Over the past few months, as the market for AI detection has steadily grown, a similar market for ways to evade them has emerged.

Today, there are a whole class of tools available on the market that call themselves “humanizers”. Under the hood, each of these tools is a little bit different, but they all add up to a single promise: if you use our tool, no one will be able to tell that your text was written by AI. These tools have quickly become popular among students, SEO professionals, and others hoping to evade detection of their AI use.

However, we’re excited to announce that today we have shipped a model that can detect over 90% of high-quality humanized content.

Results

Here’s how our new model performs:

To achieve these results, our team underwent a rigorous research process that involved closely auditing 19 publicly available humanizers, creating a classification system to rank their quality, and retraining our production model with high-quality humanized text included in it's training corpus.

This work was recently featured at the Workshop on Detecting AI-Generated Content at the International Conference on Computational Linguistics. To read the technical paper, click here.

For more information, demo requests, or other inquiries, reach out to info@pangram.com!