How Many Honest Essays Get Flagged? A Base-Rate Check for AI Detectors (2026)

Verdict: A flag from an AI detector is a probability, and the probability depends on how many AI-written documents you expect to see. When AI text is rare, even a good detector can flag more honest writers than AI writers. Read every score with the base rate in mind.

The worked example

FavTutor’s 2026 field review uses a simple case to explain false positives. Suppose you check 1,000 documents, and 10% of them are AI-written. Suppose the detector catches 90% of the AI documents, which is its sensitivity, and clears 95% of the human documents, which is its specificity. FavTutor’s pricing notes uses these round numbers.

Then 90 of the 100 AI documents are flagged, which is the detector working as intended. But 5% of the 900 human documents are also flagged, which is about 45 honest writers. In total, 135 documents are flagged, and only 90 of them are AI. The other 45 are people who wrote their own work. Source: eesel AI’s feature table.

Source: the FavTutor’s price guide cited above presents this example.

What the example teaches

The useful number is not the accuracy figure on the vendor’s page. It is the share of flags that are correct, and that depends on how common AI writing is in the pool you are checking. Change the share of AI documents and the same detector produces a different share of correct flags.

A school, publisher or hiring team should estimate its own base rate before it reads a single score. If you expect very little AI writing, a flag should carry less weight than the headline accuracy suggests. If you expect a lot, the same flag carries more.

Why non-native writers need extra care

FavTutor reports that historical research found disproportionate false-positive risks for some non-native English writers. It cites a 2023 Patterns study that found 97% of 91 TOEFL essays were flagged by at least one of seven detectors tested, per the FavTutor’s plan breakdown. FavTutor says that figure should not be read as the current error rate of every 2026 commercial detector. It still counts as a reason to require human review and to give the writer a chance to respond.

eesel AI’s comparison page reports a similar concern for Turnitin, a Stanford-linked study that flagged 61.3% of non-native English writers’ essays as AI-generated. Those are two different studies from different years and tools, so the figures should not be combined.

What vendors say about their own false-positive rates

GPTZero’s pricing notes says it trains its model to reduce the false-positive rate for ESL writing to 1%. That is a vendor claim, and this article has not tested it. The claim is worth asking about: which test set, which writers, and which length of text.

A reviewer’s checklist

FavTutor’s workflow for consequential decisions is a good starting point:

• Preserve the exact text, the detector version, the result and the date.
• Review the highlighted passages in context.
• Check drafts, citations, source notes and revision history.
• Consider length, language, genre and any editing assistance.
• Ask the writer to explain their process.
• Treat the detector as one signal, never the sole verdict.

What this article does not settle

The example uses FavTutor’s round numbers, and it is an illustration, not a measurement of any product. This article did not measure any detector’s sensitivity or specificity, and it does not give a base rate for any school or publisher. The 97% and 61.3% figures are as the cited reviews report them. Before you rely on a detector, estimate your own base rate and test the tool on samples from your own writers. The site’s method is on the method page. Source: FavTutor’s tier list.