While a whole slew of startups have sprung up over the last few months claiming to detect if text was written by AI, and others are resorting to more unconventional methods, the company responsible for creating most of the text itself doesn’t seem to be able to reliably tell if some text with generated by an AI or not.
OpenAI has quietly shut down a classifier it had trained to distinguish between AI-written and human-written text. The classifier was launched on 31st January this year, but has been shuttered 6 months later. “As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy. We are working to incorporate feedback and are currently researching more effective provenance techniques for text, and have made a commitment to develop and deploy mechanisms that enable users to understand if audio or visual content is AI-generated,” a message on its website said.
Just 6 months ago, OpenAI had announced the launch of the classifier. “We’ve trained a classifier to distinguish between text written by a human and text written by AIs from a variety of providers. While it is impossible to reliably detect all AI-written text, we believe good classifiers can inform mitigations for false claims that AI-generated text was written by a human,” OpenAI had then written.
Even back then, OpenAI had cautioned that its classifier wasn’t completely accurate. “Our classifier is not fully reliable. In our evaluations on a “challenge set” of English texts, our classifier correctly identifies 26% of AI-written text (true positives) as “likely AI-written,” while incorrectly labeling human-written text as AI-written 9% of the time (false positives),” its blogpost had said.
OpenAI’s decision to shut down its AI text detection classifier comes days after top AI companies including OpenAI, Microsoft and others had told the White House that they pledged to watermark content that was generated by AI. The companies had stressed that this was an important safety measure, as it would allow users to determine if the text they were reading was written by an AI or a human.
But OpenAI shutting down its own AI text classifier highlights the challenges that fingerprinting AI text could present. For starters, if fingerprinting techniques hide codes within text which will allow companies to tell if the text was written by AI, it could be easy to break it by simply editing some parts of the text. Additionally, it might be hard to prevent false positives — there have been reports of classifiers claiming that the US constitution was written by AI, perhaps because it formed a part of their training sets. And
And the irony of OpenAI being unable to build an accurate AI classifier wouldn’t be lost on people. OpenAI created ChatGPT, and is able to create AI-written text in seconds, but is still unable to recognize if a piece of text is AI-generated. And these problems might mirror themselves in more significant and serious ways — humanity might end up creating AI, but could have a hard time controlling it in the years to come.