With the rise of students using ChatGPT and other large language model tools to cheat on their school and college assignments and even write exams now, there is an urgent need for an algorithm that can detect whether the text is created by an AI or a human. To this end, a team of PhD students and computer scientists have released a new tool called DetectGPT that aims to identify whether a given piece of text is created by an LLM. 

Using a statistical method known as probability curvature, the researchers have found a method to detect AI-generated text. This approach reportedly does not require watermarking of generated text or even a dataset of real vs generated text but instead uses statistics to determine the likelihood of text generation. 

When tested on articles generated by GPT-NeoX, DetectGPT was able to find the AI-generated text 14 out of 15 times. It was even able to detect AI-generated text when text paraphrased by humans replaced up to 25% of the text. It does so by determining the log probability of generated text on a per-token basis. Simply put, per-token log probability is a method of determining the efficacy of an LLM, as it is an indicator of how ‘good’ the LLM is at statistically predicting the next word. 

By comparing the log probability of a source passage to the log probability of the same passage with multiple changes, the team was able to pin down whether the passage was created by an algorithm or not. Text created by an LLM has a ‘negative curvature’ of the log probability function, while human-generated texture tends to have a ‘positive curvature’ in the same parameter. Notably, this approach needs a white-box approach to the model to perform at maximum efficiency, as a known model provides a more accurate per-token log probability parameter. 

This method differs from training another deep learning algorithm to detect AI-generated text, as those methods commonly result in overfitting. These models also need to be trained as newer versions come out, further diluting their usability. The approach assumes the text has been modified to bypass detection methods and instead goes to the source of LLM sampling strategies to detect generated text. 

The full criteria for determining the efficacy of the detection approach has been detailed in the paper. However, the key takeaway is that this method provides an accurate detection performance without the need for stop-gap solutions like watermarking or pre-trained neural networks. On the other hand, the scientists also suggested that combining DetectGPT with other approaches like watermarking may increase the overall likelihood of discovering AI-generated text.

The rise of methods to find out whether a piece of written content is generated by AI is an important step in preventing misinformation. Teachers can also use applications built on methods like DetectGPT and watermarking to accurately determine the origin of an essay. In the future, these kinds of methods may also be used to determine human involvement in written content — an important piece of the puzzle in the age of AI-generated content.