Are AI Detectors Accurate? What the Evidence Shows

AI detectors are not accurate enough to be treated as proof. In the most thorough independent test published so far, no tool scored above 80 percent accuracy. OpenAI withdrew its own detector for a low rate of accuracy. And the mistakes are not spread evenly: false positives land hardest on students writing English as a second language.

That is the short answer, and almost nobody writes it down in plain language. Search this question and you get journal articles, university library guides, and blog posts published by the companies selling the detectors. This guide is none of those. GeniusPal never writes prose you submit, so we have nothing to sell you about whether the thing that checks your prose works. Below is what the research actually measured, tool by tool and number by number, with the limitations of each study stated rather than buried.

How do AI detectors actually decide?

Not by comparing your essay against a database of chatbot answers. There is no such database, and there could not be one. A detector reads statistical properties of the text itself and estimates how predictable it looks.

The central measure is perplexity, which is roughly a score for how surprised a language model is by each next word. Text where every word is the obvious choice scores low perplexity, and detectors read low perplexity as machine-like, because a language model generating text is literally picking high-probability words. Most tools pair this with a measure of variation across sentences, so writing that keeps a steady rhythm and a steady sentence length looks more synthetic than writing that lurches between a six-word sentence and a forty-word one.

Sit with that for a second, because the whole false-positive problem falls out of it. Nothing in that method detects authorship. It detects predictability. Clear, orderly, plainly worded writing is exactly what a perplexity-based tool is built to distrust, and clear, orderly, plainly worded writing is also what most students are taught to produce. A well-organised five-paragraph essay with consistent sentence rhythm is a textbook low-perplexity document, and no human wrote it any less than a rambling one.

How accurate are AI detectors in independent testing?

The most comprehensive answer comes from Weber-Wulff and colleagues, whose 2023 study in the International Journal for Educational Integrity tested 12 publicly available tools plus two commercial systems, Turnitin and PlagiarismCheck, against a purpose-built document set. Their conclusion is stated without hedging: the available detection tools are neither accurate nor reliable, and they carry a main bias towards classifying output as human-written rather than detecting AI-generated text.

The numbers underneath that sentence are worth having. Even the highest accuracy values in the study land below 80 percent. Turnitin scored best across every method the researchers used, followed by Compilatio and the GPT-2 Output Detector, and only Crossplag and Go Winston also reached at least 70 percent. So of 14 tools tested, five cleared a bar of 70 percent and none cleared 80.

Then there is the detector that its own maker retired. OpenAI launched an AI Text Classifier in January 2023 and pulled it six months later, posting a note that as of July 20, 2023 the AI Classifier was no longer available due to its low rate of accuracy. On its own evaluation set the classifier correctly identified 26 percent of AI-written text as likely AI-written, while incorrectly flagging human-written text 9 percent of the time. The company with the deepest possible knowledge of how its own model writes could not build a reliable detector for it, and said so publicly.

Vendor numbers deserve the same scrutiny in the other direction. Turnitin publishes its own figures, and they are more careful than the way people quote them. Its chief product officer states that the document-level false positive rate is less than 1 percent for documents with 20 percent or more AI writing, while the sentence-level false positive rate is around 4 percent. Read the qualifier closely. The sub-1-percent claim is scoped to documents already estimated to be at least a fifth AI-written, which says nothing about a document sitting at 8 percent. For the vendor-by-vendor detail on that one product, we go deeper in our full breakdown of whether Turnitin can detect ChatGPT.

A 1 percent error rate also sounds small until it meets a real institution. Vanderbilt University did that arithmetic in public when it disabled the Turnitin AI detector for its instructors in August 2023, noting that it submitted 75,000 papers to Turnitin in 2022, so a 1 percent false positive rate would mean roughly 750 papers wrongly labelled as containing AI writing. That is the thing about rare errors at scale. They stop being rare.

Why do AI detectors flag human writing?

Because low perplexity is not a fingerprint of machines. It is a property of constrained language, and plenty of human beings write in constrained language for reasons that have nothing to do with a chatbot.

The clearest evidence here is a Stanford study by Weixin Liang, Mert Yuksekgonul, Yining Mao, Eric Wu and James Zou, published in Patterns and available as GPT detectors are biased against non-native English writers. The researchers ran seven widely used detectors over 91 TOEFL essays written by non-native English speakers and 88 essays written by US eighth graders. On the eighth-grade essays the detectors were close to perfect. On the TOEFL essays they misclassified more than half as AI-generated, an average false positive rate of 61.22 percent. All seven detectors unanimously flagged 18 of the 91 essays, and 89 of the 91 essays, or 97.80 percent, were flagged by at least one detector.

The per-tool spread in that paper is the part worth staring at. Originality.ai flagged 76 percent of the TOEFL essays and 1 percent of the eighth-grade essays. GPTZero flagged 52 percent against 0 percent. ZeroGPT flagged 48 percent against 0 percent. Same detectors, same settings, and a gap that has nothing to do with whether anyone cheated.

The researchers then proved the mechanism rather than assuming it. Asking a model to enrich the word choices in those same TOEFL essays, with no change to the ideas or the authorship, dropped the average false positive rate from 61.22 percent to 11.77 percent. Running the experiment backwards was just as revealing: simplifying the word choices in the American eighth-grade essays pushed misclassification up from an average of 5.19 percent to 56.65 percent. Vocabulary range, and nothing else, moved the verdict by fifty points. Their recommendation is blunt, and it is aimed at exactly the people using these tools on students: they strongly caution against the use of GPT detectors in evaluative or educational settings, particularly when assessing the work of non-native English speakers.

One honest caveat, since this post is about accuracy. That study is described by its own authors as a pilot with relatively small samples, and most of the detectors it assessed used GPT-2 as their underlying model. Detectors have been rebuilt since 2023. What has not changed is the design principle: the tools still key on statistical predictability, so the mechanism that produced the bias is still in the machine.

The same logic points at neurodivergent writers, and here the evidence base is genuinely thinner, so it is worth saying so plainly. Students who write in a highly regular structure, who lean on repeated phrasing, or who have been taught to lean hard on templates as a coping strategy produce low-perplexity text for reasons that have nothing to do with AI. Whitney Gegg-Harrison and Claire Quarterman set out the argument in a 2024 chapter, AI Detection High False Positive Rates and the Psychological and Material Impacts on Students, which presents a small corpus study finding much higher than advertised false positive rates and then, on the strength of that study plus news reports and first-person testimony, raises the possibility that neurodivergent writers along with second-language writers are more likely to be hit. Note the word possibility. The second-language finding has a controlled experiment with hard numbers behind it. The neurodivergent finding has a mechanism, case reports, and student testimony, and it does not yet have an equivalent measured false positive rate. Anyone quoting you a precise percentage for autistic or ADHD writers is quoting something that has not been measured.

Can you beat an AI detector by paraphrasing?

Yes, easily, and this is the finding that collapses the argument for treating detector output as an integrity control. The Weber-Wulff study measured it directly. AI-generated text that was afterwards run through a machine paraphraser was correctly identified only 26 percent of the time, meaning most of it sailed through as human writing. AI text that a person had lightly edited by hand scored 42 percent, against 74 percent for raw unmodified AI output.

The Stanford team hit the same wall from a different direction. A single follow-up prompt asking the model to elevate its own text with more literary language dropped detection of AI-written college admission essays from 100 percent to 13 percent. One prompt.

Now put the two halves of the evidence side by side, because together they say something neither says alone. The same Weber-Wulff study found 96 percent accuracy on straightforwardly human-written documents, and reports that accuracy dropping by 20 percent once human writing in another language had been machine-translated into English. So the tools are best at recognising unedited AI output and unedited native-speaker prose, and they degrade sharply at both of the edges where real students actually live.

What to do if you are wrongly flagged

The honest student is the one this whole topic actually endangers, so treat this as the practical core of the post. A flag is a claim someone has to support, and you answer it with a record rather than with indignation.

• Ask for the specifics before you respond. Which tool, what score, and did a person review the document. Most institutions require human review, and a percentage generated by software is not evidence of authorship on its own.
• Produce your version history. Google Docs and Microsoft Word both keep revision history automatically, and a document that grew over eleven days through forty revisions is very hard to argue with. This is the single strongest piece of evidence available to you.
• Bring the debris of real work. Outlines, dated drafts, handwritten notes, annotated readings, half-finished paragraphs you cut. Generated text has no history behind it, and a false accusation cannot manufacture one for you.
• Offer to talk through the argument. Ask to explain your reasoning, your sources, and why you structured it the way you did. Someone who wrote a thing can discuss it, and instructors know that.
• Cite the vendor against the vendor. Turnitin publishes a sentence-level false positive rate of around 4 percent, so some highlighted sentences are expected to be wrong. Vanderbilt turned the feature off entirely. Both facts are on the record.
• Ask about appeals in writing. If a case proceeds, get the process in writing and keep every message. Written records are what protect you if it escalates.

The prevention is simply to build that trail while you write, which costs nothing. Write in a tool that keeps history, save your outline, and never delete the messy drafts. It is also worth understanding that a detector score is rarely the whole story on the other side of the desk either, since instructors combine it with signals of their own, which we cover in how teachers know if you used ChatGPT.

What this means for how you use AI

The evidence points somewhere slightly awkward for everyone. Detectors are unreliable enough that no institution should treat a score as proof, and they are evadable enough that anyone determined to submit generated prose will get away with it. Both of those are true at once, and the student who loses under those conditions is the careful one with tidy handwriting in their prose.

Which makes the detection question the wrong one to organise your behaviour around. Handing in work you did not write is misconduct whether or not software notices, and the real cost is that you skip the thinking the assignment existed to make you do. The line between using AI to learn and using it to cheat is clearer than the discourse suggests, and we map it properly in our guide to whether using AI to study counts as cheating. The reframe that dissolves the anxiety is to keep AI on the studying side of the work and off the submitting side, which is what the prompts in our guide to using ChatGPT to study without cheating are built around.

This is the reason GeniusPal works the way it does. You upload your own notes, a lecture PDF, or a chapter, and it returns a set of questions drawn from that file, which you can work through as a quiz, as flashcards, or in active recall where you write the answer from memory and then check it against the source. It never drafts an essay, a paragraph, or a sentence you would hand in, so nothing it produces ever reaches a detector. The output is questions aimed at you, and the writing stays yours.

Worth stating the limits plainly too. GeniusPal reads text-based files you upload and works in a single pass, with no chat and no follow-up questions, and a new account includes two study-set generations for the lifetime of that account. What it gives you is the part detectors can never assess: whether you actually know the material well enough to write about it in your own words, at speed, under pressure, with nobody to check.