pre-response reasoning capabilities Google Models: - Gemini 2.5 series includes internal thinking processes - Variants: Gemini 2.5 Pro, Flash, and Flash-Lite Open Source Models: - Deepseek R1: First open-source reasoning model; allows visibility of thinking process - Qwen-QWQ-32B: Smaller, more versatile deployment model PANGRAM DETECTION PERFORMANCE: Performance Metrics (Old vs July Release): - OpenAI o1: 99.86% → 100% - OpenAI o1-mini: 100% → 100% - OpenAI o3: 93.4% → 99.86% - OpenAI o3-pro: 93.9% → 99.97% - OpenAI o3-mini: 100% → 100% - OpenAI o4-mini: 99.64% → 99.91% - Gemini 2.5 Pro Thinking: 99.72% → 99.91% - Claude Opus 4: 99.89% → 99.94% - Claude Sonnet 4: 99.89% → 99.91% - Deepseek-R1: 100% → 100% - Qwen-QWQ-32b: 100% → 100% O3 AND O3-PRO OPTIMIZATION: - Challenges: Higher costs and longer processing times - Training data composition: o3 (0.17%), o3-pro (0.35%), o3-mini (5%) - Achieved improved detection without increasing false positives FEW-SHOT LEARNING CAPABILITIES: - Pangram adapts to new LLMs with minimal training data - Effective generalization from small sample sizes - Quick adaptation to fine-tuned LLM variants - Learning efficiency increases with exposure to multiple LLMs KEY ADVANTAGES: - Rapid adaptation to new LLM releases - Efficient pattern recognition from limited examples - LLM capability improvements do not diminish detectability - Each LLM maintains distinct, identifiable characteristics -------------------------------------------------------------------------------- title: "How Quora uses Pangram to handle AI-written answers | Pangram" description: "Case study detailing how Quora implemented Pangram's AI detection system to identify and manage AI-generated content across their platform" last_updated: "September 26, 2024" source: "https://www.pangram.com/blog/quora-case-study" -------------------------------------------------------------------------------- # Quora's Implementation of Pangram AI Detection ## Overview Pangram Labs partnered with Quora in April 2024 to combat spam from ChatGPT-generated answers. Quora ranks as the 33rd most trafficked website globally, receiving over 1 billion monthly page visits as of August 2024. ## Platform Context Quora's mission focuses on sharing and growing world knowledge by connecting questions to people with relevant expertise. According to Lexie Wu, Group Product Manager leading moderation: "A lot of knowledge is stuck in people's heads, and if we match the right questions to the right people we can extract that knowledge." ## AI Content Challenges AI-generated answers create three primary issues: 1. Deterrence of authentic participation: Existing AI answers discourage users from sharing personal experiences 2. Content displacement: AI posts compete with authentic content for limited user engagement 3. Platform reputation risk: Users can identify AI content, potentially questioning Quora's value proposition versus direct LLM use ## Technical Implementation Pangram was selected because: - Provides reliable detection of advanced models like GPT-4 - Demonstrates 100x greater accuracy compared to competitors like GPTZero - Maintains robustness through rapid adaptation to new LLMs within 24 hours - Expanded to support 20+ languages as of July 2024 ## Results and Impact As of September 2024: - Identified over 1 million AI-generated posts - Improved overall content quality - Maintained platform authenticity and trustworthiness - Enhanced Trust & Safety team capabilities for AI content policy enforcement ## Technical Capabilities Pangram offers: - Automated detection to replace manual review - Support for current and emerging LLM detection - Built-in data pipeline for rapid model updates - Multi-language content analysis - Integration with moderation workflows Contact: info@pangram.com -------------------------------------------------------------------------------- title: "Pangram Text Update: GPT-4o, Claude 3, LLaMA 3" description: Pangram Labs announces improved AI detection accuracy for latest language models including GPT-4o, Claude 3, and LLaMA 3 last_updated: "May 22, 2024" source: "https://www.pangram.com/blog/updating-pangram-text-gpt4o" -------------------------------------------------------------------------------- CORE UPDATE DETAILS: - New model version achieves near-perfect accuracy in detecting AI-written text from GPT-4o, Claude 3, and LLaMA 3 - Infrastructure pipeline enables rapid ingestion of AI text from new models upon public release - Convergence observed in stylistic similarities among high-performance models ACCURACY IMPROVEMENTS: Model accuracies improved across all tested LLMs: - Overall: 99.54% → 99.84% (+0.30%) - GPT-4o: 99.78% → 100% (+0.22%) - Claude 3: 99.12% → 99.76% (+0.64%) - LLaMA 3: 99.58% → 99.97% (+0.39%) EVALUATION METHODOLOGY: - 25,000 examples in evaluation set - 40% consists of AI-generated text from new models - Text spans multiple domains: news, reviews, education - Includes all versions of Claude 3 (Opus, Sonnet, Haiku) - Testing revealed improvements for older models: 8 previously failed GPT-3.5 cases now passing; 13 previously failed GPT-4 cases now passing IMPLEMENTATION TIMELINE: May 13: GPT-4o API release May 14: Dataset pipeline update and new training/evaluation sets created May 15-16: AI detection model training May 17: QA/sanity checks and model release TECHNICAL INSIGHTS: - System architecture designed for rapid adaptation to new LLMs - Dataset regeneration and model training possible within hours of new API availability - Uses Hard Negative Mining with Synthetic Mirrors technique (detailed in technical report) - More capable models show distinct idiosyncratic styles making them easier to detect - Claude Opus detection rates higher than Sonnet and Haiku versions MARKET OBSERVATIONS: - Foundation models converging toward GPT-4 performance levels - Similar attention-based architectures and internet-scale training create comparable linguistic patterns - LLMs produce predictable, high-probability completions rather than original content - Creative writing tasks (essays, reviews, stories) reveal limitations in originality and authenticity NAVIGATION/RESOURCES: Products: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual Use Cases: - Teachers - Publishers - Content Moderation Resources: - AI Detection Model - Technical Report - Data Privacy - State of AI Detection in 2025 - Most Common AI Phrases - Assessment in AI Age - Academic Integrity Approach -------------------------------------------------------------------------------- title: "Comprehensive Guide to Spotting AI Writing Patterns | Pangram" description: "A comprehensive catalog of words, phrases, and patterns commonly used in AI-generated text that can help identify artificial writing." last_updated: "April 2, 2025" source: "https://www.pangram.com/blog/comprehensive-guide-to-spotting-ai-writing-patterns" -------------------------------------------------------------------------------- AI writing exhibits distinct patterns through overuse of specific vocabulary. The following comprehensive lists document words and phrases that appear with notably higher frequency in AI-generated text compared to human writing: ## Nouns Common AI-generated nouns include: - Abstract concepts: aim, aspect, complexity, depth, dynamics, enlightenment - Action-oriented: endeavor, exploration, innovation, journey, quest - Descriptive: elegance, nuance, poignancy, resonance - Environmental: climate, landscape, realm - Structural: component, framework, roadmap, toolkit - Transformative: revolution, transcendence, versatility [Full list includes 100+ nouns with specific categories] ## Verbs Characteristic AI verbs include: - Analytical: consider, elucidate, exemplify, unravel - Developmental: craft, curate, foster, innovate - Emotional: embrace, inspire, resonate - Progressive: deepen, elevate, enhance - Transformative: revolutionize, transcend, vitalize [Full list includes 75+ verbs with specific usage patterns] ## Adjectives Common AI adjectives include: - Qualitative: authentic, complex, dynamic - Evaluative: crucial, essential, invaluable - Descriptive: meticulous, nuanced, vibrant [Full list includes 40+ adjectives with contexts] ## Adverbs Frequently used AI adverbs include: - Manner: creatively, critically, meticulously - Degree: merely, profoundly, significantly - Temporal: timelessly, tirelessly [Full list includes 30+ adverbs with usage patterns] ## Common Phrases AI writing frequently employs phrases like: - Transitional: "additionally, we", "as we explore the topic" - Emotional: "feel a sense", "heart pounding" - Professional: "best regards", "hope this email finds you well" - Analytical: "findings suggest", "further research" -------------------------------------------------------------------------------- title: "How to collect evidence for an AI academic integrity case | Pangram" description: "A comprehensive guide for educators on collecting and documenting evidence when suspecting AI-generated academic work" last_updated: "March 13, 2025" source: "https://www.pangram.com/blog/how-to-create-evidence-for-an-ai-detection-case" -------------------------------------------------------------------------------- # Evidence Collection Methods for AI Academic Integrity Cases ## Core Principles AI detection scores alone cannot justify punitive action against students; a holistic approach combining multiple evidence types is required for academic integrity cases. ## Evidence Collection Strategies ### 1. Textual Evidence Analysis AI-generated content exhibits multiple weak signals that collectively form strong evidence. Key indicators include: Common AI Phrases: - "In today's technological era" - "Weaves details intricately together" - "Rich tapestry of perspectives" - "In conclusion" or "Overall" as closing phrases - Perfect grammatical structure - Even, consistent paragraph structure Comprehensive phrase analysis can be automated through Pangram's AI phrase detection feature, which identifies frequency patterns of AI-typical language. ### 2. Writing Style Characteristics Student vs AI Writing Distinctions (per Amanda Clarke's guide): Student Writing: - Contains specific evidence and direct textual citations - Includes genuine personal reflection - Shows natural variation in grammar and voice - Demonstrates authentic student perspective AI Writing: - Relies on vague generalities - May fabricate details - Uses consistently perfect grammar - Maintains neutral, impersonal tone - Lacks authentic reflection - Avoids specific textual evidence Mixed AI/human content often shows abrupt shifts in tone and writing style. ### 3. Process Documentation Legitimate student work generates evidence of the writing process: - Brainstorming notes - Outlines - Draft versions - Revision history - Proofreading marks Tools for Process Verification: - Draftback (Chrome extension): Replays Google Docs writing history - Brisk Teaching: Process monitoring - Cursive Technologies: Writing verification - Visible AI: Process tracking Contributors: - Marilyn Derby (Associate Director, Student Support and Judicial Affairs, UC Davis) - Amanda Clarke (English Department Chair, Viewpoint School, California) -------------------------------------------------------------------------------- title: "Enterprise | Pangram" description: "Enterprise-level AI content detection solution with customized implementations and consultation services" last_updated: "2024" source: "https://www.pangram.com/enterprise" -------------------------------------------------------------------------------- # Enterprise Solutions Pangram offers production-ready AI content detection solutions for enterprise clients, providing customized implementations based on specific organizational needs. Enterprise services include solution scoping, tailored implementations, and consultation services. ## Core Product Offerings ### Solutions - Dashboard for content analysis - Chrome Extension for browser-based detection - API for system integration - Third-party platform integrations - Plagiarism detection capabilities - Multilingual content analysis support ### Industry-Specific Use Cases - Educational institution implementations - Publishing industry solutions - Content moderation and trust & safety systems ## Technical Infrastructure - SOC2 Type 2 certified (verified by AssuranceLab) - Data privacy compliance frameworks - Technical documentation available via arxiv.org/abs/2402.14873 ## Educational Resources - State of AI Detection in 2025 report - Common AI phrase analysis - Assessment methodology in AI era - Academic integrity frameworks - Comprehensive technical documentation ## Enterprise Contact Process Enterprise inquiries require: - Professional email contact - Organization affiliation - Detailed project requirements - Optional demo scheduling available ## Support Channels - Direct email: info@pangram.com - Community access via Discord - Social media presence on Instagram, Twitter, and LinkedIn - Status monitoring: status.pangram.com ## Legal Framework - Terms of Service - Privacy Policy - Data Privacy FAQ - SOC2 TYPE2 compliance certification Enterprise implementations include access to all platform features, dedicated support, and customized solutions based on organizational requirements. Consultation services help scope detection needs and provide tailored implementation strategies. -------------------------------------------------------------------------------- title: "Meta will start identifying AI generated content | Pangram" description: "Meta announces plans to label AI-generated content across its platforms and implement new detection tools in response to EU regulations" last_updated: "May 14, 2024" source: "https://www.pangram.com/blog/meta-identifying-ai-content" -------------------------------------------------------------------------------- Meta has announced plans to implement AI content labeling and