Are AI Detectors Accurate? How to Read an Accuracy Claim

Every AI detector on the market quotes an accuracy figure, and almost none of those figures answer the question a worried student or a careful assessor is actually asking. This piece is the arithmetic, worked out, so you can read any of those claims for yourself.

Why a single accuracy number is close to useless

Accuracy is the share of all judgements that were correct. Its weakness is easy to demonstrate. Feed a detector a hundred documents of which two are AI-written, and a tool that simply says “human” every single time scores 98% accurate while catching nothing at all.

Accuracy also depends entirely on the test set. A detector evaluated against the output of one model family, at one length, in one genre, will report a very different figure from the same detector evaluated against a newer model, on shorter texts, in a different discipline. Unless a vendor tells you the mix, the number is unfalsifiable.

The four numbers that do mean something

Recall and false positives trade against each other, always. Lower the threshold and you catch more cheating and accuse more innocent people. Raise it and you accuse fewer people and miss more. There is no setting that escapes the trade — there is only a decision about which error you would rather make, and that decision is a policy question, not a technical one.

The base rate, and why the same detector changes character

Here is the part that surprises people. Take a detector with genuinely good numbers: it catches 80% of AI text, and wrongly flags only 1% of human text.

Now run it over a cohort of 1,000 essays in a course where 5% of students used AI undisclosed:

That is a usable tool. Roughly four in five flags are real. Now run the identical detector over a cohort where only 1% used AI:

Nothing about the detector changed. Same model, same threshold, same published accuracy. But in the second cohort, more than half of everyone flagged did nothing wrong. Push the base rate down to a fraction of a percent — a well-behaved postgraduate seminar, say — and the majority of flags are false.

What this means for the numbers vendors publish

Turnitin, to its credit, publishes both a document-level false positive rate (under 1% for documents marked 20% or more AI) and a sentence-level rate (around 4%), and states that a score should not be the sole basis for action against a student. That is more disclosure than most.

Run the sentence-level figure through the same arithmetic and you can see why it matters: on an essay with forty qualifying sentences, a 4% sentence-level error rate means you should expect one or two highlighted sentences to be wrong even when the overall reading is correct. Arguing about those individual sentences is a losing strategy; arguing about the document-level conclusion is the real conversation.

Two things that break the arithmetic entirely

1. New models. A detector is trained on the output of the models that existed when it was built. Text from a newer model can score as thoroughly human, because it genuinely does not look like the training data. Recall against yesterday's models tells you nothing about recall against today's.
2. Short texts. Every one of these measures is a statistic over a sample. On 80 words there is barely a sample. Detectors are least reliable exactly where people most often paste a paragraph in to check it.

This is why we say on the detector's own page that no detector, ours included, is accurate enough to act on alone. It is not a disclaimer bolted on by a lawyer. It falls out of the arithmetic above.

Questions to ask before trusting any detector

• What is the false positive rate, and at what threshold was it measured?
• What text was it evaluated against — which models, which genres, which lengths?
• What is the minimum length below which you do not report a result?
• What does the tool do when it is unsure? Does it say so, or does it show a number anyway?
• Does the vendor state, in writing, that the score is not sufficient grounds for action?

A vendor that answers all five is being straight with you. A vendor that answers none and leads with a single headline percentage is selling confidence, which is the one thing this technology cannot honestly supply.

For the mechanism behind the numbers, see How AI detectors work. For why the errors land unevenly on particular writers, see why AI detectors flag human writing.

Are AI detectors accurate?

Accurate enough to be a signal, never accurate enough to be evidence. Their usefulness depends on the threshold, the length of the text, the model that produced it, and how common AI use actually is in the group being tested.

What does '99% accurate' mean for an AI detector?

On its own, very little. Accuracy is the share of all judgements that were correct, which can be made to look excellent simply by testing on a set that is mostly human text. Ask for the false positive rate and the threshold instead.

What is a false positive in AI detection?

Human-written text that the detector flags as AI-generated. It is the error that harms an individual, and the one that vendors report least prominently.

Why do detectors get less reliable on short text?

Because every signal they use is a statistic over a sample of words. A short passage does not contain enough sentences for the measurement to be stable.

Can a detector prove someone used AI?

No. It produces an estimate about how text reads. Proof, where it exists at all, comes from process evidence: drafts, version history, notes and a conversation with the writer.