OpenAI has decided to discontinue its AI text classifier, a tool originally designed to assess the likelihood of a text passage being authored by another AI. Despite being used by many to detect low-effort forgeries, OpenAI decided to retire the tool due to its much-criticized “low rate of accuracy.”
The assumption that AI-generated text possesses some identifiable characteristic or pattern that can be consistently detected seems plausible. However, the reality so far has not matched the theory. Although some AI-produced text may exhibit telltale signs, the swift development and varied nature of large language models have rendered these signs unreliable.
TechCrunch's own evaluation of various AI-written-detection tools found them to be, at best, inconsistent, and at worst, entirely futile. Out of seven AI-generated text snippets subjected to multiple detectors, GPTZero correctly identified five, while OpenAI’s classifier managed to detect only one. This was achieved with a language model that was not even at the cutting edge during the test.
Regrettably, some accepted the claims of detection at face value, even though OpenAI launched the classifier tool with a list of significant limitations that called into question its usefulness. Individuals concerned about the authenticity of texts submitted by students, job applicants, or freelancers used the classifier for validation. Despite the tool's questionable reliability, its results were sometimes accepted without skepticism.
As language models continue to advance and become more widespread, it appears that OpenAI deemed it time to retire this capricious tool. The company announced on July 20, "We are working to incorporate feedback and are currently researching more effective provenance techniques for text."
It's noteworthy that OpenAI's decision coincided with its involvement in a White House–led "voluntary commitment" to develop AI ethically and transparently, alongside several other companies. This commitment includes attempts to develop reliable watermarking and/or detection methods. However, so far, we have yet to see any watermark or detection method that cannot be easily circumvented.
Undoubtedly, the first company to succeed in developing a truly reliable detection tool will be rewarded handsomely, as such a tool would be invaluable in numerous situations. This fact alone likely makes it unnecessary to include in any AI accords.
