How-To

Facts About OpenAI's AI Text Classifier (and Why AI Detectors Are So Unreliable)

FFacts On Tap Editorial Team Fact-checked Updated October 9, 2025 3 min read

In early 2023, OpenAI released a tool meant to spot AI-written text — and just six months later, the company pulled the plug on it. That short, messy history says a lot about why every "AI detector" you've seen since should be treated with a healthy dose of skepticism.

Here are the verified facts about OpenAI's AI Text Classifier, how AI detection tools work, and why trying to "beat" them is a much shakier idea than the internet makes it sound.

OpenAI launched its AI Text Classifier on January 31, 2023, as an experimental tool to help distinguish human-written text from AI-generated text.
OpenAI discontinued the classifier on July 20, 2023, citing its 'low rate of accuracy.'
By OpenAI's own numbers, the tool correctly flagged only about 26% of AI-written text as 'likely AI-generated.'
The classifier also incorrectly labeled genuinely human-written text as AI-generated about 9% of the time (a false positive).
Independent researchers found that AI detectors, including OpenAI's, were especially likely to misclassify writing from non-native English speakers as AI-generated.
The AI Text Classifier worked by estimating a probability score, not by giving a definitive yes-or-no answer, which made its results easy to misread as more certain than they were.
OpenAI said it was discontinuing the tool while it researched 'more effective provenance techniques,' such as cryptographic watermarking of AI-generated content.
OpenAI Text Classifier is a specific, now-retired OpenAI product — it is not the same thing as the general-purpose GPT models like ChatGPT.
Most modern AI-detection tools (including those still on the market) work by analyzing statistical patterns in word choice and sentence structure, not by directly 'reading the mind' of the writer.
Because AI detectors rely on probability and pattern-matching rather than certainty, none of them — including tools marketed by third-party companies — can reliably prove a specific piece of text was or wasn't written by AI.
Simple word-swapping or inserting unusual Unicode characters into text does not reliably fool modern AI detectors and can also make writing look garbled or spammy to human readers and search engines.
Academic institutions including Vanderbilt University and the University of Pittsburgh publicly advised instructors against relying on AI-detection tools for grading decisions because of high false-positive rates.
Turnitin, a major plagiarism-detection company used by schools, has acknowledged its AI-writing indicator can produce false positives and has cautioned against using its score as the sole basis for academic discipline.
GPTZero, one of the best-known third-party AI detectors, was originally built by a Princeton student in early 2023 in direct response to concerns about AI-generated schoolwork.

Frequently asked questions

Is OpenAI's AI Text Classifier still available?

No. OpenAI discontinued it on July 20, 2023, less than six months after launch, because of its low accuracy rate.

Can AI detectors accurately tell if ChatGPT wrote something?

Not reliably. Even the best-known tools produce both false positives (flagging human writing as AI) and false negatives (missing real AI writing), so results should be treated as a rough signal, not proof.

Do special characters or unicode substitutions help text 'pass' as human-written?

Not reliably, and it can backfire — garbled substitutions often look like spam to search engines and readers, and modern detectors are generally not fooled by simple character swaps.

Why did OpenAI shut its detector down instead of improving it?

OpenAI said it wanted to focus on more effective, longer-term approaches to identifying AI-generated content, such as cryptographic watermarking, rather than keep a tool with a documented high error rate available to the public.

Should schools or employers rely on AI-detection scores to make decisions about someone?

Most experts and even detection vendors themselves advise against using a single AI-detection score as definitive proof, given documented false-positive rates and bias against non-native English writers.

Sources & further reading

Reviewed by the Facts On Tap Editorial Team

We write every article to be genuinely fun — and genuinely true. Each fact is checked against reputable sources before it goes live. Spotted something off? Tell us and read our fact-check policy.