How to detect AI writing

ChatGPT writes 2000 words of slop in a matter of seconds

ChatGPT and other large language models (LLMs), commonly referred to in aggregate as "AI," have become increasingly popular tools. Their most common use case is as an assistant, where a user can chat directly with the language model. However, LLMs are increasingly used to create slop, a term for unwanted or lazily-generated AI content. An example of AI slop is a blog post fully written by ChatGPT and posted on the internet for the purpose of driving search traffic. Another example of AI slop is a business trying to boost their ratings by using AI to write several lengthy, glowing reviews.

Pangram Labs has trained a high-accuracy machine learning model to classify AI-generated text. By the end of training, the model has seen hundreds of millions of examples of human-written and AI-generated text. This post will help you build your own intuition for spotting AI text, ideally with fewer examples.

I've spoken to educators, admissions counselors, and publishers who all were pushed into learning this skill in the last year due to the massive surge in AI usage. This article is a combination of information I've learned from them, as well as techniques we use internally at Pangram to get new hires up to speed. After all, you can't train a good machine learning model if you're unfamiliar with the task itself.

TL;DR

• Look for common AI phrases.
• Ask ChatGPT to generate multiples of the type of document you are trying to detect and look for similarities.
• Try your hand at making AI writing undetectable and look for what sticks out.
• Use AI detection tools to hone your intuition.

Components of writing

Topic

Generally, AI doesn't get to choose its own topic - it will be prompted to write about something specific. Even so, AI will show bias towards the subtopics it chooses for itself. For example, if I prompt ChatGPT to write me an essay about the symbolism in Moby Dick, it will likely choose the same set of most obvious topics. However, any student could prompt ChatGPT to write specifically about a certain form of symbolism in the book, and ChatGPT will be able to write about it. For that reason, I don't put too much consideration into the topic when I'm evaluating whether text is AI or not.

Structure

When asked to write full documents, AI generally has a certain structure it will default to. Ask it to write a blog post and it will start with an intro, 3-4 paragraphs, some bullet points and a conclusion summarizing the post. Similarly, if I want AI to write me a restaurant review, it will often start by saying "I recently had the pleasure of dining at ____", talk about the food, ambiance, and service, and conclude with an enthusiastic "Highly recommend". While the default structure may also be prompted away ("Write me a 20-paragraph blog post", "Keep the review short, don't mention the restaurant name."), this takes effort that many prompters are unwilling to expend. When writing is structured similar to other AI writing, that's a strong initial indication to dig deeper.

Style

In poker, the word "tell" refers to a cue someone unknowingly gives when they bluff. For example, when someone is nervous they may fiddle with their glasses, leaking information to about their bets. Similarly, at Pangram we refer to certain words or phrases as "AI tells" - these are words or phrases that ChatGPT or other language models favor disproportionately, and they're a very strong signal of AI origin once you start noticing them.

"Delve" is an extreme example of an AI tell because ChatGPT uses it much more often by ChatGPT than it would be in normal American English.

I've also noticed that "I recently had the pleasure" is a common phrase for restaurant reviews and ChatGPT always manages to insert the word "keen" if you ask it to write a cover letter. Even with heavy prompting, LLMs are strongly driven by internal stylistic tendencies. For that reason, style is a strong signal that I index heavily on.

Here is a list of common AI phrases to use as a starting point. (Sources: reddit, aiphrasefinder.com)

• "It's important to note"
• "Delve into"
• "Tapestry"
• "Vibrant"
• "Bustling"
• "In summary" or "In conclusion"
• "Remember that...."
• "Take a dive into"
• "Navigating" i.e. "Navigating the landscape" "Navigating the complexities of"
• "Landscape" i.e. "The landscape of....", especially "ever-evolving landscape"
• "Testament" i.e. "a testament to..."
• "In the world of"
• "Realm"
• "Embark"
• "Symphony"
• "Embrace"
• "Whether you're X or Y"
• "When it comes to"

Techniques to improve your intuition

Phrases are a good baseline, but often AI will generate some text that is obviously AI-written but doesn't use any of these phrases. To improve this skill, we'll need to spend some time actually studying AI writing.

Study AI writing

It may be distasteful to some or exciting to others, but if you want to identify AI you need to know what AI writing looks like. It helps if you have a particular type of text you want to be able to identify. Let's say I want to get really good at identifying AI emails. The first thing to do is to ask ChatGPT to write a cold introduction email. A thank you email. A marketing email. Repeat a few times, changing up the prompt and writing down the similarities between each iteration. Next, try again with Claude and Gemini, continuing to look for patterns.

Subject: Exploring Collaboration Opportunities with OpenAI

Dear [Founder's Name],

I hope this message finds you well. My name is [Your Name], and I am the [Your Title] at [Your Company/Organization], where we specialize in [briefly describe your field or expertise]. I have been following OpenAI's groundbreaking advancements in artificial intelligence, and I'm truly impressed by your commitment to ethical AI development and the transformative impact of your projects.

I am reaching out to explore potential collaboration opportunities between our organizations. I believe that combining OpenAI's AI capabilities with our expertise in [your specific field or technology] could lead to innovative solutions that further advance the field and create substantial value.

I would appreciate the opportunity to discuss this further and explore ways we can work together towards our mutual goals. Could we schedule a brief call or meeting at your convenience? I am flexible with timings and would be happy to accommodate your schedule.

Thank you for considering this opportunity. I look forward to the possibility of working together and contributing to the exciting developments at OpenAI.

Warm regards,

[Your Full Name]

[Your Position]

[Your Company/Organization]

[Contact Information]

[LinkedIn Profile or Website] (optional)

Play adversarial games

When we were starting Pangram Labs, we did a ton of what we called The Human Baseline. The rules are simple. You choose a topic. Our first human baseline was admission essays. Person A procures five college admissions essays written by humans. Then, Person A uses ChatGPT or another LLM to create five AI-written essays. They mix them up, and present the ten essays as a test for Person B to classify which are AI. Person B can take the test and give their reasoning. This creates a baseline for how detectable these essays are and helps develop heuristics.

To take it a step further, Person B can use their findings to make a more difficult human baseline, prompting the LLM with more input or asking it to avoid the phrases that make it obvious. These back-and-forths can further improve your intuition and also help you learn which AI tells continue to crop up.

For an example human baseline, take a look at these two reviews. One is from Yelp, and one is from GPT-4.

If it's not obvious, the first example is AI and the second one is human. I specified both the topic (Textbook Cafe in Brooklyn) as well as the structure (one paragraph), but ChatGPT's style is so strong that it still comes out obvious. Try to make one of your own and share with a friend!

Use tools built for AI detection

Pangram Labs has dedicated countless hours to training machine learning models that have learned from hundreds of millions of documents to best differentiate AI- and human-written content. So I might as well self-promote a bit. Our dashboard allows you to paste in text or upload documents, so if you're unsure and want something more concrete than intuition to back you up, this will give you a classification and a confidence score. The dashboard also contains deeper analysis that shows which words or phrases provided the most signal in the classification. I personally use our tool all the time when I see something on the internet that rings some alarm bells, and I hope it can be useful to you as well.

Questions? Comments? Found some egregious AI slop and want to share? Reach out on Twitter/X at @max_spero_.