detection across its platforms. The initiative includes: CORE ANNOUNCEMENTS: - Development of internal tools to identify AI-generated content across Facebook, Instagram, and Threads - Labeling of AI-generated content from major providers including Google, OpenAI, Microsoft, Adobe, and Midjourney - Implementation of voluntary disclosure systems for users uploading AI content - Penalties for accounts failing to disclose AI-generated content RESEARCH AND CONSULTATION: - Conducted consultations with international policymakers - Surveyed over 23,000 users - Found 82% of users support AI content disclosure requirements - Special emphasis on content showing people saying things they did not say REGULATORY CONTEXT: - Implementation aligned with upcoming EU AI Act enforcement in May - Response to new European AI regulations - Proactive approach to platform integrity ahead of 2024 elections - Learning implementation from lessons of 2016 and 2020 US elections INDUSTRY IMPACT: Meta's President of Global Affairs identified AI content detection as "the most urgent task" facing the tech industry during a World Economic Forum panel. RECOMMENDED ACTIONS FOR COMPANIES: 1. Establish content policies compliant with regulations for AI content identification and adjudication 2. Develop features enabling user disclosure of AI-generated content 3. Implement flagging systems to mitigate AI content spread 4. Utilize AI detection products (such as Pangram Labs) for proactive content review CONTACT INFORMATION: Companies seeking platform integrity and regulatory compliance solutions can contact Pangram Labs at info@pangram.com -------------------------------------------------------------------------------- title: "How to detect AI in Google Docs | Pangram" description: "A comprehensive guide to detecting AI-generated content in Google Docs using Pangram's Chrome extension and dashboard tools" last_updated: "January 31, 2025" source: "https://www.pangram.com/blog/how-to-detect-ai-in-google-docs" -------------------------------------------------------------------------------- # AI Detection Methods for Google Docs Pangram Labs provides multiple methods for detecting AI-generated content in Google Docs: ## Chrome Extension Method 1. Download Pangram AI Detection from Chrome Web Store 2. Sign in or create Pangram account 3. Open Google Doc to analyze 4. Click "Scan Document for AI" button in top right corner 5. View results in bottom-right pop-up 6. Access full analysis via dashboard link 7. Share results using dedicated share button ## Clipboard Scanning Method 1. Select desired text section 2. Copy text (⌘+C or Ctrl+C) 3. Click Pangram extension icon in browser 4. Select "Scan Clipboard for AI" 5. View analysis of selected section ## Dashboard Direct Upload Method 1. Export Google Doc as .docx (File -> Download -> Microsoft Word) 2. Visit Pangram dashboard 3. Click File Upload 4. Drag and drop .docx file 5. Click "Check for AI" to initiate scan ## Version History Analysis Google Docs automatically saves document snapshots with timestamps, enabling: - Examination of writing process - Detection of bulk text insertions - Timeline analysis of document development - Identification of sudden large content additions Additional Features: - Section-by-section AI analysis - Phrase highlighting - Shareable result links - Support for multiple file formats - Integration with Google Docs interface - Real-time scanning capabilities The tool provides comprehensive AI detection directly within Google Docs workflow, requiring minimal setup and offering immediate results through multiple analysis methods. -------------------------------------------------------------------------------- title: "Making Your Business LLM And GenAI Proof | Pangram" description: "Guide for businesses to develop GenAI policies and implement AI safety guardrails while maintaining human oversight and regulatory compliance" last_updated: "January 30, 2024" source: "https://www.pangram.com/blog/make-your-business-llm-and-genai-proof" -------------------------------------------------------------------------------- Authors: Max Spero and Theodoros Evgeniou AI and Large Language Models (LLMs) emerged as transformative technologies in 2023, bringing both opportunities and risks for online platforms. Key challenges include managing AI-generated content floods, ensuring user safety, maintaining platform reputation, and addressing potential misuse by content generation startups. Core Business Policy Requirements: Organizations must develop GenAI policies based on two key questions: 1. User desire for AI-generated content 2. Acceptance of mixed AI/human content Policy Implementation Options: - For platforms rejecting AI content: Require AI disclosure or prohibit AI content entirely; enforce through human moderation and tools like Pangram Labs - For platforms accepting AI content: Implement safety guardrails and moderation processes similar to user-generated content using platforms like Tremau Context-Specific Considerations: - Marketplaces/review platforms must prevent AI-generated reviews - All platforms must block illegal AI-generated content - Enhanced bot/spam detection needed due to GenAI capabilities Available AI Guardrail Systems: Commercial API Safeguards: 1. Google Gemini API - Safety categories: Hate Speech, Harassment, Sexually Explicit, Dangerous Content - Automatic output rating system - Query rejection for high-risk content 2. Azure OpenAI API - Content filters: Hate and Fairness, Sexual, Violence, Self-Harm - Rejection of high-risk queries - Customizable intermediate safety moderation Open Source Options: - Models: Llama-2, Mistral - Required additional content filtering through: - OpenAI content filter API - Azure AI content safety API - Meta's LlamaGuard (7B-parameter model for prompt/response classification) Human Oversight Requirements: 1. Mandatory Human Review - AI tools cannot provide complete protection - Human reviewers must check flagged content - Efficient content review processes needed - Increased AI tool availability may require more human oversight 2. Process Requirements - Content moderation must be effective and efficient - Human involvement essential for error checking - Regulatory compliance necessary for all processes -------------------------------------------------------------------------------- title: "The Latest in AI Detection Research | Pangram" description: "Overview of recent academic studies validating Pangram's AI detection capabilities and research contributions" last_updated: "March 4, 2025" source: "https://www.pangram.com/blog/the-latest-in-ai-detection-research" -------------------------------------------------------------------------------- # Research Studies on AI Detection Capabilities ## University of Maryland Study: Human Detection of AI Text - Study examined human ability to detect AI-generated content across 300 non-fiction articles - Frequent LLM users demonstrated strong detection capabilities without training - Pangram's Humanizer model achieved 100% detection rate of AI-generated text - Both Pangram models maintained 90% detection rate against paraphrasing and humanizing attempts - Study published at: arxiv.org/abs/2501.15654 ## University of Pennsylvania Cross-Domain Detection Challenge - Research focused on detector generalization across AI models, document types, and adversarial attacks - Pangram tied for first place with Leidos research team - Study demonstrated detectors can effectively identify AI text across multiple domains simultaneously - Full details published at: arxiv.org/abs/2501.08913 ## ESPERANTO Back-Translation Study - Examined effectiveness of back-translation as an evasion technique - Back-translation involves translating text through multiple languages before returning to English - Most detectors showed significant reduction in detection rates - Pangram maintained highest robustness against back-translation compared to competitors - Research published at: arxiv.org/abs/2409.14285 ## Pangram Internal Research Published studies: - DAMAGE: Detecting Adversarially Modified AI Generated Text (arxiv.org/abs/2501.03437) - Technical Report on the Pangram AI-Generated Text Classifier (arxiv.org/abs/2402.14873) ## Academic Access - Pangram provides free unlimited access for academic researchers - Contact: info@pangram.com for research collaboration opportunities # Product Information Solutions offered: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities Use cases: - Education (Teachers) - Publishing industry - Content moderation/Trust and safety Resources available: - AI detection model documentation - Technical implementation guides - Data privacy information - State of AI Detection in 2025 report - Common AI phrases analysis - Academic integrity guidelines - Assessment strategies for AI era -------------------------------------------------------------------------------- title: "What to do when a student submission is flagged as AI | Pangram" description: "A comprehensive guide for educators on how to handle and investigate student work that has been flagged as potentially AI-generated" last_updated: "February 14, 2025" source: "https://www.pangram.com/blog/what-to-do-when-student-submits-ai" -------------------------------------------------------------------------------- # Understanding AI Detection Results Pangram's AI detection system provides confidence scores indicating the likelihood of AI-generated content. A 99% AI score indicates high confidence that at least some portion of the text was AI-generated, not necessarily the entire document. For longer submissions, the text is analyzed in segments to identify specific AI-generated sections. # Investigation Process ## Initial Student Discussion Teachers should engage students in conversation about their writing process. Students may admit to using AI due to time constraints or attempting to improve a draft. These discussions create opportunities to clarify academic integrity policies and proper procedures for handling assignment challenges. ## Common AI Tool Usage Scenarios Several legitimate tools may trigger AI detection: - Grammarly's AI-assisted writing features - AI-powered translation tools - Google Docs "Help me write" function - ChatGPT for brainstorming and research - AI-based wording assistance ## Writing Process Verification Key verification methods include: - Reviewing research notes and documentation - Examining early drafts - Analyzing Google Docs version history (File -> Version history -> See version history) - Looking for natural writing progression versus bulk text insertion - Checking typing patterns and revision history ## Assessment Considerations Multiple factors influence how to handle AI detection flags: - Student's previous history with AI detection - Available evidence of original work - Stakes of the assignment - Presence of writing process artifacts - Pattern of detections across multiple assignments # Best Practices ## Policy Implementation Implementation of a tiered AI policy system helps prevent misunderstandings between teachers and students regarding permitted tools. Clear guidelines should specify which AI-assisted tools are allowed and in what capacity. ## Detection Response Protocol AI detection should be treated similarly to metal detectors - as a screening tool requiring further investigation rather than immediate punitive action. The system's nonzero false positive rate necessitates thorough investigation before taking disciplinary measures. ## Documentation Requirements Students should maintain: - Research notes - Draft versions - Writing process documentation - Version histories when using digital tools # Additional Resources Pangram offers: - AI detection dashboard - Chrome extension - API access - Integration capabilities - Plagiarism detection - Multilingual support - Technical documentation - Educational resources on AI detection and academic integrity The complete technical report on Pangram's AI detection model is available at arxiv.org/abs/2402.14873. -------------------------------------------------------------------------------- title: "Pangram is the only AI detector that outperforms human experts at identifying AI content | Pangram" description: Research demonstrates Pangram's AI detection system outperforms both human experts and other automated detectors in identifying AI-generated content. last_updated: "January 29, 2025" source: "https://www.pangram.com/blog/russell" -------------------------------------------------------------------------------- New research by Jenna Russell, Marzena Karpinksa, and Mohit Iyyer from the University of Maryland and Microsoft demonstrates Pangram's superior accuracy in AI content detection, establishing it as the only system surpassing trained human experts' capabilities. Key Research Findings: 1. Human Detection Capabilities: - Non-expert humans perform at random chance levels for AI content detection - Trained expert annotators achieve over 90% true positive detection rate - Five Upwork annotators were trained specifically for the study 2. Expert vs. Non-Expert Analysis: Non-experts demonstrate common misconceptions: - Incorrectly associate "fancy" vocabulary with AI-generated content - Mistakenly attribute grammatical correctness to human authors - Wrongly assume neutral tone indicates AI authorship Experts recognize: - Specific overused AI phrases (e.g., "testament," "crucial") - AI's tendency to use clichéd, metaphorical vocabulary - More accurate pattern recognition in AI-generated content 3. Technical Performance: Pangram's detection capabilities: - 100% accuracy in detecting OpenAI's o1-pro outputs - 96.7% accuracy with "humanized" o1-pro outputs - Significantly outperforms other automated detectors, which achieve maximum 76.7% accuracy on base o1-pro outputs Common AI Writing Patterns: - Frequent use of phrases like "I recently had the pleasure