OpenAI’s AI Classifier is no longer available. OpenAI says it removed the experimental web tool on July 20, 2023 because of its low accuracy. The classifier was never a definitive authorship test: in OpenAI’s English-language challenge-set evaluation, it identified 26% of AI-written samples as “likely AI-written” and incorrectly labeled 9% of human-written samples as AI-written.
What OpenAI shut down
The retired product was called AI Classifier, a public, work-in-progress web service announced by OpenAI on January 31, 2023. It was designed to estimate whether text had been written by a human or by an AI system from a range of providers.
The interface returned graded judgments rather than a forensic verdict. Contemporaneous Associated Press coverage described labels running from “very unlikely” to “likely” AI-generated. OpenAI’s announcement said the release was intended partly to gather feedback on whether an imperfect classifier could still be useful.
OpenAI’s announcement now carries the shutdown notice: “As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy.” That is the company’s explanation for this specific tool; it does not establish that every AI-detection product performs identically.
#1 Best Overall
How accurate was the classifier?
OpenAI reported two separate results from a challenge set of English texts. They should not be merged into one overall accuracy percentage.
| Measure | OpenAI’s reported result | What it means |
|---|---|---|
| True positives | 26% | Share of AI-written samples that the classifier identified as “likely AI-written.” |
| False positives | 9% | Share of human-written samples that the classifier incorrectly labeled as AI-written. |
These figures came from OpenAI’s particular evaluation set, not a universal test across languages, genres, text lengths, or newer models. A 26% true-positive figure means most AI-written samples in that challenge set were not flagged at the stated threshold. A 9% false-positive figure means a meaningful minority of human writing could be accused incorrectly.
Rank #2
How OpenAI said it worked
OpenAI described a language model fine-tuned on pairs of human-written and AI-written text about the same topic. Researchers collected material they believed was human-authored, then generated paired responses with models from OpenAI and other organizations. The announcement did not provide a complete technical specification or explain an internal chain of reasoning that could verify authorship.
Because the system learned statistical patterns from its training examples, its output was a probability-style classification. It was not a record of who actually typed a passage, and it could not establish intent or prove that a person had used a particular chatbot.
Free tools Windows power users keep installed
One-click scans. No signup required.
Why the tool was risky in real decisions
OpenAI’s own limitations list explains why a detector score should not be treated as proof:
- Short text: performance was very unreliable below 1,000 characters, and longer passages could still be mislabeled.
- Language: OpenAI recommended English-only use because results were significantly worse in other languages.
- Code: the classifier was unreliable on programming code.
- Evasion: editing could help AI-generated text avoid detection.
- Predictable writing: formulaic or highly predictable text was difficult to attribute reliably.
- Calibration: neural classifiers can behave poorly when an input differs from the data used to train them.
OpenAI advised using the classifier only as one input alongside other evidence, never as the primary decision-making tool. That warning matters wherever an uncertain label could cause harm—for example, a school disciplinary case, an employment decision, a publishing dispute, or an allegation of academic dishonesty. A detector result should prompt a fuller review of drafts, revision history, citations, oral explanation, and other context, rather than serve as a verdict.
What the classifier was intended to detect
OpenAI presented several possible uses: helping challenge false claims that AI-written material was human-written, identifying possible automated misinformation campaigns, supporting investigations of academic dishonesty, and spotting cases where an AI chatbot was presented as a person. These were proposed use cases, not a guarantee that the tool could reliably settle any of them.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What happens to OpenAI’s detection work now?
OpenAI said it was researching more effective text-provenance techniques and hoped to share improved methods in the future. That statement describes an ongoing research direction; it does not confirm that a replacement text classifier was subsequently launched. The retirement of AI Classifier therefore answers the availability question clearly—this web tool is gone—without establishing that provenance research has ended.
Quick Recap
What readers should conclude
- OpenAI’s AI Classifier was announced on January 31, 2023 and became unavailable on July 20, 2023.
- OpenAI attributed the shutdown to low accuracy.
- Its published challenge-set results were 26% true positives and 9% false positives for English text.
- The company documented serious weaknesses involving short passages, non-English text, code, predictable writing, edited AI text, and out-of-distribution inputs.
- Its score was an uncertain signal about text patterns, not proof of authorship.
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
