OpenAI's AI text classifier, 2023: the published accuracy record

This record is one of eight in the studies index and the only one written by a vendor about its own withdrawn tool. The classifier was tested by four of the other studies indexed here before OpenAI pulled the tool for low accuracy.

One detector: OpenAI's AI Text Classifier, offered at platform.openai.com/ai-text-classifier from January to July 2023 and described by OpenAI as a language model fine-tuned on pairs of human-written and AI-written text on the same topic.

Detector as described on the OpenAI 2023 announcement page, openai.com.

The evaluation sample is a "challenge set" of English texts; OpenAI states no size or composition for that set, so this record carries none.

Sample as described in the OpenAI 2023 announcement, openai.com.

The metrics are a true positive rate (AI text marked "likely AI-written") and a false positive rate (human text marked AI). The web app's threshold was set to keep false positives low, so text was marked likely AI-written only when the classifier was very confident.

Two numbers, as published on the OpenAI 2023 page (openai.com).

Values as published in the OpenAI 2023 announcement, openai.com.

OpenAI's own wording in the 2023 announcement is that the classifier "is not fully reliable", identifying 26% of AI-written text while mislabelling human text 9% of the time (OpenAI announcement). The post adds that reliability typically improved as input length increased.

The update came on the same page. As of July 20, 2023, OpenAI 2023 states, the classifier was no longer available "due to its low rate of accuracy", with the company turning to research on provenance techniques for text. No later accuracy figure was published for the tool (openai.com).

Independent records in this index tested the classifier before the withdrawal. Weber-Wulff 2023 scored the OpenAI Text Classifier at 29 of 54 documents correct (54%) (DOI 10.1007/s40979-023-00146-z). Walters 2023 recorded 78% correct on 126 documents, with a footnote that the tool was discontinued on July 20, 2023 (DOI 10.1515/opis-2022-0158). Elkhatat 2023 found the classifier flagged all 5 human controls, and Liang 2023 included the OpenAI detector among seven tools with a 61.22% average false positive rate on TOEFL essays. Each figure is sourced on its own record page. Liang 2023 names its OpenAI tool by the address openai-openai-detector.hf.space, which is the hosted demo of the GPT-2 Output Detector, OpenAI's separate 2019 RoBERTa model; that card carries the Liang result, and the classifier of 2023 was a different tool.

The withdrawal note is one of seven primary documents the results as evidence question, added 9 November 2026, reads on acting on a detector score, and the minimum word count question reads its 1,000-character reliability line beside the vendor floors.

OpenAI's post says the tool "should not be used as a primary decision-making tool". OpenAI 2023 calls the classifier very unreliable on texts below 1,000 characters, and longer texts were sometimes mislabelled too (openai.com).

Human-written text could be labelled AI "incorrectly but confidently". OpenAI recommended English only, saying the classifier performed significantly worse in other languages and was unreliable on code. Very predictable text could not be reliably identified, AI-written text could be edited to evade the classifier, and neural-network classifiers were acknowledged to be poorly calibrated outside their training data.

A vendor that publishes a 26% catch rate, as OpenAI did in 2023, and then withdraws the tool for low accuracy has written the clearest caveat in this index, and in its own words (openai.com).

No detector card: the classifier is withdrawn and has no entry in the index. This record is linked instead from the four study records that tested the tool, and the question page on withdrawn detectors sets this withdrawal beside Writer's sunset of 22 December 2025 and the paused GPT-2 demo.

The ranking of detectors by independent test record covers current tools only. How this record is sourced explains why a vendor statement is indexed with its date rather than as a card score, and the home page lists which AI detectors exist and who has tested them.