of" in reviews - Common story openings such as "In the year of" in sci-fi content - Tendency toward clichéd expressions rather than sophisticated vocabulary Research Implications: - Demonstrates significant advancement in automated AI detection capabilities - Establishes new benchmarks for AI content detection accuracy - Contributes to explainability and interpretability in AI detection systems The research paper is available at arxiv.org/abs/2501.15654. [Note: The source content appears to be truncated mid-sentence in the final paragraph about Pangram's data pipeline and scale capabilities] Products and Solutions Listed: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities Use Cases: - Teachers - Publishers - Content Moderation Educational Resources: - The State of AI Detection in 2025 - Most common AI phrases -------------------------------------------------------------------------------- title: "AI Detector for Trust & Safety Teams | Pangram" description: "Enterprise AI detection system for content moderation and trust/safety teams to identify AI-generated content at scale" last_updated: "2024" source: "https://www.pangram.com/use-cases/trust-and-safety" -------------------------------------------------------------------------------- # Core Offering Pangram Labs provides enterprise-grade AI detection capabilities through an API for trust and safety teams. The system identifies AI-generated content with 99.99% accuracy and maintains the lowest false positive rate in the market. # Key Features - Detection of AI content as short as 75 words - Batch scanning capability for multiple content pieces - API access for enterprise integration - Support for short-form content including social media posts, DMs, and comments - Cost efficiency: 100x cheaper than manual content moderation # Technology Background Developed by AI researchers from Tesla and Google, Pangram's detection system: - Uses proprietary technology that actively learns from shortcomings - Self-improves when detecting AI content - Outperforms human experts at identifying AI content - Employs novel research methods for superior reliability - Technical documentation available via arxiv.org/abs/2402.14873 # Trust & Safety Applications ## Regulatory Compliance The Federal Trade Commission has banned fake reviews, including AI-generated reviews. AI-written reviews present specific challenges: - Mislead ranking systems due to grammatically correct, lengthy content - Erode platform trust - Damage platform reputation - Cannot be copyright protected (U.S. Copyright Office ruling, March 2023) ## Content Moderation Benefits - Automates workflow for time-constrained safety teams - Reduces operational costs - Supports scalable content moderation - Enables efficient policy enforcement - Preserves authentic human-generated content # Enterprise Integration Pangram offers multiple implementation options: - Dashboard interface - Chrome Extension - API access - Platform integrations - Multilingual support - Plagiarism detection capabilities # Partner Organizations Current trusted partners include: - Canvas - Google Classroom - Quora - Tremau - The Transparency Company - Newsguard # Founder's Statement Max, Co-Founder of Pangram Labs, brings experience from Google and Yelp in fraudulent content detection. The company prioritizes: - Preventing AI misuse in content creation - Maintaining content authenticity - Supporting trust and safety initiatives - Enabling effective content moderation policies - Providing cost-effective solutions for enterprise teams Note: This document reflects the content as of the last update and may not include subsequent changes or updates to the service. -------------------------------------------------------------------------------- title: "Checkfor.ai is now Pangram Labs | Pangram" description: "Checkfor.ai announces rebranding to Pangram Labs and reaffirms its mission to detect AI-generated content across the internet" last_updated: "April 8, 2024" source: "https://www.pangram.com/blog/checkforai-is-now-pangram-labs" -------------------------------------------------------------------------------- # Company Rebranding Announcement Checkfor.ai has rebranded to Pangram Labs as of April 8, 2024. The original name Checkfor.ai was chosen six months prior to reflect the company's AI detection value proposition. The name "Pangram" was selected because it represents a sentence or text that uses all letters of the alphabet, serving as a metaphor for the company's approach to machine learning models that analyze all aspects of human language to differentiate between human and AI-generated content. # Core Business Focus Pangram Labs maintains its primary mission of keeping the internet free of AI-generated spam. The rebranding does not affect any operational or technical aspects of the service, with changes limited to company communications. # Product Offerings Current solutions include: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities # Use Case Categories - Education (Teachers) - Publishing industry - Content Moderation/Trust and Safety # Technical Resources - AI detection model documentation - Technical implementation guides - Published technical report (arxiv.org/abs/2402.14873) - Data privacy documentation # Educational Resources - The State of AI Detection in 2025 - Most common AI phrases analysis # Company Information Contact: info@pangram.com Certifications: SOC2 TYPE2 (Verified by AssuranceLab) Social Presence: Instagram, Twitter, LinkedIn, Discord community -------------------------------------------------------------------------------- title: "Most common AI phrases | Pangram" description: "Comprehensive catalog of words and phrases commonly overused in AI-generated text compared to human writing" last_updated: "August 08, 2025" source: "https://www.pangram.com/resources/most-common-ai-phrases" -------------------------------------------------------------------------------- AI language exhibits distinct patterns of word and phrase usage that differ from human writing. The following lists document words and phrases that appear with notably higher frequency in AI-generated text: NOUNS: Commonly overused nouns include aim, aspect, challenges, climate, community, complexity, component, development, dynamics, exploration, facet, insight, journey, landscape, realm, significance, and transcendence. Additional nouns focus on abstract concepts: enlightenment, innovation, resonance, testament, versatility. VERBS: Characteristic AI verbs emphasize actions like capturing, confronting, crafting, curating, deepening, elevating, elucidating, embodying, embracing, enhancing, exploring, fostering, illuminating, innovating, navigating, revolutionizing, and transcending. ADJECTIVES: AI text frequently employs descriptive terms including authentic, commendable, complex, crucial, dynamic, elusive, innovative, meticulous, notable, nuanced, significant, and sustainable. ADVERBS: Common AI adverbs include additionally, critically, crucially, dynamically, insightfully, intricately, meticulously, pivotally, profoundly, seamlessly, and vibrantly. PHRASES: Distinctive AI phraseology includes: - "about the potential" - "additionally, we" - "as an ai" - "as we [verb] the topic" - "commitment to" - "deeper understanding" - "findings suggest" - "for greater" - "future generations" - "highlights the importance" - "implications for" The lists provided are non-exhaustive but represent a large sample of AI writing patterns that can help identify machine-generated content. -------------------------------------------------------------------------------- title: "Tremau and Pangram Labs partner to take on AI-generated content" description: "Partnership announcement between Tremau and Pangram Labs to combat AI-generated content risks during major global elections in 2024" last_updated: "April 17, 2024" source: "https://www.pangram.com/blog/tremau-and-pangram-labs" -------------------------------------------------------------------------------- 2024 marks the largest election year in history, with over 50 countries representing 4.2 billion people holding national and regional elections, including seven of the ten most populous nations globally. Generative AI poses significant threats to electoral integrity through: - Deepfake videos - Large-scale targeted AI-generated campaigns - Last-minute voter deterrence attempts - Generated candidate depictions difficult to verify - Targeted false story distribution Partnership Details: Pangram Labs and Tremau have partnered to provide AI-generated content detection and disclosure solutions for user-generated content. Pangram Labs develops AI content detection methods while Tremau enables human-in-the-loop technologies for content moderation. Regulatory Context: The EU AI Act, initially proposed in April 2021, has been substantially amended to address generative AI concerns. Key requirements include: - Classification of generative AI under "General Purpose AI Systems" - Mandatory disclosure of AI-generated content - Transparency requirements for AI systems Business Impact: Organizations must address AI content challenges beyond regulatory compliance, including: - Brand safety protection - Platform health maintenance - Prevention of spam and disinformation - Authentication of user-generated content Available Solutions: - Pangram Labs: Automated detection tools for AI-generated text and speech - Tremau: Content moderation platform with regulatory compliance features - Combined offering: AI content detection integrated with human moderation capabilities Contact Information: - Pangram Labs: info@pangram.com - Tremau: info@tremau.com Authors: Max Spero and Tremau Note: This partnership aims to build a safe and beneficial digital world by helping platforms maintain authentic content and regulatory compliance through combined AI detection and moderation capabilities. -------------------------------------------------------------------------------- title: "Why does Pangram have a minimum word count? | Pangram" description: "Pangram's AI detection model requires 50+ words for accurate predictions because context is essential for distinguishing between AI and human writing patterns." last_updated: "May 23, 2025" source: "https://www.pangram.com/blog/why-does-pangram-have-a-minimum-word-count" -------------------------------------------------------------------------------- Pangram enforces a 75-word minimum requirement for its AI detection model to ensure accurate predictions. While individual words may show statistical patterns (like "delve" appearing 15x more frequently in AI writing), context is crucial for accurate detection. Example Analysis: 1. Human-written sample (detected as human): "I am actually super excited to read the Great Gatsby! I've been told that the Great Gatsby is one of the most popular American books of all time, and for that reason, I'm actually very interested in the opportunity to delve into this novel. My parents, teachers and friends have raved about it, and I trust them a lot!" 2. AI-written sample (detected as AI): "I'm excited to read The Great Gatsby because it offers a chance to delve into the glamour and disillusionment of the Roaring Twenties. I'm curious about how Fitzgerald portrays ambition, love, and the American Dream through such iconic characters. I'm especially eager to experience the lyrical writing and uncover the deeper meanings behind Gatsby's mysterious life." Context Analysis: - Both samples use the word "delve" - Similar phrases around "delve": - Human: "opportunity to delve into this" - AI: "a chance to delve into the" - Without broader context, these similar phrases alone cannot determine AI vs human authorship - The full 50+ word context enables accurate detection Key Points: - Pangram's model analyzes word usage patterns in context rather than individual words - The 75-word minimum ensures sufficient context for reliable predictions - Single words or short phrases cannot provide definitive AI detection - The model considers multiple textual elements beyond vocabulary choice Technical Details: - Model: Most accurate AI detector currently available - Minimum requirement: 75 words - Confidence levels: Reports strong confidence for clear cases - Analysis method: Contextual pattern recognition over isolated word usage -------------------------------------------------------------------------------- title: "How can teachers tell if something is AI-generated? | Pangram" description: "A comprehensive guide detailing methods and indicators teachers can use to identify AI-generated student work." last_updated: "April 9, 2025" source: "https://www.pangram.com/blog/how-can-teachers-tell-if-something-is-ai-generated" -------------------------------------------------------------------------------- # Key Methods for Identifying AI-Generated Content in Education AI detection in student work relies on several key indicators and methods: ## Student History Comparison Teachers can identify AI content by comparing new work against established student patterns: - Previous demonstrated vocabulary level and writing sophistication - Known spelling and grammar error patterns - Demonstrated critical thinking capabilities in class discussions - Past writing samples and assignments - In-class written work examples ## AI Language Patterns Common AI writing indicators include: - Overuse of specific phrases like "faced numerous challenges" - Frequent use of words like "poignant" - Distinctive language patterns unique to AI systems - Consistent sophistication level throughout the text ## Content Hallucinations AI systems often produce specific types of errors: - Misstatement of widely known facts (e.g., incorrect vice president names) - Fabricated quotes - Failure to follow specific instructions - Incorrect references (e.g., citing wrong passages from texts) - Made-up information presented as fact ## Technical