How it works
AI Detector Accuracy and Limits
AI detector accuracy explained: how accurate AI detectors are, where they fail, and how to use a result responsibly.
Last updated Oct 29, 2026
How accurate are AI detectors?
No AI detector is always right. AI detection looks for statistical patterns. It can miss AI-generated content and it can flag human work. We publish accuracy figures only for checks we have measured on our own test set, and we describe the test so you can judge it. So far that is the free quick check, below. Be wary of any tool that claims near-perfect accuracy without saying how it was measured.
Measured: the free quick check (October 2026)
We tested the quick check's model on texts we collected ourselves. The table below uses the compressed model exactly as the site runs it.
- Human writing: 1,141 English texts written before mid-2021, before AI chatbots were common. They include Wikipedia, Wikinews, Wikibooks and Wikivoyage pages, Stack Exchange answers, arXiv abstracts and questions from English learners.
- Essays by English learners: 879 essays of 150 words or more. They come from two openly licensed corpora, ArCC (Arabic-speaking school and university students) and SCoCESLE (adult learners in Colombia).
- AI writing: 270 texts from ChatGPT, Gemini, Grok, Claude, DeepSeek, GLM (z.ai), Doubao and Meta AI, on the same topics as the human texts.
- Rewritten AI writing: 80 of those texts rewritten by a second chatbot, either paraphrased or told to sound human and avoid detection.
| Texts | Likely AI-generated | Can't tell | Likely human-written |
|---|---|---|---|
| Human writing (1,141) | 0.3% (3) | 0.6% | 99.1% |
| Essays by English learners (879) | 0.1% (1) | 2.8% | 97.0% |
| AI writing (270) | 86.7% | 5.6% | 7.8% |
| Rewritten AI writing (80) | 71.2% | 13.8% | 15.0% |
What this means:
- False alarms are rare but not zero. About 3 in 1,000 human texts were called AI. A result is evidence, not proof.
- It misses some AI text. About 1 in 8 AI texts, and more after rewriting, were not called AI. Text written by Grok, Claude, Meta AI and DeepSeek was caught less often (73–77%) than text by ChatGPT, Gemini and Doubao (98–100%).
- Mixed text gets one answer for the whole. We joined human and AI passages into one text. When half of a text was AI-written, the verdict was "Likely AI-generated" 36–49% of the time and "Likely human-written" 34–47% of the time. The quick check cannot say which part is which.
- It is English only. The model was trained on English. In our tests it was not reliable enough for German, French or Spanish, so the quick check refuses other languages.
- Short text is not checked. Below 150 words the scores in our tests were too unsteady to give a verdict.
- Limits of the test. It is our own test set and modest in size. New AI models and dedicated "humanizer" tools, which we did not test, may be harder to detect.
Why AI detector accuracy figures disagree
Different detectors publish very different accuracy numbers, and the same detector can look excellent in one test and weak in another. The figure depends on how it was measured as much as on the detector:
- What the test texts are. Wikipedia articles, student essays, emails and fiction behave differently. A detector tested on one kind of writing tells you little about another.
- Which AI models wrote the samples. A detector can be strong on one chatbot and weak on a newer one. A test that uses only a few older models will overstate accuracy on today's text.
- Whether the AI text was edited. Raw chatbot output is the easiest case. Paraphrased, translated or hand-edited text is harder, and many published figures leave it out.
- Where the threshold sits. Any detector can raise its catch rate by flagging more texts as AI, at the cost of more false alarms on human writing. A single "accuracy" number hides that trade-off, so look for the false positive rate on human writing as well.
- Who ran the test. Figures measured by the vendor on its own data are not the same as independent tests, and a test set the detector was trained on gives inflated results.
When you compare AI detector accuracy, ask for both rates: how much AI text was caught, and how much human text was wrongly flagged, on content like yours. That is why we publish our test set and both rates above, including the cases the quick check gets wrong.
What a false positive means
A false positive is human work that a detector calls AI-generated. Even a low rate matters at scale: at 3 in 1,000, a teacher who checks 1,000 genuine essays should still expect a few false alarms. The person affected usually has no way to prove a negative, so the cost of a false accusation falls on them.
A false positive does not mean the writer did anything unusual. Clear, formal or formulaic writing, writing in a second language and text polished with grammar tools can all resemble machine output. A false negative, AI text that passes as human, is also common, especially after rewriting. Neither kind of error can be seen from the verdict alone, which is why a result should start a conversation rather than end one.
Where text detection struggles
- Short text. Under about 300 words there is little signal, and scores move more between drafts.
- Edited or mixed text. A human-edited AI draft, or AI-polished human writing, can land in any zone. One verdict for a whole text cannot say which parts are which; sentence scores in a Deep check can help.
- Second-language writing. Plain, formulaic writing by people writing in a second language is more likely to be flagged as AI. Treat such results with extra care; see do AI detectors flag non-native English writers?
- Paraphrasing tools. Text run through "humanizer" or paraphrasing tools can avoid detection.
- Lists, code and quotations. These do not read like natural prose and give unreliable scores.
Where image detection struggles
- Screenshots, compression and resizing remove the detail detectors rely on.
- Edited photos. A real photo with an AI-filled area, or an AI image retouched by hand, can land anywhere on the scale.
- New generators can produce images the engine has not learned to recognise yet.
- Metadata. Missing camera data or credentials proves nothing, because apps strip them; present credentials say what a tool claims, not that the claim has been verified.
Using a result responsibly
- Treat the verdict as one piece of evidence.
- Look at the report's evidence and limits, not just the label.
- Follow the next steps in the report: ask for drafts or the original file, compare with earlier work, run a reverse image search.
- Never use a detector result alone to accuse, grade, reject or punish someone.
Reporting a problem
If you think a result is clearly wrong, email support@isthisai.pro with the report code. We review these cases when we evaluate engines.