How I Improved the GPT2 Output Detector

The GPT2 Output Detector is a powerful classifier that detects ChatGPT text with an impressive accuracy. I extended the detector’s frontend to work for large texts, support custom file uploads, and allow parallel processing of many files.

These new features make the GPT2 detector more usable and accessible. In this blog post I will describe my changes and how they work.

You can try out the application at: https://michaelgithubhype.github.io/Extended-GPT2-Output-Detector/index.html

Here is a full demo:

Background

This project is based on my previous post where I tested the accuracy of the GPT-2 Output Detector. I found that the tool is accurate — it effectively detects the difference between human written vs ai-generated text.

However, there is an issue. The GPT2 Output Detector is limited to 510 tokens. For long posts, the tool is sensitive to small changes at the beginning. This restricts content detection support to social platforms with short texts. Ideally, the tool would take a holistic view of the content.

The problem with the GPT-2 Output Detector is that the API limits the amount of allowed tokens. Here is the response format once a user sends more than 510 tokens: