Are AI detectors accurate? The newest research says they punish honesty.

The short answer

No, not accurately enough to accuse anyone on.

That is not a contrarian take. It is the position of the company that built ChatGPT, which quietly killed its own detector, and it is what the newest peer-reviewed and preprint research keeps finding. Detectors produce a number that looks like evidence and is closer to a weather forecast, and the people most likely to be wrongly flagged are the ones least able to argue back.

Here is what the actual studies say, and what the vendors say about themselves when you read past the marketing page.

OpenAI shut down its own detector and published the reason

In January 2023, OpenAI released an AI Classifier to spot AI-written text. The page is still online, and it now carries this notice: "As of July 20, 2023, the AI classifier is no longer available due to its low rate of accuracy."

The numbers behind that decision are on the same page. 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)."

Read those together. It caught about a quarter of the AI text it was shown, and wrongly accused nearly one in eleven human writers. The limitations list is more damning than the score, including the flat admission that "sometimes human-written text will be incorrectly but confidently labeled as AI-written by our classifier."

Confidently. That is the whole problem in one word. These tools do not return "unsure," which is the honest answer most of the time.

What the 2026 research found, and it is worse than inaccuracy

The most useful recent work does not just measure accuracy. It asks who gets caught.

An August 2026 study by Jonathan A. Karr Jr, Grigorii Khvatskii, Ting Hua and Nitesh V. Chawla tested commercial detectors against the way students actually use AI. Light AI editing, the kind most university policies explicitly permit, was flagged as AI-written between 38% and 80% of the time. Unmodified academic abstracts from 2023 to 2025, written before or without AI, were flagged 9% to 15% of the time. Text that had been deliberately laundered through humanizer tools was detected less than 4% of the time, meaning the tools missed over 96% of actual evasion.

The authors put the conclusion bluntly: "Honest AI-editing results in a higher sanction risk than humanizer-assisted evasion."

Sit with that. Under a detector-based regime, a student who follows the rules and discloses light AI assistance is in more danger than one who runs the whole essay through a paraphraser. The tool inverts the incentive it was bought to enforce. The same study also found that flagging rates were significantly higher for non-STEM than STEM writing.

Benchmark scores do not survive contact with reality

Detector vendors quote impressive accuracy figures. Those figures are usually real and usually irrelevant, for a reason a March 2026 paper by Shushanta Pudasaini, Luis Miralles-Pechuán, David Lillis and Marisa Llorens Salvador pinned down precisely.

Their detector hit an F1 score of 0.9734 on standard benchmark datasets, then degraded significantly under cross-domain evaluation. Using SHAP-based explanations to see what the model was keying on, they found "the most influential features differ markedly between datasets, indicating that detectors often rely on dataset-specific stylistic cues rather than stable signals of machine authorship."

In plain terms: the detector learned what the test set looks like, not what machine writing is. Move it to a new subject, a new writing style, or a newer model, and the score it earned on the benchmark no longer applies to your essay. If you have read our explainer on how AI benchmarks work and why leaderboard scores mislead, this is that problem in its most consequential form, because the output is a disciplinary hearing rather than a press release.

The bias is documented and it has a name

In 2023, Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou published work in Cell Patterns that remains the single most important finding in this field. Their conclusion: "These detectors consistently misclassify non-native English writing samples as AI-generated, whereas native writing samples are accurately identified."

The mechanism is not mysterious. Detectors largely measure how predictable text is. Writing in a second language tends to use more common words and simpler constructions, which is exactly what a predictability model reads as machine-generated. The clearer and more careful a non-native speaker's prose, the more suspicious it looks.

Three years on, the 2026 work above still finds systematic skew by discipline. Nothing has fixed this, because the thing measured is a proxy for fluency, not a fingerprint of authorship.

What the vendors admit when you read their own pages

To Turnitin's credit, it publishes its error rates rather than only its accuracy. Annie Chechitelli, its chief product officer, wrote in June 2023 that the document-level false positive rate is "less than 1% for documents with 20% or more AI writing," while "our sentence-level false positive rate is around 4%."

Note the conditional buried in the first number. The sub-1% figure applies to documents that already contain 20% or more AI writing, so it is not a claim about a fully human essay. And at 4% per sentence, a 40-sentence essay will contain wrongly flagged sentences as a matter of arithmetic, not malfunction. Turnitin says 54% of false-positive sentences sit directly next to real AI writing, which is useful for interpreting a mixed document and no comfort at all to someone who wrote every word.

Our take

Treat an AI detection score as a prompt to have a conversation, never as proof. The evidence supports exactly one defensible use: a flag that makes a human being look more closely at the work, the drafts, and the writer's other submissions. It supports no version of a score as a verdict.

If you are an educator, the most robust available method is the one detectors were supposed to replace: process evidence. Version history, outlines, in-person discussion of the argument. If you have been wrongly flagged, the strongest material is your draft history, and the strongest arguments are on the record and citable, including OpenAI's own retirement notice and the 2026 finding that compliant AI use is flagged more often than laundered text.

The deeper point is that detection is the wrong shape of solution. A model trained to predict text cannot reliably identify text from a model trained to predict text, and it fails hardest on careful, plain, second-language prose. It is the same reason AI systems state false things confidently: confidence and correctness are separate properties, in the detector as much as in the thing it detects.