Persistent Challenges and Humbling Realizations: OpenAI Acknowledges the Complexity of Plagiarism Detection in ChatGPT
OpenAI, the creators of ChatGPT for artificial intelligence (AI), has locked down a tool it developed to detect content generated by AI rather than humans. The tool, dubbed AI Classifier, was shut down just six months after launch due to its “low accuracy rate,” OpenAI said.
Since the popularity of ChatGPT and competing services has skyrocketed, there has been a coordinated reaction from various groups concerned about the consequences of using unmonitored AI. For one thing, teachers were particularly annoyed that students could use ChatGPT to write their essays and assignments, and then pass them off as their own.
OpenAI’s AI classifier was an attempt to assuage the concerns of these and other groups. The idea was that it could determine whether a piece of text was written by a human or an AI chatbot, giving people a tool to rate students fairly and to combat misinformation.
However, even from the start, OpenAI didn’t seem to have much confidence in its own tool. In a blog post announcing the tool, OpenAI declared that “our classifier is not completely reliable”, stating that it correctly identified AI-typed texts from a “challenging set” only 26% of the time.
The decision to drop the tool didn’t get much fanfare, and OpenAI didn’t publish a dedicated post on its website. Instead, the company updated the post in which it revealed the AI classifier, stating that “the AI classifier is no longer available due to its low accuracy rate.”
The update continued: “We are working on integrating feedback and are currently researching more efficient source technologies for text, and we have committed to developing and deploying mechanisms that enable users to understand whether audio or visual content is generated by AI.”
Better tools are needed
AI Classifier isn’t the only tool developed for AI-driven content discovery, as competitors like GPTZero exist and will continue to work, despite OpenAI’s decision.
Previous attempts at AI handwriting recognition have backfired in spectacular fashion. For example, in May 2023, an entire class professor accidentally enlisted ChatGPT to detect plagiarism in their students’ papers. Needless to say, ChatGPT got a lot wrong, and so did the prof.
It’s a cause for concern even when OpenAI admits that it can’t properly recognize the plagiarism generated by its chatbot. This comes at a time of growing concern about the disruptive potential of AI-powered chatbots and calling for a temporary suspension of development in this area. If AI has as much of an impact as some people expect, the world will need tools more powerful than OpenAI’s failed AI rating tool.