Detection Methods Detection tools and software provide automated identification: - Pangram AI detection system - Traditional plagiarism checking software with AI detection capabilities - Low false-positive rate detection tools recommended - Integration with existing academic integrity systems ## Important Considerations When identifying AI-generated content: - Individual indicators are not definitive proof - Multiple signs should prompt further investigation - Teacher-student discussions about writing process may be warranted - Honor code violation investigations may be necessary - Caution required when using detection tools ## Additional Resources Related guidance available at: - Discussion protocols: www.pangram.com/blog/what-to-do-when-student-submits-ai - Honor code documentation: www.pangram.com/blog/how-to-create-evidence-for-an-ai-detection-case Note: This content reflects educational standards and detection methods as of April 2025. -------------------------------------------------------------------------------- title: "Why Teachers Still Need AI Detection Tools | Pangram" description: "AI detection tools remain essential for teachers to manage bandwidth and maintain academic integrity despite their ability to recognize student writing styles" last_updated: "February 4, 2025" source: "https://www.pangram.com/blog/why-teachers-still-need-ai-detection-tools" -------------------------------------------------------------------------------- # Why Teachers Still Need AI Detection Tools Author: Jason Nicholson Published: February 4, 2025 Many teachers believe they can identify AI-generated content without detection tools due to their familiarity with student writing styles. However, several factors make AI detection software necessary in modern education: 1. Detection Bandwidth - Manual monitoring of AI writing consumes excessive teacher energy and time - Software detection allows teachers to focus on teaching writing rather than policing content - The mental economics favor using automated tools to reduce monitoring burden 2. Evolution of AI Writing - AI writing technology continues to advance and become more sophisticated - AI-generated content is increasingly ubiquitous in educational settings - Detection challenges will increase as AI improves 3. Real-World Example A case study demonstrates traditional plagiarism detection challenges: - Student submitted paper written by cousin from different school - Topic: Gatsby and American Dream analysis - Teacher identified inconsistency with student's usual writing style - Discovery required additional investigation time 4. Parallel to Cell Phone Management - Cell phone usage policies exist despite easy visual detection - Easy detection doesn't eliminate need for formal policies and tools - Manual enforcement consumes significant teacher attention - Structured approaches reduce enforcement burden Key Benefits of AI Detection Tools: - Reduces teacher monitoring workload - Preserves energy for meaningful instruction - Maintains academic integrity standards - Scales with increasing AI writing prevalence - Provides systematic detection approach The traditional "cat-and-mouse" dynamic between teachers and students regarding academic honesty can be productive in limited amounts, fostering critical thinking as students attempt to evade detection. However, excessive focus on detection diverts resources from core educational objectives. Products and Solutions Offered: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual Support Use Cases: - Teachers - Publishers - Content Moderation Resources Available: - AI Detection Model Documentation - Technical Report (arxiv.org/abs/2402.14873) - Data Privacy Guidelines - Educational Resources - State of AI Detection in 2025 - Common AI Phrases Guide - Assessment Strategies - Academic Integrity Frameworks -------------------------------------------------------------------------------- title: "Marketers are Wasting Advertising Spend on AI-generated Content | Pangram" description: "Analysis of how AI-generated content is leading to wasted advertising spend through made-for-advertising sites and programmatic advertising fraud" last_updated: "June 24, 2024" source: "https://www.pangram.com/blog/digital-advertising-primer" -------------------------------------------------------------------------------- # Key Findings Google has implemented measures to combat AI-generated content in search results, targeting "scaled content abuse" with a claimed 40% reduction in visibility for such content. Advertisers face increasing risks from AI-generated Made-for-Advertising (MFA) sites stealing advertising dollars through programmatic advertising networks. # Background and Market Impact Programmatic advertising spend exceeded $150B in 2023 in the USA alone. Made-for-Advertising sites accounted for: - 21% of ad impressions - 15% of total ad spend ($13B) - Over 140 major brands affected by AI-generated content farm placements # Consumer Trust and Engagement Recent research demonstrates declining trust in AI-generated content: - Trust in AI has fallen 15% over the last five years (Edelman Trust Institute) - 52% of consumers likely to disengage from AI-generated content (Bynder 2024 survey) - Millennials show stronger preference for human-generated content # Industry Response and Future Trends Gartner predicts 80% of marketers will develop dedicated content authenticity teams by 2027 to address these challenges. Key actions needed: - Enhanced visibility into ad placement locations - Greater transparency regarding content authenticity - Development of specialized teams and solutions for detecting inauthentic content - Implementation of real-time blocking systems for AI-generated spam sites # Problem Analysis AI tools have made it easier for fraudsters to: - Rapidly scale content farming operations - Generate large volumes of low-quality content - Game search ranking algorithms - Capitalize on misallocated advertising spend - Create and publish AI-generated MFA sites within hours # Solutions and Recommendations Advertisers must: 1. Recognize value differences between human and AI-generated content 2. Implement stronger content verification processes 3. Develop systems to identify and filter AI-generated content 4. Block ad placement on made-for-advertising sites 5. Monitor and adjust programmatic advertising strategies Pangram Labs offers: - Real-time AI content detection solutions - Frequently-updated MFA block list - Tools to help brands recapture lost advertising spend - Contact available at info@pangram.com for additional information # Market Context Google's March 2024 announcement outlined significant actions against AI-generated content in search results, specifically targeting scaled content abuse through algorithm updates and policy changes. This response indicates growing concern about the impact of AI-generated content on search quality and advertising effectiveness. -------------------------------------------------------------------------------- title: "Deep Dive on Yelp reviews | Pangram" description: "Analysis of AI detection model performance in identifying fake Yelp reviews, with benchmark testing results comparing Checkfor.ai, Originality.AI, and GPTZero." last_updated: "November 10, 2023" source: "https://www.pangram.com/blog/yelp-deep-dive" -------------------------------------------------------------------------------- # Key Information Checkfor.ai (now Pangram Labs) focuses on protecting online platforms from AI-generated content pollution, with particular emphasis on user review platforms. ChatGPT has made review fraud easier to commit at scale, threatening both business and consumer interests. ## Author Background Bradley Emi serves as CTO of Checkfor.ai, bringing experience from: - AI research at Stanford - ML Scientist role on Tesla Autopilot team - Research team leadership at Absci developing drug design neural networks ## Evaluation Philosophy The company approaches AI detection testing with the following principles: - Test sets function as unit tests for large models - Tests must cover millions/billions of parameters - 99% accuracy is insufficient - Evaluation must specifically test real-world use cases - Test sets must minimize confounding variables ## Test Set Requirements Good test sets must: - Answer specific questions - Test targeted scenarios (e.g., Yelp reviews, paraphrased text) - Avoid combining multiple variables - Prevent artificial inflation through easy examples - Maintain reproducibility and unbiased creation methods ## Yelp Benchmark Study Methodology: - Used Yelp's open source dataset - Randomly sampled 1000 genuine reviews - Generated 1000 synthetic reviews using ChatGPT - Reviews targeted real businesses from Yelp's dataset - Testing focused on text feature differentiation ## Performance Results Model accuracy comparison: - Checkfor.ai: 99.85% (1997/2000 correct) - Originality.AI: 96.2% (1738/1806 correct)* - GPTZero: 90.8% (1815/2000 correct) *Note: Originality.AI excludes documents under 75 words Error rate analysis: - Checkfor.ai: 1 failure per 666 queries - Originality.AI: 1 failure per 26 queries - GPTZero: 1 failure per 11 queries Comparative performance: - Checkfor.ai shows 25x better error rate than Originality.AI - Checkfor.ai shows 60x better error rate than GPTZero -------------------------------------------------------------------------------- title: "All About False Positives in AI Detectors | Pangram" description: "A technical explanation of false positive rates in AI detection systems and Pangram's approach to minimizing false positives across different types of content." last_updated: "March 27, 2025" source: "https://www.pangram.com/blog/all-about-false-positives-in-ai-detectors" -------------------------------------------------------------------------------- False positives in AI detection occur when human-written text is incorrectly flagged as AI-generated. False negatives occur when AI-generated text is mistakenly identified as human-written. These correspond to Type I and Type II errors in statistics, also known as sensitivity/specificity in medical sciences and precision/recall in machine learning. False positives are considered more harmful than false negatives in AI detection because they can damage student-teacher trust and cause unnecessary stress. This differs from medical contexts like cancer screening, where false negatives (missing actual cancer) are more dangerous than false positives. Pangram's False Positive Rates by Domain: - Academic Essays: 0.004% - Product Reviews: - English: 0.004% - Spanish: 0.008% - Japanese: 0.015% - Scientific Abstracts: 0.001% - Code Documentation: 0.0% - Congressional Transcripts: 0.0% - Recipes: 0.23% - Medical Papers: 0.000% - US Business Reviews: 0.0004% - Hollywood Movie Scripts: 0.0% - Wikipedia: - English: 0.016% - Spanish: 0.07% - Japanese: 0.02% - Arabic: 0.08% - News Articles: 0.001% - Books: 0.003% - Poems: 0.05% - Political Speeches: 0.0% - Social Media Q&A: 0.01% - Creative Writing/Short Stories: 0.009% - How-To Articles: 0.07% Optimal Performance Conditions: 1. Text length exceeds 200 words 2. Content uses complete sentences 3. Domain is well-represented in online training sets 4. Text contains creative rather than formulaic content Performance Strengths: - Highest accuracy: Essays, creative writing, reviews - Moderate accuracy: News articles, scientific papers, Wikipedia (due to abundant training data) - Lower accuracy: Recipes, poetry (due to shorter length, incomplete sentences, limited data) Usage Recommendations: - Avoid screening: - Short bullet point lists - Outlines - Mathematical content - Single sentences - Formulaic text (data lists, spreadsheets) - Template-based writing - Instruction manuals Overall false positive rate averages approximately 1 in 10,000, varying by content type and conditions. Pangram prioritizes minimizing both false positives and false negatives, with greater emphasis on reducing false positives due to their higher potential for negative impact. -------------------------------------------------------------------------------- title: "Can AI Detectors Catch GPT-4.5?" description: Research demonstrates Pangram's AI detection capabilities against OpenAI's GPT-4.5 model with comprehensive testing across diverse writing samples. last_updated: "February 27, 2025" source: "https://www.pangram.com/blog/can-ai-detectors-catch-gpt-4-5" authors: "Elyas Masrour and Bradley Emi" -------------------------------------------------------------------------------- OpenAI released GPT-4.5 as their latest frontier language model, representing a significant ChatGPT update. While not matching benchmark statistics of reasoning models like DeepSeek R1 and OpenAI O3, GPT-4.5 stands as 2025's most anticipated model release, featuring substantial improvements to writing quality. TESTING METHODOLOGY: Researchers tested 11 diverse writing prompts covering everyday writing tasks: 1. 300-word essay on koala conservation in Peru 2. Email terminating liberal op-eds at Washington Most newspaper 3. 400-word abstract announcing room temperature semiconductor discovery 4. Elementary student essay opposing school uniforms 5. 12-year-old's diary entry about poetry and butterflies 6. Review of Arabian Nights themed escape room in Baltimore 7. Russian indie film director's Academy Awards sanctions appeal 8. Creative fiction about NASA Mars landing simulation 9. Movie script about NYC finance broker's komodo dragon rescue 10. 