Since OpenAI launched ChatGPT in November last year, the potential of ChatGPT is a question of great concern to people. In addition to helping users answer questions and write code, ChatGPT also passed the Wharton MBA exam last week (although In the face of such results, Wharton professors believe that it is necessary for the school to make reforms in examination policy, curriculum design and teaching.

However, if you want to change the curriculum and teaching, it is not something that can be done overnight. So is there any way to meet the challenges of AI tools such as ChatGPT and prevent students from using AI to complete homework, cheat in exams, or generate papers?

Recently, Stanford University has developed a model called DetectGPT, which can detect whether a paragraph is generated by a given large language model (LLM).

According to the research team, text generated by LLM usually hovers in a specific region of the negative curvature region of the model’s log probability (log probability) function (inference: text written by humans may lie in positive curvature, the announcement does not have this to explain). Using this observation, they defined a new curvature-based criterion to judge whether the text was generated by LLM. The team called this new method “zero-shot”, which allows DetectGPT to know nothing about the text used to generate the text. AI can also do text detection without training or relying on large data sets.

According to the Stanford development team, DetectGPT is more discriminative than existing similar methods, and its evaluation results have also improved significantly when it undertakes the task of detecting fake news articles.

As public acceptance of LLMs increases and usage grows, the importance of corresponding systems for detecting machine-generated text will also increase.

At present, the development team only released a general introduction of the project on the website, and said that the code, data and other content of the project will be released soon (coming soon).