OpenAI kills its AI detection tool over terrible accuracy

Back in January, OpenAI launched a tool meant to save the sanity of educators worldwide: an official AI Classifier that could sniff out whether an essay was written by ChatGPT or an actual human.

Six months later, OpenAI quietly pulled the plug. The company disabled the tool on July 20, adding a low-key update to its original launch post stating that the classifier was shut down due to a "low rate of accuracy."

The math never worked

From day one, the tool struggled to do the very thing it was built for.

According to OpenAI's own benchmark data, the classifier correctly flagged AI-written English text just 26% of the time. Worse, it carried a 9% false-positive rate, meaning roughly one in every ten human-written pieces was incorrectly accused of being generated by an AI.

It was also exceptionally picky. The system required at least 1,000 characters to make a guess, performed terribly on non-English text, and fell apart whenever inputs drifted outside its training data.

Even as OpenAI warned users not to use it as a primary decision-making tool, desperate educators clung to it. But in a classroom setting, a 9% chance of wrongfully accusing an honest student of cheating is simply unacceptable.

Back to the drawing board

When OpenAI first launched the detector alongside research with Stanford and Georgetown faculty, it warned that automated misinformation and cheating were rapidly scaling threats. That paper noted that no reasonable mitigation could fully prevent AI influence operations.

Now, OpenAI says it is researching more effective text provenance techniques. It has also committed to developing mechanisms to identify AI-generated audio and visual content, like images from DALL-E. However, the company gave no timeline for when an improved text solution might return.

Why it matters

If the creator of ChatGPT cannot reliably identify its own text, nobody else can either.

For teachers, employers, and content moderators, this should be a definitive wake-up call to stop relying on AI detectors. The cottage industry of third-party detection apps that popped up over the last year is built on shaky ground; if OpenAI maxed out at a 26% catch rate, third-party claims of near-perfect detection are borderline fiction.

We think shutting the tool down was the responsible move. Keeping a tool with a 9% false-positive rate live only invited false accusations and unnecessary stress. But quietly burying the shutdown in an edited blog post shows just how awkward this technological bottleneck is for frontier labs.

Ultimately, post-hoc statistical detection looks like a dead end. The future will have to rely on watermarking and cryptographic provenance at the point of generation, though even those will struggle against open-source models.