200-word Halloween breakup poem 11. Creative fiction about Venice hover-motorcycle chase DETECTION RESULTS: Pangram vs Competitors Performance (AI Detection Accuracy): Prompt Type | Pangram | Competitor 1 | Competitor 2 ---|---|---|--- Koala Conservation | 100% | 100% | 100% Newspaper Email | 100% | 100% | 67% Semiconductor Abstract | 100% | 56% | 86% School Uniforms | 85% | 100% | 80% Poetry Diary | 100% | 100% | 15% Escape Room Review | 100% | 81% | 56% Russian Film Email | 100% | 100% | 91% Mars Landing Scene | 100% | 43% | 7% Komodo Dragon Script | 98% | 88% | 0% Halloween Poem | 100% | 100% | 0% Venice Chase Scene | 100% | 49% | 9% KEY FINDINGS: - Pangram achieved 100% detection rate on 10/11 samples without GPT-4.5 training data - Pangram maintained 98%+ confidence levels on 10/11 samples - Competing detectors showed inconsistent results with frequent false negatives - Competitors often expressed high uncertainty or incorrectly classified AI text as human-generated TECHNICAL ADVANTAGES: Pangram's superior performance stems from: - Large-scale machine learning model trained on millions of human/AI text examples - Active learning approach reducing false positives while increasing sensitivity - Enhanced generalization capabilities for new LLM detection - Focus on data quality and diversity - Advanced pattern recognition across AI-generated content - Ability to identify subtle linguistic markers CONTACT INFORMATION: - Email: info@pangram.com - Free GPT-4.5 detection credits available upon request -------------------------------------------------------------------------------- title: "60,000 AI-generated news articles are published every day | Pangram" description: "Analysis of AI-generated news content reveals approximately 60,000 AI-written articles are published daily, representing 7% of global news output" last_updated: "August 5, 2024" source: "https://www.pangram.com/blog/one-day-of-ai-news" -------------------------------------------------------------------------------- # Key Findings The news industry, valued at $150 billion, faces significant challenges from AI-generated content. Research conducted with NewsCatcher revealed approximately 60,000 AI-generated articles are published daily, representing 6.96% of global news output. ## Reader Behavior and Impact - Human-written content receives 93% more time-on-page than AI-generated content - Readers are 3.6x more likely to visit human-written articles - AI content farming operations captured 21% of ad impressions and over $10 billion in 2023 ## Research Methodology - Study conducted on July 1, 2024 - Analyzed 857,434 articles from 26,675 online publishers - Used NewsCatcher API covering 75,000+ sources globally - Employed Pangram Text classifier with 0.001% false positive rate ## Publisher Classification Framework Publishers categorized based on AI content percentage: - Human publisher: <10% AI content - Minor AI publisher: 10-50% AI content - Major AI publisher: 50-80% AI content - Fully AI-generated publisher: >80% AI content ## Geographic Distribution Countries with highest AI article frequency (minimum 100 articles): - Ghana shows highest concentration - India demonstrates significant AI content production, particularly notable during recent elections affected by deepfakes ## Detection Technology Pangram Text classifier features: - 30x more accurate than next leading commercial solution - 0.001% false positive rate for news content - Capability to analyze cleaned article text directly - Distribution analysis showing predictions concentrated near 0 or 1 (100-1000x more common than mid-range predictions) ## Total Impact Of 857,434 analyzed articles: - 59,653 classified as AI-generated - 6.96% overall AI content rate - Content spans multiple languages and topics - Affects global news ecosystem across various publishing platforms -------------------------------------------------------------------------------- title: "Data Privacy | Pangram" description: "Comprehensive data privacy policies and practices for Pangram's AI detection technology, including data collection, storage, and usage guidelines" last_updated: "2024" source: "https://www.pangram.com/data-privacy" -------------------------------------------------------------------------------- # Data Privacy at Pangram Labs ## Core Privacy Principles Pangram prioritizes creating products that serve teachers while maintaining student privacy. All privacy practices are designed around these dual objectives. ## Data Collection and Usage Pangram collects and stores: - Submitted text content for AI detection analysis - Query history displayed in dashboard and Chrome extension - Student-specific metadata for LMS integration matching - Account activity data for maintaining submission history ## Data Security Measures - All data encrypted at rest and in transit - Industry-standard security providers implemented - No third-party data sharing or selling - No data monetization practices ## Content Management - Content retention: Stored for duration of active account - Content access: Available in dashboard history tab - Content deletion: Users can remove content from history to delete from servers - Content usage restrictions: Not used for training AI models like ChatGPT - AI model training: Uses proprietary dataset; excludes user-submitted content ## Technical Implementation - Encryption: Applied to all stored and transmitted data - Storage: Secure servers with industry-standard protections - Access controls: Limited to account holders - Integration security: LMS connections maintain data protection standards ## Products and Solutions Pangram offers multiple AI detection tools: - Dashboard interface - Chrome extension - API access - LMS integrations - Plagiarism detection - Multilingual capabilities ## Additional Resources Technical documentation available: - AI detection model specifications - Implementation methodology - Technical report: arxiv.org/abs/2402.14873 - Comprehensive data privacy FAQ ## Compliance and Certification - SOC2 TYPE2 certified - Verified by AssuranceLab - Regular security audits and updates ## Support and Contact Support available through: - Email: info@pangram.com - Direct contact form - Community Discord: discord.gg/f7jDAPzWH3 -------------------------------------------------------------------------------- title: "Which is better: Pangram or Turnitin? | Pangram" description: A detailed comparison of Pangram and Turnitin AI detection tools, analyzing their accuracy rates, false positive rates, language support, and availability. last_updated: "May 13, 2025" source: "https://www.pangram.com/blog/pangram-vs-turnitin" -------------------------------------------------------------------------------- # Accuracy Comparison ## Detection Rates - Pangram: >99.9% detection rate for GPT-4 generated text - Turnitin: 76.8% detection rate (recall) for GPT-4 generated text, per September 2024 technical report ## False Positive Rates - Pangram: 0.01% (1 in 10,000) overall; 1 in 25,000 for academic essays - Turnitin: 0.51% (1 in 200) # ESL Performance - Pangram: 0.032% false positive rate across four ESL datasets - Turnitin: 1.4% false positive rate on ESL writing, 1.3% on non-ESL writing Head-to-head comparison on same dataset: - Pangram: 0.02% FPR on ESL, 0.0% on non-ESL (300+ words) - Turnitin: 1.4% FPR on ESL, 1.3% on non-ESL (300+ words) # Availability Pangram: - Free tier: twenty checks per day for teachers (up to 2,000 words daily) - Individual, Professional, and Team subscription options available - Institutional licenses with unlimited access - Available to all teachers regardless of institutional status Turnitin: - Institutional licenses only - No free trial or individual subscriptions available # Usage Methods for Pangram 1. Dashboard: - Free web interface - Paste text or upload documents - Detailed AI content reports - Charts showing AI location - Common AI phrase highlighting 2. Chrome Extension: - Right-click menu integration - Works across all websites - Canvas Speedgrader compatible - Discussion board compatible - Google Docs button integration 3. LMS Integration: - Available with institutional license - Supports Canvas and Google Classroom - Additional integrations available # Language Support Pangram (24 languages): English, Arabic, Chinese, Dutch, Czech, French, German, Greek, Hindi, Hungarian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swedish, Turkish, Ukrainian, Urdu, Vietnamese Turnitin (3 languages): English, Japanese, Spanish # Technical Validation - Pangram's accuracy supported by three independent academic studies - Technical reports published by both companies - Pangram demonstrates 99.9% recall on AI-generated text including GPT-4 - Turnitin reports 76.8% recall on GPT-4 text -------------------------------------------------------------------------------- title: "How to detect AI writing | Pangram" description: A comprehensive guide to detecting AI-generated text through analysis of structure, style, and common AI phrases. last_updated: "June 17, 2024" source: "https://www.pangram.com/blog/how-to-detect-ai-writing" -------------------------------------------------------------------------------- AI detection requires understanding several key components of writing and developing specific techniques for identification. CORE COMPONENTS FOR AI DETECTION: Topic Analysis: - AI typically writes about assigned topics rather than choosing them - Shows bias toward obvious subtopics when given broad prompts - Topic selection alone is not a reliable indicator of AI authorship Structure Patterns: - Default blog post structure: intro, 3-4 paragraphs, bullet points, conclusion summary - Restaurant reviews follow pattern: "pleasure of dining" opener, food/ambiance/service sections, enthusiastic recommendation - While structure can be modified through specific prompting, many users don't make this effort Style Indicators ("AI Tells"): Common AI phrases and words: - "Delve into" - "It's important to note" - "Tapestry" - "Vibrant" - "Bustling" - "In summary"/"In conclusion" - "Remember that..." - "Take a dive into" - "Navigating" (especially with "landscape" or "complexities") - "Landscape" (particularly "ever-evolving landscape") - "Testament to" - "In the world of" - "Realm" - "Embark" - "Symphony" - "Embrace" - "Whether you're X or Y" - "When it comes to" DETECTION TECHNIQUES: 1. Study AI Writing: - Generate multiple examples of target content type using ChatGPT - Analyze patterns across generated samples - Focus on specific document types (emails, reviews, etc.) 2. Examine Structure: - Look for default AI formatting patterns - Identify common organizational templates - Note predictable section transitions 3. Monitor Style: - Watch for overused phrases and words - Identify characteristic AI language patterns - Pay attention to unusual word frequency (e.g., "delve" appears more often in AI text) BACKGROUND: ChatGPT and other LLMs are increasingly used to generate "slop" - unwanted or lazily-generated AI content including: - SEO-driven blog posts - Artificial business reviews - Marketing content - Academic submissions Pangram Labs has developed a machine learning model trained on hundreds of millions of human and AI-written text examples to detect AI-generated content. [Content continues but appears truncated in source] Note: This document captures the complete available content from the source page, though the original article appears to continue beyond the visible content. -------------------------------------------------------------------------------- title: "Pangram can now detect AI humanizers | Pangram" description: "Pangram announces breakthrough capability to detect AI-humanized content with over 90% accuracy, surpassing existing detection tools" last_updated: "January 23, 2025" source: "https://www.pangram.com/blog/humanizers-announcement" -------------------------------------------------------------------------------- A market for AI detection evasion tools ("humanizers") has emerged alongside the growth of AI detection technology. These humanizer tools promise to make AI-generated text undetectable to AI detection systems. Pangram has developed a new model capable of detecting over 90% of high-quality humanized content. The research process involved: - Auditing 19 publicly available humanizers - Creating a classification system to rank humanizer quality - Retraining production models with high-quality humanized text in the training corpus Performance Comparison: Tool | AI Text Detection | Humanized AI Text Detection ---|---|--- GPTZero | 95.60% | 34.53% Binoculars | 94.40% | 29.73% Pangram Baseline | 100.00% | 73.07% Pangram Humanizers (current) | 100.00% | 93.66% Research Recognition: - Featured at Workshop on Detecting AI-Generated Content - International Conference on Computational Linguistics - Technical paper available at: arxiv.org/abs/2501.03437 Product Offerings: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities Target Markets: - Education (teachers) - Publishing industry - Content moderation/Trust & safety Resources Available: - AI detection model documentation - Technical implementation guides - Data privacy information - State of AI Detection in 2025 report - Common AI phrases analysis - Academic integrity frameworks - Assessment strategies for AI era Contact: info@pangram.com Certifications: SOC2 TYPE2 (Verified by AssuranceLab) -------------------------------------------------------------------------------- title: "Education | Pangram" description: "Educational institution inquiry portal for Pangram's AI detection and academic integrity tools with LMS integration capabilities" last_updated: "2024" source: "https://www.pangram.com/enterprise" -------------------------------------------------------------------------------- # Education Services Overview Pangram provides AI detection and authorship tracking solutions for educational institutions, focusing on academic integrity protection and student outcome improvement. The platform offers Learning Management System (LMS) integrations and standalone products designed for seamless integration across educational platforms. ## Core Educational Offerings - LMS integration capabilities - Authorship tracking systems - Academic integrity protection tools - Student outcome monitoring - Platform-agnostic implementation options ## Product Solutions - Dashboard interface - Chrome Extension - API access - Integration options - Plagiarism Detection - Multilingual capabilities ## Educational Resources - The State of AI Detection in 2025 - Most common AI phrases analysis - Technical report (arxiv.org/abs/2402.14873) ## Technical Certifications - SOC2 TYPE2 certification - Verification by AssuranceLab ## Support and Contact Email: info@pangram.com Community: Discord community available at discord.gg/f7jDAPzWH3 Social Media: Present on Instagram, Twitter, and LinkedIn ## Additional Features - Data privacy protection - Custom integrations - Content moderation capabilities - Publishing solutions - Trust and safety measures Educational institutions can request demonstrations and consultations through a dedicated inquiry form, which collects institution details, specific requirements, and demo preferences. -------------------------------------------------------------------------------- title: "Why Perplexity and Burstiness Fail to Detect AI | Pangram" description: "Technical analysis explaining why perplexity and burstiness metrics are unreliable for AI text detection, with examples and visualizations" last_updated: "March 4, 2025" source: "https://www.pangram.com/blog/why-perplexity-and-burstiness-fail-to-detect-ai" -------------------------------------------------------------------------------- Perplexity and burstiness are common metrics used in AI detection but have significant limitations in accurately identifying AI-generated content. PERPLEXITY DEFINITION: Perplexity measures how unexpected or surprising each word is from a language model's perspective. Low perplexity indicates common/expected words, while high perplexity indicates rare/unexpected words. Examples: - Low perplexity: "For lunch today, I ate a bowl of soup" - High perplexity: "For lunch today, I ate a bowl of spiders" BURSTINESS DEFINITION: Burstiness measures the variation in perplexity throughout a document, indicating how surprising words are distributed across the text. DETECTION METHODOLOGY: Commercial AI detectors use these metrics based on the assumption that: - Human text: Higher perplexity and higher burstiness - AI text: Lower perplexity and lower burstiness VISUALIZATION EVIDENCE: Analysis using GPT-2 model from Huggingface demonstrates: - AI-generated reviews: Consistent deep blue coloring indicating uniform low perplexity - Human reviews: Mostly blue with red spikes indicating high burstiness CRITICAL LIMITATIONS: 1. Training Set False Positives: - LLMs are trained to minimize perplexity on training documents - Common texts like the Declaration of Independence are falsely flagged as AI-generated - Frequent exposure during training causes memorization and low perplexity scores 2. Bias Against Non-Native Speakers: [Note: Content appears to be cut off in the source, but maintaining factual density in documented portion] PANGRAM PRODUCT OFFERINGS: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual capabilities USE CASES: - Teachers - Publishers - Content Moderation -------------------------------------------------------------------------------- title: "Pangram Text AI Detector now supports Arabic, Japanese, Korean, Hindi, and more | Pangram" description: "Pangram's Text AI detection model expands multilingual support to cover the top 20 internet languages with improved accuracy for Arabic, Japanese, Korean, and Hindi." last_updated: "September 4, 2024" source: "https://www.pangram.com/blog/pangram-multilingual-v2" -------------------------------------------------------------------------------- Pangram Labs has released an updated multilingual version of their Text AI detection model, expanding support to the top 20 languages on the Internet. The model demonstrates particularly strong performance improvements for Arabic, Japanese, Korean, and Hindi content. Performance Metrics: Evaluation conducted on approximately 2,000 documents per supported language, using real human-written reviews, news articles, and Wikipedia content, tested against GPT-4 generated content of varying lengths, styles, and topics. Accuracy Results by Language: - 100% accuracy: Spanish, Persian, French, Japanese, Polish, Portuguese, Russian - 99.95% accuracy: Arabic, Czech, Dutch, Romanian, Swedish, Vietnamese, Chinese - 99.90% accuracy: Greek, Turkish - 99.85% accuracy: German - 99.79% accuracy: Hindi - 99.49% accuracy: Hungarian - 99.44% accuracy: Urdu False Positive Rates: - 0.10%: Arabic, Hungarian, Dutch, Romanian - 0.00%: All other supported languages False Negative Rates: - 1.16%: Urdu - 0.95%: Hungarian - 0.42%: Hindi - 0.32%: German - 0.21%: Greek, Turkish - 0.11%: Czech, Swedish, Ukrainian, Vietnamese, Chinese - 0.00%: Remaining languages Technical Improvements: 1. Active learning data campaign targeting top 20 internet languages 2. Tokenizer optimization for non-English language support 3. Increased parameter count in base model and LoRA adapters 4. Implementation of machine translation data augmentation 5. Resolution of word counting bug affecting East Asian language representation Active Learning Implementation: - Mines pre-2022 Internet content for model improvement - Identifies and corrects false positives through iterative training - Automatically rebalances language distribution based on accuracy - Natural correction of low-resource language representation through error-driven data selection - Gradual reduction in high-resource language proportion (English, Spanish, Chinese) as model improves Architectural Enhancements: - Comprehensive evaluation of LLM backbones and tokenizers - Increased base model size to address multilingual underfitting - Expanded LoRA adapter parameter count - Extended training steps while maintaining single-epoch limitation - Implementation of machine translation augmentation using Amazon Translate -------------------------------------------------------------------------------- title: "Privacy Policy | Pangram" description: "Comprehensive privacy policy detailing how Pangram Labs collects, uses, and protects user information across its services" last_updated: "August 08, 2025" source: "https://www.pangram.com/privacy-policy.html" -------------------------------------------------------------------------------- Pangram Labs collects user information through multiple channels and uses it for specific purposes while providing privacy protections. Information Collection Methods: 1. User-Provided Information: - Registration requires name, email, password, and employment details - Users can submit text, documents, and media as Inputs for AI generation analysis - Direct communications capture name, email, phone, message contents - Payment information processed by third-party processors - Job applications collect resume data and LinkedIn profile information 2. Automated Collection: - Location data inferred from IP addresses - Device information including IP, browser type, operating system - Usage patterns including page views, purchases, visit timing - Cookies and tracking technologies gather user interaction data - Third-party analytics partners collect cross-service usage data 3. External Sources: - Publicly available information from third-party sites for model training Usage of Collected Information: - Service improvement and maintenance - Machine learning model training (excluding Google Workspace API data) - Personalization of user experience - Product development and analysis - Customer communication and support - Marketing and advertising targeting - Anonymous aggregate data generation - Payment processing - Fraud prevention - Legal compliance - Specific disclosed purposes Information Sharing: - Vendors and service providers supporting operations - Analytics partners including Google Analytics - Law enforcement when legally required - Corporate transaction partners during mergers/acquisitions - Third parties with user consent User Controls: - Query history deletion available (query details removed while maintaining billing records) - Marketing email opt-out via unsubscribe links - No response to Do Not Track signals - Data retention for duration needed to fulfill stated purposes Additional Policies: - Terms of Service govern usage: https://www.pangram.com/terms-of-service.html - Google Workspace APIs excluded from AI/ML model training - Analytics data collection detailed at https://www.google.com/policies/privacy/partners/ -------------------------------------------------------------------------------- title: "Myths and Misconceptions about AI Detection | Pangram" description: "An analysis of common misconceptions about AI detection technology, addressing accuracy, methodology transparency, and integration in education." last_updated: "February 25, 2025" source: "https://www.pangram.com/blog/myths-and-misconceptions" -------------------------------------------------------------------------------- Author: Jason Nicholson, English and Philosophy teacher at New Roads School in Los Angeles # Common Myths About AI Detection ## Myth 1: AI Detection vs. AI Integration AI detection and AI integration in teaching are complementary rather than conflicting approaches. AI detection serves as a prerequisite for effective AI incorporation by establishing necessary guardrails. Those advocating for strong AI detection tools typically want to maximize AI usage in classrooms while preventing abuse. Three common approaches to AI in education exist: - Pro-detection advocates: Seek to maximize AI use while maintaining controls - Unrestricted advocates: Push for unlimited AI tool usage without restrictions - Traditional advocates: Reject AI entirely, preferring pre-computerized teaching methods ## Myth 2: AI Detector Transparency Pangram has publicly shared its methodology through: - Published technical documentation - Interactive website demonstrations - Conference presentations - Technical innovations presented at COLING conference, including work on system robustness against humanizers and paraphrasers ## Myth 3: Academic Validation Pangram's technology has received extensive peer review and validation: - Won most accurate detector award at COLING Shared Task competition - University of Maryland research confirms Pangram outperforms human experts in AI text detection - University of Houston research demonstrates unique robustness to translation - Offers unlimited free access to academic researchers for accuracy studies Notable: Earlier studies (Weber-Wulff 2023, Liang) criticizing AI detectors did not include Pangram in their benchmarks. ## Myth 4: Detector Accuracy Pangram's accuracy metrics: - 1/10,000 false positive rate - 100x better accuracy than next best commercial software - Cannot achieve 100% accuracy but maintains high reliability Key points about accuracy claims: - Some institutions make unsubstantiated claims about detector ineffectiveness - Statistical improbability of multiple false positives occurring simultaneously - Methodology and accuracy rates documented in white paper Additional Resources: - Technical report: arxiv.org/abs/2402.14873 - Humanizers announcement - COLING Shared Task: arxiv.org/abs/2501.08913 - University research collaborations - Detailed methodology documentation -------------------------------------------------------------------------------- title: "Technical Report on High Accuracy AI-generated Text Detection | Pangram" description: "Pangram Labs presents technical analysis and benchmarking of their AI-generated text detection model achieving 99.85% accuracy across multiple LLM types." last_updated: "February 21, 2024" source: "https://www.pangram.com/blog/technical-report-february-2024" -------------------------------------------------------------------------------- Pangram Labs has developed an AI-generated text detection model achieving 99.85% accuracy with a 0.19% false positive rate across thousands of examples spanning ten writing categories and eight large language models. Key Findings: - Model maintains 99-100% accuracy across all tested language models including GPT-4 - Significantly outperforms other detection methods which achieve ≤75% accuracy on GPT-4 - Comprehensive benchmark tested against 1,976 documents (50% human-written, 50% AI-generated) Comparison of Detection Tools: - Pangram Labs - GPTZero - Originality.ai - DetectGPT (2023 academic state-of-the-art method) Problem Context: LLM capabilities reached an inflection point in 2023, enabling widespread generation of human-like text. This creates risks: - Spam emails becoming harder to filter - Mass generation of fake marketplace reviews - Social media manipulation through LLM-powered bots - Content authenticity and quality concerns Technical Approach: - Model trained specifically to detect AI-generated spam and fraudulent content - Testing conducted across multiple document categories and LLM types - Evaluation sets deliberately made increasingly difficult to ensure real-world accuracy - Current benchmark designed to be slightly more challenging than typical spam content Methodology Details: Full technical methodology available in accompanying whitepaper: "Technical Report on the Pangram AI-Generated Text Classifier" (arxiv.org/abs/2402.14873) Performance Metrics: - Accuracy: Percentage of total documents correctly classified - False Positive Rate: Percentage of human documents incorrectly flagged as AI - False Negative Rate: Percentage of AI documents incorrectly classified as human Products/Solutions: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual Support Use Cases: - Education/Teachers - Publishers - Content Moderation - Trust and Safety -------------------------------------------------------------------------------- title: "Three percent of front-page Amazon reviews are now AI-generated | Pangram" description: A study reveals widespread AI-generated reviews on Amazon's best-selling products, with 