Why AI detectors flag human writing, and what changed in 2026

An AI detector flags human writing because it never sees an author. It sees a finished text and estimates how closely the style matches what it learned to call machine-made. Plain, even prose, the kind taught in composition classes and the kind a second-language writer produces with care, can sit close to that pattern. Our article “What an AI detector score actually measures” covers the statistics. This one covers what happened in 2026 when institutions treated the score as proof.

How often does a detector flag the wrong person?

The rate for a given classroom is not published, and the memo from Washington State University shows why the number matters even when it is small. The memo says Turnitin claims a false positive rate of 1 to 2 per cent, calls that figure disputed, and notes that other studies placed it higher, especially for neurodivergent students and students who write in English as a second language.

The memo then does the multiplication. Turnitin analysed 148,547 assessments at the university in fall 2024. Even at a rate of 1 per cent, that is almost 1,485 assessments flagged as AI-generated when a person wrote them. Turnitin’s own chief product officer wrote in 2023 that documents with less than 20 per cent detected as AI writing carry a higher incidence of false positives, and that the score is less reliable there.

What happened when a court looked at one flag?

Orion Newby, a student at Adelphi University, submitted a paper in a fall 2024 history course. His professor ran it through Turnitin, which labelled it wholly AI-generated. According to Plagiarism Today’s report of February 2026, two other AI checks called the same paper human-written. The judge’s finding, as reported by Newsday and relayed by Plagiarism Today, was that the accusation was “completely false”. The judge ordered the university to reverse the discipline and clear his record.

The order concerned the sanction, not the technology. The report says the judge found the decision to sanction him had no valid basis and no reason behind it.

Which institutions stopped relying on the score, and which have been taken to court over it?

Curtin framed its decision as being about trust and clarity in academic culture. Washington State gave three reasons: the false positive arithmetic, the hearing board outcomes, and the fact that the report goes to the instructor and not the student, so a student cannot raise a problem before an accusation. The memo says 33 per cent of hearing board cases on alleged AI use between 2023 and 2025 ended in a finding of not responsible, because the detector result had no other evidence beside it.

The Palo Alto case is a complaint, not a ruling. The San Francisco Standard reports that Turnitin gave an essay a 76 per cent likelihood of AI authorship and that the company says its scores can vary by plus or minus 15 percentage points. What the family wants is an order stopping the district from treating a score as decisive without a review by an educator and a chance for the student to respond.

Does the same text keep the same score?

A score belongs to the model and the report that produced it, and Turnitin changes both. Its release notes record that on 4 August 2026 the AI Writing Report merged its AI-paraphrased category into a single AI-generated category, and that the update applies to newly submitted files. On 6 October 2026 Turnitin added AI writing detection for Turkish, German and Dutch, built on a separate model from its English one.

Neither entry says old results were wrong. They show that a percentage is a reading taken at one moment, by one version of a tool, and that the reading is not stored in the text. A student accused months later has no way to rerun the identical measurement, which is one more reason a score cannot carry a case on its own.

What can a flagged author do?

• Ask what evidence exists beyond the score, and ask for the report itself. Washington State’s policy is that no detector is the sole support for a misconduct case.
• Open the version history of the file. Word keeps one for files stored in OneDrive or SharePoint, shows what changed between saved versions and lets you compare them.
• Gather the material around the text: notes, outlines, earlier drafts, the sources you read. A draft that grew over days is hard evidence of a process, and a detector cannot see it.
• Ask for a review by a person who will read those drafts. The Palo Alto family’s request is the same one, written as a court order.
• Do not rewrite a text to push a number down. A new version leaves a new history, and the question is who wrote the first one.

GenPolish’s detector reads register and lists the phrases behind its reading, and its page says a verdict is never proof of who wrote something. The text diff checker helps with the drafts: it shows what changed between two of them, which is the record a reviewer wants to see.

What should a teacher do with a score?

Turnitin’s chief product officer advised educators to “use the AI score to start a meaningful and impactful dialogue with their students”. That is the use the vendor describes, and it is compatible with Washington State’s policy, which does not allow a detector as the sole support for a misconduct case: a score can open a conversation about the draft, and the draft settles it. Our guide to what teachers can and cannot check explains what a missing watermark means for the same decision.

Editing a text is a separate matter from contesting a flag: the version history above is the record to bring to a review. If you also want to rework a draft that came from a model, the rewriter below is there for that, not to push a score down, and the text diff checker can compare the result with your original.