OpenAI shut down its AI Text Classifier on 20 July 2023 and said plainly that it did so because of the tool's low rate of accuracy. It has not come back. If you are searching for that classifier, you need a replacement, and OpenAI's own published limitations are the best checklist for picking one.
Paste some text or load the example to estimate AI likelihood.
OpenAI announced it on 31 January 2023 in a post titled "New AI classifier for indicating AI-written text," credited to Jan Hendrik Kirchner, Lama Ahmad, Scott Aaronson and Jan Leike. OpenAI described it as a free, publicly available web app and invited readers to "try our free work-in-progress classifier yourself." You gave it text, and it returned a verdict such as "likely AI-written."
Technically, OpenAI described it as "a language model fine-tuned on a dataset of pairs of human-written text and AI-written text on the same topic." The human side came from sources OpenAI believed were human-written, including its pretraining data and human demonstrations on InstructGPT prompts. Each text was split into a prompt and a response, and responses were then generated from a range of models, OpenAI's own and other organisations'.
One design decision matters more than the rest, because most people never saw it. OpenAI wrote: "For our web app, we adjust the confidence threshold to keep the false positive rate low; in other words, we only mark text as likely AI-written if the classifier is very confident." The tool was deliberately tuned to under-accuse.
The notice OpenAI added to its own announcement is one sentence, and it is unambiguous:
"As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy."
The numbers behind that decision were published by OpenAI from the start, six months before the shutdown. In its own evaluation on a challenge set of English texts, the classifier "correctly identifies 26% of AI-written text (true positives) as 'likely AI-written,' while incorrectly labeling human-written text as AI-written 9% of the time (false positives)."
Those are OpenAI's figures for OpenAI's tool, reported here as theirs. Read them together with that conservative threshold and the picture is coherent rather than mysterious: a detector tuned hard against false accusations will miss most of the AI text it sees. OpenAI never hid this. The announcement stated flatly: "Our classifier is not fully reliable."
The same notice added that OpenAI was "currently researching more effective provenance techniques for text," and separately committed to mechanisms for identifying AI-generated audio and visual content. Checked on 16 August 2026, that page still frames text provenance as ongoing research and does not point readers to a replacement classifier.
If you found this page through the domain aitextclassifier.com, be aware of what that address is. It is not OpenAI's, it never was, and TextSight has no connection to OpenAI.
Checked on 16 August 2026: the domain was registered on 20 April 2026 through DropCatch.com, its registrant organisation is HugeDomains.com, and requests to it redirect to a HugeDomains sale listing for the name. Its HTTPS endpoint does not present a working certificate at all. There is no product there, no paste box and no company behind it. It is a domain for sale that happens to carry the name of a tool people still search for.
Nowhere, because it no longer runs. That is the honest answer, and anyone telling you otherwise is describing a tool they have not opened.
It is still worth saying what it did well, because those qualities are rarer than they should be. It was free with no signup and no upsell. It came from the organisation that built the model it was detecting. Most importantly, it was unusually candid: the announcement led with its own weakness, published its own poor accuracy figures, carried an explicit Limitations section, and told readers that "it should not be used as a primary decision-making tool, but instead as a complement to other methods of determining the source of a piece of text."
Very few detectors on the market today publish anything like that. If you valued the classifier for being honest about itself rather than for being right, you have lost something real, and the sensible response is to hold its replacement to the same standard.
The limitations OpenAI listed still apply to detectors built the same way. Use them as questions to ask of anything you are considering.
OpenAI: "The classifier is very unreliable on short texts (below 1,000 characters)." Its reliability "typically improves as the length of the input text increases." Any tool that returns a confident verdict on two sentences is telling you more about its interface than about your text. TextSight requires a minimum of 25 words before it will score anything, and longer passages give the ensemble more to work with.
OpenAI: "Sometimes human-written text will be incorrectly but confidently labeled as AI-written by our classifier." A 9% false positive rate on a challenge set is not an abstraction when the text belongs to a student or a freelancer.
TextSight's detector is an English-only ensemble of DeBERTa, RoBERTa and ELECTRA models, and it carries a known false-positive bias against writers using English as a second language. We would rather state that than let you discover it in a disciplinary meeting. Treat any score as one input, exactly as OpenAI advised.
OpenAI: "We recommend using the classifier only for English text. It performs significantly worse in other languages and it is unreliable on code." The same constraint applies to TextSight. Our detector is English-only. If your text is in another language, no honest score is available from either tool.
OpenAI's example is a good one: nobody can tell whether a list of the first 1,000 prime numbers was written by a person or a model, because there is only one correct answer. Formulaic writing, boilerplate and heavily conventional structures sit in the same category. OpenAI also noted that neural classifiers are "poorly calibrated outside of their training data" and can be "extremely confident in a wrong prediction" on unfamiliar input.
TextSight is a paid AI detection and rewriting product, so it is a different kind of thing from a free research demo, and it is worth being clear about what you actually get.
The paste box on this page runs the detector without an account. Anonymous checks are limited to 3 per day, pooled across the AI Detector and the Document Detector, with a 1,000 word cap per check. A free account raises that to 60 checks per day at 5,000 words, and every scan you run while signed in is saved to your Scan History, including on the free tier. Starter allows 200 checks per day. Pro and above are unlimited.
Scores are reported in three bands rather than as a single number pretending to precision: under 30 is "likely human-written," 30 to 69 is "mixed signals," and 70 or above is "likely AI-generated."
Paid tiers add the parts a free demo never had. Sentence-level analysis and PDF export start at Starter. A Chrome extension and a REST API start at Pro. Teams and branded reports start at Business. File upload, URL scanning and bulk scanning are available on paid plans.
Monthly, or with 25% off when billed annually:
https://openai.com/index/new-ai-classifier-for-indicating-ai-written-text/. Checked 16 August 2026. Because openai.com returns HTTP 403 to automated requests, the text was verified against an Internet Archive capture of OpenAI's own page dated 3 February 2024, which carries the withdrawal notice.https://www.hugedomains.com/domain_profile.cfm?d=aitextclassifier.com. Checked 16 August 2026.The deeper head-to-head: feature table, where GPTZero wins, where TextSight wins, and migration steps.
Read the comparePerplexity, burstiness, and transformer scoring explained, so you can read any detector's verdict critically.
Read the guideWhy detectors over-flag ESL and formal writing, and how sentence-level evidence helps you defend real work.
Read the guideThe full ranking with detection approach, pricing, and use-case fit side by side.
See the rankingStart with TextSight's free tier. No card, no signup, no commitment. Your first scan in about six seconds, with the evidence shown line by line.
Honest comparisons vs other tools.