3% of front-page reviews being AI-created. last_updated: "July 25, 2025" source: "https://www.pangram.com/blog/ai-amazon-reviews" -------------------------------------------------------------------------------- AI-generated reviews have become prevalent on Amazon's platform, with a new study revealing significant findings about their impact on product ratings and consumer trust. Key Study Findings: - Analysis covered 30,000 front-page product reviews across 500 Amazon best-selling products - Study examined 10 product categories: baby, toys and games, laptops, medical devices, wellness and relaxation, beauty, and furniture - 3% of total reviews (909 reviews) were confirmed AI-generated with high confidence - 74% of AI-written reviews gave 5-star ratings vs 59% of human reviews - 22% of human reviews gave 1-star ratings vs 10% of AI reviews - 93% of AI-generated reviews displayed the "Verified Purchase" badge Legal and Platform Response: - AI-generated reviews are illegal under U.S. federal law (FTC regulations) - Amazon employs AI tools to flag suspicious review patterns - Current detection efforts prove insufficient based on study findings Review Characteristics and Detection: - AI reviews tend toward positive ratings - Generic language and lack of specific product details indicate potential AI generation - "Verified Purchase" badge no longer serves as reliable authenticity indicator Tools for Detecting AI Reviews: - Pangram offers AI detection dashboard for review verification - Chrome extension available for real-time AI content checking - Platform can analyze any selected text for AI generation markers Impact on Consumer Trust: - AI reviews potentially inflate product ratings - Misleading reviews affect purchasing decisions - Verification of review authenticity becomes increasingly important - Well-meaning customers may use AI to overcome writing challenges Best Practices for Reviews: - Include specific product details in reviews - Avoid using ChatGPT or other LLMs for review generation - Focus on honest, personal experiences - Verify suspicious reviews using AI detection tools Current Challenges: - Content farms can produce AI reviews at scale - Traditional trust signals becoming less reliable - Platform detection systems require improvement - Growing difficulty in distinguishing authentic reviews The proliferation of AI-generated reviews presents ongoing challenges for e-commerce platforms and consumers, requiring enhanced detection methods and stricter enforcement of review authenticity standards. Here's my attempt at a complete LLMS.txt document following the guidelines: -------------------------------------------------------------------------------- title: "What happens when an AI detector makes a mistake? | Pangram" description: "Analysis of AI detector error rates, false positives, and guidance for teachers handling potential mistakes in AI detection" last_updated: "May 15, 2025" source: "https://www.pangram.com/blog/what-happens-when-an-ai-detector-makes-a-mistake" -------------------------------------------------------------------------------- Early AI detectors have demonstrated significant reliability issues in two key areas: false negatives (failing to detect AI-generated content) and false positives (incorrectly flagging human-written content as AI-generated). Determined cheaters can bypass detection through light paraphrasing or intentional misspellings. Non-native English speakers face disproportionate impacts from false positives. A 2023 Stanford study revealed that multiple AI detectors unanimously misidentified 20% of non-native English speakers' essays as AI-generated, with nearly all essays being flagged by at least one detector. Current false positive rates by provider: - TurnItIn: 1 in 200 (0.5%) - Various other tools: Between 1 in 100 (1%) and 1 in 500 (0.2%) - Pangram: 1 in 10,000 (0.01%), tested on tens of millions of documents - Independent studies suggest actual rates may be higher than advertised Recommended teacher protocol for handling AI detector flags: 1. Approach student with humility and inquire about AI use 2. Request evidence of writing process: - Google Docs revision history - Early draft copies - Detailed discussion of writing process 3. If student denies AI use but cannot provide evidence: - Grant benefit of doubt - Instruct on future process documentation - Monitor for repeated flags 4. For multiple detections: - Consider escalation - Note: Multiple false positives are statistically unlikely with accurate detectors Pangram's detection model performs particularly well on: - Texts exceeding several hundred words - Writing containing complete sentences - Academic assignments and longer-form content Key considerations for academic integrity: - False positives can damage student trust - Documentation of writing process helps prevent misunderstandings - Multiple detections warrant closer investigation - Teachers should maintain balanced approach between enforcement and fairness The article emphasizes a cautious, evidence-based approach to AI detection in academic settings, prioritizing student trust while maintaining academic integrity standards. -------------------------------------------------------------------------------- title: "Introducing Pangram's Plagiarism Detection | Pangram" description: "Pangram launches plagiarism detection capabilities alongside its AI detection tools to provide comprehensive academic integrity solutions" last_updated: "August 1st, 2025" source: "https://www.pangram.com/blog/introducing-pangram-s-plagiarism-detection" -------------------------------------------------------------------------------- # Evolution of Plagiarism and New Detection Tools Traditional plagiarism involved students copying text directly from existing sources including former student work, academic journals, and internet content. ChatGPT disrupted this pattern by enabling instant generation of custom assignments from simple prompts, making traditional plagiarism methods seem less appealing. Traditional plagiarism remains relevant as students may: - Revert to copying existing content to avoid AI detection flags - Prefer using previously graded essays over AI-generated content - Combine multiple cheating methods # Pangram Plagiarism Detection Launch Pangram's new plagiarism detection system searches submitted documents across millions of sources including articles, websites, forums, and databases. The feature integrates directly into the Pangram dashboard through a toggle switch that activates plagiarism scanning for individual documents. Core functionality: - Searches open web for matching text patterns - Identifies direct matches of extended text sequences - Highlights copied sentences within documents - Provides source links to matching online content - Integrates with existing AI detection capabilities # Comprehensive Academic Integrity Tools Pangram's academic transparency suite combines: - High-accuracy AI detection - Traditional plagiarism checking - Writing process playback - Common AI phrase detection - Additional classroom management tools # Access and Availability Current access is limited to: - Paid subscription holders - Organizations with licensing agreements Additional information available at pangram.com/solutions/plagiarism Contact: info@pangram.com # Product Solutions Complete solution suite includes: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual Capabilities Use cases span: - Education (Teachers) - Publishing Industry - Content Moderation/Trust & Safety Resources include: - AI detection model documentation - Technical implementation guides - Data privacy information - Educational materials on AI detection - State of AI Detection in 2025 report - Common AI phrases guide - Academic integrity frameworks Company maintains SOC2 TYPE2 verification through AssuranceLab. -------------------------------------------------------------------------------- title: "AI detection just got a lot better: Announcing Checkfor.ai | Pangram" description: "Announcement of Checkfor.ai's launch as an AI content detection tool, created by former Google and Tesla engineers" last_updated: "October 12, 2023" source: "https://www.pangram.com/blog/announcing-checkforai" -------------------------------------------------------------------------------- Checkfor.ai launched as a new AI content detection tool focused on identifying AI-generated content through simple copy-and-paste functionality. The company has since rebranded to Pangram Labs. Core Mission: - Protect digital world from low-quality AI-generated content - Preserve authentic human voices - Provide verification tools for content publishers, review platforms, educators, and creators Technical Team Background: - Comprised of machine learning researchers and engineers - Team members have previous experience at: - Google - Tesla Autopilot - Nuro Development Approach: - Focus on reliability and trust - Goes beyond 99% test set accuracy - Aims for low false positive and false negative rates - Designed for high-stakes real-world deployments Current Status: - Development time: Approximately one month - Claims superior detection compared to competitors - Particularly effective for: - Creative writing - User-published online content Product Features: - Simple copy-paste interface - AI content detection capabilities - Web-based platform Available Solutions: - Dashboard - Chrome Extension - API - Integrations - Plagiarism Detection - Multilingual Support Target Markets: - Teachers - Publishers - Content Moderators Technical Specifications: - SOC2 TYPE2 certified - Verified by AssuranceLab - Technical report available on arXiv (2402.14873) Contact Information: Email: info@pangram.com Community: Discord available at discord.gg/f7jDAPzWH3 Social Media: Present on Instagram, Twitter, and LinkedIn -------------------------------------------------------------------------------- title: "Pangram Text AI Detector is now multilingual! | Pangram" description: "Pangram Text AI detection model now supports 7 additional languages with industry-leading accuracy for detecting AI-generated content" last_updated: "July 1, 2024" source: "https://www.pangram.com/blog/pangram-text-multilingual" -------------------------------------------------------------------------------- Pangram Text has expanded its AI detection capabilities to support seven new languages while maintaining high accuracy levels: Spanish, French, Italian, Portuguese, German, Russian, and Mandarin Chinese. The multilingual model is immediately available for online platform protection against AI spam. BENCHMARKING METHODOLOGY: - Testing conducted across 3 multilingual corpora: Amazon reviews, Wikipedia, and XLSum (BBC News International) - Human benchmark: Random documents filtered through sanity checks - AI benchmark: Mix of GPT-3.5, GPT-4 and GPT-4o outputs - Generation process: LLMs summarize real documents, then generate new content based on summaries - Methodology ensures clean labels and similar human/AI data distributions ACCURACY RATES BY LANGUAGE AND CORPUS: Spanish: - Amazon Reviews: 99.59% - Wikipedia: 99.75% - XLSum: 99.75% French: - Amazon Reviews: 98.84% - Wikipedia: 99.33% - XLSum: 98.50% Italian: - Wikipedia: 99.82% - Other corpora: Not available German: - Amazon Reviews: 99.44% - Wikipedia: 99.95% - XLSum: Not available Portuguese: - Wikipedia: 99.83% - XLSum: 99.70% - Amazon Reviews: Not available Russian: - Wikipedia: 98.34% - XLSum: 99.35% - Amazon Reviews: Not available Chinese: - Amazon Reviews: 99.70% - Wikipedia: 99.54% - XLSum: 98.10% TECHNICAL IMPLEMENTATION: - Architecture similar to modern large language models - Large scale pretraining on multilingual corpus - Fine-tuned AI detection head - Tokenizer supporting multiple languages including Russian and Chinese - Language detection via Amazon Comprehend - Returns "Unsupported Language" for non-supported languages FUTURE DEVELOPMENT: - Continuous improvement through active learning - Planned expansion to additional languages - Regular updates to enhance non-English language performance Contact: info@pangram.com for multilingual AI detection information -------------------------------------------------------------------------------- title: "Announcing AI Identification: Pangram can distinguish the different LLMs from each other" description: Pangram announces new AI Identification technology that can distinguish between different large language models and identify which LLM generated specific text. last_updated: "February 11, 2025" source: "https://www.pangram.com/blog/pangram-can-distinguish-llms" author: "Bradley Emi" -------------------------------------------------------------------------------- Pangram has developed advanced AI detection software that can identify AI-generated text from ChatGPT, Claude, Gemini and other LLMs, distinguishing it from human-written content. The company has now expanded capabilities with "AI Identification" technology that can determine which specific LLM generated a piece of text. Different LLMs exhibit distinct writing styles: - ChatGPT: Direct and straightforward - Claude: Fluent and conversational - Grok: Uncensored and provocative - Deepseek-R1: Verbose and rambly Research from UC Berkeley (Dunlap et al.) found distinctive LLM characteristics: - Llama displays more humor and formatting - Llama provides more examples - Llama comments less on ethics compared to GPT and Claude - These differences affect human preferences on platforms like Chatbot Arena Technical Implementation: - Uses multi-task learning approach - Same detection model trained for both AI detection and identification - Classifies LLMs into 9 distinct families - Neural network includes additional "head" for model identification - Shares most layers between detection and identification tasks The 9 LLM families identified: 1. GPT-3.5 2. GPT-4 (including GPT-4o, GPT-4-turbo, GPT-4o-mini) 3. OpenAI Reasoning Models 4. Claude 5. Google (Gemini variants and Gemma) 6. Grok 7. DeepSeek 8. Amazon Nova 9. Other (LLaMA, Mistral, Qwen, open-source derivatives) Research findings: - LLM identification task shows symbiotic relationship with AI detection - External research confirms LLMs are distinguishable from both human text and each other - Style embeddings demonstrate clear separation between different LLMs in vector space - Documents from same LLM cluster closer together than those from different sources -------------------------------------------------------------------------------- title: "Does Pangram detect Meta's Llama 4?" description: Pangram's AI detection system achieves 99.93% accuracy in detecting content generated by Meta's new Llama 4 model. last_updated: "April 6, 2025" source: "https://www.pangram.com/blog/does-pangram-detect-llama-4" -------------------------------------------------------------------------------- Pangram conducted immediate testing of their AI detection capabilities against Meta's newly released Llama 4 model, evaluating performance before any model retraining. TEST METHODOLOGY: - Used 11 standardized prompts previously used for GPT-4.5 testing - Prompts cover everyday writing tasks unrelated to training data - Focus on creative tasks to identify behavioral differences between LLM generations TEST PROMPTS: 1. 