Are AI Detectors Accurate? What the Research Shows and How to Read a Score
AI detectors are not accurate enough to prove that someone used AI. The best commercial tools now catch most text pasted straight from a chatbot and rarely flag longer human texts, but weaker tools flag a lot of human text, and AI text that has been rewritten or written to imitate a real author can slip past many tools. Treat any score as a hint that tells you where to look, never as evidence on its own.
In recent independent tests the best tools scored far better than OpenAI's own detector did in 2023, when OpenAI withdrew it for low accuracy, but not well enough to make a score proof.
How AI detectors decide
AI detectors judge the text itself, not where it came from. Simpler tools look at how predictable the word choices are and how even the sentences look. Commercial tools train classifiers on large sets of human and AI writing. Either way, the output is an estimate such as "82% AI".
Predictability is where many false alarms come from. James Zou, the senior author of a 2023 Stanford study on detectors, said they "typically score based on a metric known as 'perplexity'", which correlates with how sophisticated the writing is. Plain, careful writing can look machine-like to that kind of measure, even when a person wrote every word.
What the research found
Studies checked in October 2026. The first three are from 2023; the last three test newer tools and models.
| Source | What was tested | What it found |
|---|---|---|
| OpenAI's own AI classifier (2023) | OpenAI's tool on a challenge set of English texts | It caught 26% of AI-written texts and labeled human text as AI 9% of the time. OpenAI withdrew it on July 20, 2023 "due to its low rate of accuracy". |
| Stanford study, Liang et al. (2023) | 7 detectors on essays by non-native and native English writers | On average, the detectors labeled 61% of 91 TOEFL essays by non-native writers as AI-generated; 89 of the 91 were flagged by at least one. US eighth-graders' essays were judged almost perfectly. |
| Weber-Wulff et al. (2023) | 12 public tools plus Turnitin and PlagiarismCheck | The tools were "neither accurate nor reliable", leaned toward calling text human-written, and did significantly worse once the AI text was disguised, for example by paraphrasing. |
| Erol et al. (2025) | GPTZero, ZeroGPT and Corrector on the abstracts and introductions of 250 human-written neurosurgery papers and 750 ChatGPT-written versions | The tools separated unedited ChatGPT text from human text moderately to very well overall, but 30% of the human-written texts scored above 50% AI on Corrector and 16% on ZeroGPT. GPTZero put none above 50%. Edited AI text was not tested. |
| Jabarian and Imas, University of Chicago (2025) | Pangram, GPTZero, OriginalityAI and an open-source model on 1,992 human passages in six genres and matching AI passages from four models | Pangram had "essentially zero" false positives and false negatives on medium and long passages. The open-source model wrongly flagged roughly 30-78% of human text. After the StealthGPT "humanizer" rewrote the AI text, GPTZero missed around half of it or more in most genres, while Pangram still caught nearly all of it. |
| Epoch AI (July 2026) | Pangram, GPTZero and Originality.ai on 495 human passages of about 500 words and AI passages from three current models | Pangram and GPTZero flagged none of the human passages; Originality.ai flagged 3.8%. All three caught nearly all plainly prompted AI text, but missed 10-18% of AI text written to imitate a real author, and about a quarter of such imitations in scientific writing. |
Read together, the studies say two things: which tool you use matters enormously, and style imitation is where even the best tools slip, while a humanizer defeated some of them.
For the details, start with OpenAI's note on its classifier, Stanford HAI on the bias study, the University of Chicago research brief and Epoch AI's 2026 test. The Weber-Wulff paper is in the International Journal for Educational Integrity, and the 2025 neurosurgery study in Acta Neurochirurgica.
Worked example 1: the same honest class, six different tools
Take 1,000 essays written entirely by students, with no AI at all, and apply the false positive rates reported above. This is our arithmetic, not a new test, and the studies used different texts and settings, so it is not a ranking.
| Tool and study | Reported false positive rate | Honest essays wrongly flagged, out of 1,000 |
|---|---|---|
| Pangram or GPTZero, Epoch AI 2026 (about 500-word passages) | 0 of 495 | about 0 |
| Originality.ai, Epoch AI 2026 | 3.8% | about 38 |
| OpenAI classifier, 2023 | 9% | about 90 |
| ZeroGPT, neurosurgery papers 2025 (human texts scored above 50% AI) | 16% | about 160 |
| Corrector, neurosurgery papers 2025 (human texts scored above 50% AI) | 30% | about 300 |
| Open-source model, Chicago 2025 | roughly 30-78% | roughly 300-780 |
Depending on the tool, the same honest class gets anywhere from no false flags to several hundred. Zero errors in 495 tries is also not proof of zero errors on every kind of writing.
Worked example 2: why even a 1% rate matters
Vanderbilt University did a version of this math when it switched off Turnitin's AI detector in 2023: it had submitted about 75,000 papers in 2022, so a 1% false positive rate could have meant around 750 papers wrongly labeled.
There is a second effect. For a flagged student, the question is not the error rate but the chance that their flag is wrong, and that depends on how many people really used AI. Here we assume a detector that catches 90% of AI-written essays, and vary the rest. The numbers are illustrative, not measured.
| Scenario, 1,000 essays | Essays written with AI | AI essays flagged (90%) | Human essays wrongly flagged | Share of flags that are wrong |
|---|---|---|---|---|
| A: 10% used AI, 1% false positives | 100 | 90 | 9 (1% of 900) | about 9% (9 of 99) |
| B: 10% used AI, 5% false positives | 100 | 90 | 45 (5% of 900) | about 33% (45 of 135) |
| C: 2% used AI, 1% false positives | 20 | 18 | about 10 (1% of 980) | about 35% (10 of 28) |
Scenario C shows the catch: with the same "1%" detector, when few people used AI, roughly one flag in three lands on an honest writer. For the best tools on long text the rate may be closer to 0%, so this is the pessimistic case for them. For weaker tools, short texts and formal writing, rates of several percent or more are realistic, as the first example shows.
How to read a detector score
- Check what the percentage measures. Turnitin's number is the share of "qualifying text" it judges likely AI-generated, not a verdict on the whole paper. Read the tool's own explanation before acting on a number.
- Low scores are the least reliable. Turnitin itself shows only an asterisk instead of a number between 1% and 19%, because its testing found false positives more common in that range.
- Scores from different tools are not comparable. In the 2025 neurosurgery study, the same human-written texts averaged 37% AI on Corrector, 28% on ZeroGPT and 6% on GPTZero.
- Length and language matter. OpenAI said its classifier was very unreliable below 1,000 characters, a short paragraph or two, and much worse outside English.
- One run on one tool is one data point. A high score on a long text from a well-tested tool deserves a closer look. A high score on a short paragraph from a free tool deserves very little weight.
When a false positive is more likely
The risk of a wrong "AI" label goes up when:
- The writer is not a native English speaker. In 2023 the seven detectors in the Stanford study labeled 61% of TOEFL essays as AI on average, which the authors linked to simpler word choices that make text more predictable. Newer tools may do better, but treat such a flag with extra care.
- The text is short or not in English, as OpenAI warned about its own tool.
- Human and AI text are mixed. Turnitin says its sentence-level false positives are more common in mixed documents, especially where human and AI passages meet.
- The writing is formal and formulaic. Scientific papers follow strict conventions, one reason so many human-written neurosurgery texts scored high on Corrector and ZeroGPT in the 2025 study.
A wrong "human" label, a false negative, is more likely when AI text has been paraphrased, run through a humanizer, or written to imitate a real author, as the Weber-Wulff and Epoch AI tests found. The Chicago study found the same for GPTZero after a humanizer, though not for Pangram.
What to do if your writing is flagged as AI
If a teacher, editor or client says a detector flagged work you wrote yourself:
- Ask for details. Which tool was used, what score did it give, and which passages were highlighted?
- Show your process. Version history in Google Docs or Microsoft Word, outlines, notes, earlier drafts and the sources you read are much stronger evidence than any score.
- Offer to talk it through. Explaining your argument, your sources and your choices is something the author of a text can usually do easily.
- Point to the evidence. OpenAI withdrew its own detector, Turnitin advises using its score "to initiate a conversation, not to draw a conclusion", and Vanderbilt University turned off Turnitin's AI detector.
- Use the formal process. Read your school's or employer's AI and integrity policy and use its appeal route. If the stakes are high, ask a student adviser, your students' union or an ombudsperson for help.
To avoid a repeat, keep drafts and version history from the start, check the rules before using AI for anything, even grammar fixes, and disclose it where asked.
How teachers and editors should use a detector
- Use a score to start a conversation, not to reach a verdict. Ask the writer to explain what the piece is about and how they wrote it.
- Set clear rules up front. MIT Sloan's teaching technology team recommends clear AI policies, open discussion and asking students to describe how they used AI, rather than relying on detection.
Where a free AI checker fits
A detector is most useful on your own writing, as a quick check of whether a draft reads as generic before you submit or publish. If it does, rewrite in your own voice and add specific details only you know.
Clever AI has a free AI detector for that use. You paste 20 to 4,000 characters and get an AI score, a human score and a verdict of likely AI, mixed or likely human, with no account. Results on short passages are the least reliable, for the reasons above. It is a language-model estimate, not a scan of Turnitin, GPTZero or any other detector, and it was not one of the tools in the studies above. As the page itself says, use it to decide what to rewrite, not as proof about someone else's work.
If you use AI to brainstorm or draft, in Clever AI Hub or anywhere else, the safest habit is the same: keep your drafts, follow the rules you write under and say when AI helped.
The bottom line
In the newest tests the best detectors were rarely wrong about longer, unedited text, but none can prove anything about one piece of writing: tools differ enormously, rewritten or style-imitating AI text slips past some of them, and when real AI use is rare, many flags are wrong. Use a score as a reason to look closer, then decide on real evidence such as drafts, notes and a conversation.