ChatGPT text detection tool exists, but OpenAI won’t release it: report
OpenAI has reportedly built a tool that can detect when text has been generated using its artificial intelligence (AI) chatbot, but is not ready to release it publicly. According to the report, the tool has been ready to be released for some time, but the AI company is concerned that such a detection tool could make the chatbot unpopular among users. However, the failure to release such a tool has left teachers struggling to figure out when an assignment or essay was written using AI.
OpenAI reportedly has an AI text detector
According to a report by The Wall Street Journal, OpenAI has been debating the release of such a tool for two years. The text detection tool has also been ready for release for about a year, the report claimed, citing unnamed people familiar with the matter. One of the sources told the publication that releasing the tool was as easy as pressing a button.
Such a tool could help teachers and other similar institutions where humans have resorted to generating content such as essays and research papers using AI. Earlier this year, a peer-reviewed scientific paper paper published in Elsevier’s Surfaces and Interfaces journal was found to be written by AI and was subsequently retracted after it gained online attention. This AI-based plagiarism has become a major concern for academia due to the lack of a reliable method to detect AI use.
According to the report, OpenAI’s refusal to release its AI detection tool stems from fear of losing existing users. The company reportedly conducted a survey and found that nearly a third of users would be less likely to use ChatGPT if an anti-cheating mechanism were released.
Another fear was that making the watermarking technology available to a select group of users, such as teachers, could reduce the usability of the tool. If the watermarking technology were made available to a large group of users, it could lead to malicious actors being able to decipher the technology and create sophisticated masking tools.
Based on internal documents reviewed by the publication, the report claims that the AI text detection tool is 99 percent effective at finding text written using ChatGPT. This allegedly works because the tool is essentially a watermarking technology that embeds the text with an invisible watermark that cannot be seen, but is highlighted when run through the AI checker.
The tool’s operation was also explained in internal documents. According to the report, ChatGPT generates text by predicting which word or phrase, also known as a token, should follow in a string. The selection is based on a small pool, so that the sentence is coherent.
The AI detection tool, in turn, has a slightly different algorithm for selecting the tokens. The changes between ChatGPT generated text and the tool text leave a pattern, which helps in judging whether AI was used to generate the text or not.