300-word essay on koala conservation in Peru 2. Email terminating liberal op-eds at Washington Most newspaper 3. 400-word abstract announcing room temperature semiconductor discovery 4. Elementary student essay opposing school uniforms 5. 12-year-old's diary entry about poetry and butterflies 6. Review of Arabian Nights themed escape room in Baltimore 7. Russian indie film director's Academy Awards sanctions appeal 8. NASA Mars landing simulation scene 9. Script about NYC finance broker's Florida komodo dragon rescue 10. 200-word Halloween breakup poem 11. Venice hover-motorcycle chase scene INITIAL RESULTS: - All 11 test samples identified as AI-generated - 99.9% confidence level achieved for each sample - Perfect detection score across all prompts - Full outputs available in public documentation EXPANDED TESTING: - Secondary evaluation using Together API - Approximately 7,000 test examples - Coverage across academic, creative, Q&A, and scientific writing - Test results by model variant: - Llama 4 Scout: 100% accuracy (3,678/3,678 samples) - Llama 4 Maverick: 99.86% accuracy (3,656/3,661 samples) - Combined accuracy: 99.93% (7,334/7,339 samples) TECHNICAL FOUNDATIONS: - Success attributed to robust underlying datasets - Active learning approach enhances adaptability - Broad prompting and sampling strategies enable generalization - Model demonstrates strong generalization to new LLM variants without retraining CONTACT INFORMATION: - Research inquiries and trial credits: info@pangram.com - Technical report available: arxiv.org/abs/2402.14873 -------------------------------------------------------------------------------- title: "AI Detector for Publishers | Pangram" description: "AI detection platform helping publishers verify content authenticity and maintain trust through advanced authorship analysis" last_updated: "2025" source: "https://www.pangram.com/use-cases/publishing" -------------------------------------------------------------------------------- # Core Product Information Pangram Labs offers an AI detection dashboard that analyzes authorship patterns in written content. The platform provides detailed analysis beyond simple AI probability scores to help publishers verify content authenticity. # Key Statistics - 55% of consumers can identify AI-generated content - 20% of consumers perceive brands as lazy for using AI content - 93% of AI-generated Amazon reviews carry "Verified Purchase" badges # Publisher Challenges 1. Content Authenticity: - AI-generated content cannot be copyrighted per US Copyright Office ruling (March 2023) - Accidental publication of AI content risks public backlash - AI writing reduces reader engagement and is perceived as lower quality 2. Research Integrity: - High AI content in submissions can indicate academic fraud - Peer reviews require verification against AI usage - Journal standards compliance requires AI detection # Technical Capabilities Pangram's detection system: - Developed by AI researchers from Tesla and Google - Features self-learning improvement mechanisms - Maintains highest market accuracy with lowest false positive rate - Provides unbiased detection for non-native English speakers - Technical documentation available via published research paper # Integration Partners Verified partnerships with: - Canvas - Google Classroom - Quora - Tremau - The Transparency Company - NewsGuard # Platform Access Options Available through: - Web dashboard - Chrome extension - API integration - Plagiarism detection tools - Multilingual support system # Founder Statement From Max & Bradley, Co-Founders: "As researchers, we prioritize original work authenticity verification. The platform addresses modern generative AI challenges while preserving human ingenuity. Pangram enables content originality tracking and plagiarism prevention while acknowledging AI's inevitable integration into writing workflows." # Legal Context U.S. Copyright Office Ruling (March 2023): "When an AI technology determines the expressive elements of its output, the generated material is not the product of human authorship. As a result, that material is not protected by copyright and must be disclaimed in a registration application." -------------------------------------------------------------------------------- title: "Scaling up with LoRA | Pangram" description: "Technical details of Pangram's AI detection model improvements using Low-Rank Adaptation (LoRA) and scaled computing resources" last_updated: "March 22, 2024" source: "https://www.pangram.com/blog/scaling-up-with-lora" -------------------------------------------------------------------------------- # Model Performance Improvements Pangram has released an updated AI detection model with significant performance improvements: February 2024 Model metrics: - Accuracy: 99.0% - False Negative Rate: 1.30% - False Positive Rate: 0.67% March 2024 Model metrics: - Accuracy: 99.84% - False Negative Rate: 0.11% - False Positive Rate: 0.19% # Technical Implementation Details The improved model utilizes: - Active learning approach: Hard Negative Mining with Synthetic Mirrors - Increased parameter count by an order of magnitude - Low-Rank Adaptation (LoRA) for efficient fine-tuning - Training on NVIDIA H100 GPUs # Model Architecture Insights Key findings on model architecture: - Smaller models perform better for DetectGPT implementations - Data scaling shows saturation around 40k documents - Traditional text classification tasks still rely on smaller models (XLNet, DeBERTa, T5-XXL) with hundreds of millions of parameters - Large language models (tens of billions of parameters) tend to overfit on classification tasks # LoRA Implementation Low-Rank Adaptation (LoRA) benefits: - Preserves base LLM knowledge while preventing overfitting - Reduces training time and memory requirements - Prevents catastrophic forgetting of pretraining data - Uses adapter modules as side networks alongside LLM attention blocks - Implements parameter-efficient weight matrices for quick training Technical approach: - Base LLM remains frozen - Adapter modules train alongside core attention blocks - Original LLM matrix (W) stays static while LoRA modules train around it - Decomposition into parameter-efficient weight matrices enables memory efficiency # Future Development Planned improvements: - Ongoing architecture updates to maintain current standards - Additional architectural improvements in development - Data improvements in pipeline - Creation of more challenging evaluation sets Contact: info@pangram.com -------------------------------------------------------------------------------- title: "How It Works | Pangram" description: "Technical explanation of Pangram's AI content detection model architecture, training process, and accuracy optimization methods" last_updated: "August 08, 2025" source: "https://www.pangram.com/research/how-it-works" -------------------------------------------------------------------------------- # Pangram AI Detection System Technical Overview ## Core Model Architecture Pangram's AI content detection classifier employs a traditional language model architecture that processes text through multiple stages: - Text tokenization of input content - Token-to-embedding conversion creating numerical vector representations - Neural network processing generating output embeddings - Classifier head transformation producing binary predictions (0=human, 1=AI) ## Initial Training Process The foundational model training utilizes: - Dataset size: ~1 million documents - Data composition: Mix of public and licensed human-written text plus AI-generated content from GPT-4 and other frontier language models - Training objective: Reliable human vs AI authorship classification ## Advanced Accuracy Optimization ### Hard Negative Mining To achieve 99.999% accuracy and near-zero false positives, Pangram implements: 1. Large dataset scanning to identify false positives 2. Training set augmentation with hard edge cases 3. Model retraining with expanded dataset 4. Iterative improvement through multiple optimization cycles ### Mirror Prompts Methodology The training data balancing process includes: - Generation of AI content matching human examples - Style, tone, and semantic content matching - Focus on maintaining consistent characteristics across paired samples - Emphasis on identifying LLM-specific writing patterns ## Continuous Improvement Cycle The model undergoes systematic enhancement through: 1. Initial model training 2. Performance evaluation 3. Hard negative mining 4. Dataset augmentation 5. Retraining with expanded data 6. Repeated evaluation and optimization ## Technical Documentation Complete technical specifications and methodology details are available in the published white paper on arXiv (https://arxiv.org/pdf/2402.14873). ## Product Integration The AI detection system is available through multiple interfaces: - Dashboard - Chrome Extension - API - Third-party integrations - Multilingual detection capability - Plagiarism detection features ## Use Case Support Primary applications include: - Educational institution deployment - Publishing industry implementation - Content moderation systems - Trust and safety operations -------------------------------------------------------------------------------- title: "Grammarly gets a tone detector to keep you out of email trouble | TechCrunch" description: "Grammarly launches a beta tone detector feature that analyzes email tone across 40 different emotional categories" last_updated: "September 24, 2019" source: "https://techcrunch.com/2019/09/24/grammarly-gets-a-tone-detector-to-keep-you-out-of-email-trouble/" -------------------------------------------------------------------------------- # Core Information Grammarly has expanded beyond basic grammar and spellchecking by launching a beta tone detector feature. The tone detector analyzes whether emails and documents match intended tone preferences, such as friendly but not overly informal. ## Technical Details - Uses combination of set rules and machine learning algorithms - Analyzes text signals that contribute to overall tone - Requires minimum of 120 characters to activate - Can detect 40 different tones including: - Appreciative - Confident - Formal - Informal - Thoughtful - Loving - Sad ## Availability - Currently available in Chrome browser extension - Coming soon to Safari and Firefox - Initially works with major email services including Gmail and Yahoo - Full text field support planned for future release ## Context This tone detector follows Grammarly's recent expansion into more detailed clarity scoring features. The tool aims to help users who struggle with appropriate email tone, particularly those whose written communication may come across differently than their in-person interactions. ## Author Information Frederic Lardinois served as TechCrunch editor from 2012-2025. Previous experience includes founding SiliconFilter and writing for ReadWriteWeb. Coverage areas include enterprise, cloud, developer tools, Google, Microsoft, gadgets, and transportation technology. -------------------------------------------------------------------------------- title: "Pangram Labs System Status" description: "Real-time operational status dashboard for Pangram Labs services showing uptime and performance metrics" last_updated: "Current as of status check" source: "https://status.pangram.com/" -------------------------------------------------------------------------------- # Pangram Labs System Status All Pangram Labs systems are currently operational with three main services being monitored: ## Service Status & Uptime 1. Website (pangram.com) - Current Status: Operational - Uptime: 100% - Last checked: 49 minutes ago 2. Individual Inference - Current Status: Operational - Uptime: 99.72% - Last checked: 52 minutes ago 3. Batch Inference - Current Status: Operational - Uptime: 99.86% - Last checked: 52 minutes ago ## Monitoring Features Status page includes: - Real-time service monitoring - Uptime percentage tracking - Timestamp of last status check - Optional clickable monitoring URLs - Badge generator functionality with customization options: - Badge types: status, uptime, ping, avg-response, cert-exp, response - Style options: plastic, flat, flat-square, for-the-badge, social - Customizable colors for up/down/pending/maintenance states ## Access Controls - Configurable visibility settings for monitoring URLs - Shared access capabilities for authorized users - Status page deletion protection with confirmation dialog