OpenAI Built a Detector, Then Quietly Turned It Off

There’s a short, telling chapter in the history of AI detection that gets less attention than it deserves. It begins in January 2023, two months after ChatGPT launched, with OpenAI publishing a tool to detect AI-written text. It ends in July of the same year, when the tool was withdrawn.

If any organization was positioned to detect ChatGPT’s output, it was the one that built ChatGPT.

What OpenAI released

The tool was a classifier: a model trained to distinguish human-written text from AI-generated text. OpenAI was unusually candid about its limits from day one.

By the company’s own evaluation, the classifier correctly identified only about a quarter of AI-written text as “likely AI-written,” while wrongly labeling human-written text as AI about one time in eleven. OpenAI warned that it was unreliable on short texts, worse in languages other than English, and could be evaded with editing. It said the tool should not be used as a primary decision-making tool.

That’s an extraordinary set of caveats for a product launch. It was, in effect, a research preview with a big warning label.

What happened next

On July 20, 2023, OpenAI updated the original announcement with a short note: the classifier was no longer available, “due to its low rate of accuracy.” The company said it was working on more effective provenance techniques and was committed to helping users understand whether audio or visual content was AI-generated.

There was no press event. Many people who had bookmarked the tool simply found it gone.

Why it matters now

The shutdown is sometimes cited as proof that detection is impossible. That overstates it. Detection methods have improved since 2023, and some commercial tools perform considerably better on long, unedited text than OpenAI’s classifier did.

But the episode established a few things that remain true:

• Building the generator doesn’t make detection easy. OpenAI had every advantage — the model, the data, the research team — and still concluded its classifier wasn’t good enough to keep running.
• Honest error rates are rare. OpenAI published both its detection rate and its false positive rate. Many products since have led with the first and buried the second.
• The hard cases are the real cases. Short text, edited text and non-English text were the classifier’s stated weaknesses. They’re also what most real-world detection involves.

If the people who built the model couldn’t reliably detect its output, the problem isn’t effort.

The pivot to provenance

OpenAI’s note pointed toward a different approach: instead of guessing from the text, attach information about where content came from. That’s the logic behind watermarking and content credentials, which we’ve covered separately.

It’s a quieter idea than a detector with a big percentage on the screen. It’s also a more honest one. A detector asks you to trust its guess. Provenance, done right, lets you check.