What Turnitin's Own Margin of Error Means for the Palo Alto AI-Cheating Lawsuit

Quick answer

A sophomore in Palo Alto had an essay on "The Crucible" analyzed by Turnitin. The tool returned a 76% score suggesting the text may be AI-generated. The student's parents felt the score was treated as a verdict instead of a lead. They sued the school district for $150 million. The suit was settled in September 2026, with no ruling on the merits, so the authorship question remains open. Turnitin's own published documentation already answers what a 76% score actually promises. The company's math shows a score like that carries an uncertainty of about plus-or-minus 15 percentage points. Turnitin also notes a 50% score could contain up to 65% AI writing. The company says a single percentage was never meant to serve as standalone proof, and says so in writing.

What happened in Palo Alto

In October 2025, a Palo Alto High School sophomore turned in an essay on Arthur Miller's "The Crucible" to his teacher, Sarah Bartlett. She ran it through Turnitin. The tool returned a 76% likelihood of AI influence. The student had to rewrite the essay in class. The rewrite earned a D, and his semester grade slipped from a strong A or B to a C (SF Standard).

His father, Takashi Kato, who works in AI professionally, pushed back. The family submitted more than 1,000 pages of drafts, timestamps, and Google Docs revision history. By March 2026 they offered the district a trade: restore the grade to a B, and the dispute ends there. The district declined. In May 2026, Kato filed a federal civil rights lawsuit in the Northern District of California, seeking $150 million and alleging the school treated a Turnitin score as "dispositive without educator-driven evaluation or a meaningful opportunity for the student to respond before sanctions were imposed" (Palo Alto Online).

The complaint also alleged a pattern beyond this one essay: at least one other Asian student in the same class was flagged and forced into the same in-class rewrite, and male students in the class were four to five times more likely than female students to be subjected to a Turnitin check and a punitive grade replacement at all (SF Standard).

In September 2026, Kato dropped the suit. Both sides agreed to cover their own costs, and the case ended without a court ever ruling on whether the essay was AI-written or whether the district's process was fair (Palo Alto Online). That's an unsatisfying ending if you wanted a verdict. It's a useful one if you wanted to understand what a "76%" is actually capable of telling anyone in the first place, because that question has an answer that doesn't depend on how the lawsuit ended.

What Turnitin's own documentation says a score means

Turnitin publishes its own accuracy framework, and it is more specific than most people assume. According to the company's own FAQ for its AI writing detection capability, the model only flags text as AI-written when it reaches 98% confidence, specifically to minimize false accusations. The tradeoff for that caution is real: Turnitin's own words are that it will "likely miss up to 15% of text written by AI," with a false-positive rate the company puts under 1% in lab testing.

The company gives its own worked example: "if we identify that 50% of a document is written by AI, we are 98% sure that at least 50% is written by AI... but it could contain as much as 65% AI writing" (Turnitin FAQ, hosted by University of Kansas CTE). Read that again: the number on the report is a floor, not a point estimate, and the company's own stated range runs about 15 percentage points wide. A 76% reading, by that same logic, isn't "76% AI." It's "at least 76%, quite possibly closer to 91%" — or, under real-world conditions the company itself flags as noisier than its lab tests, possibly less reliable than either number suggests.

Turnitin's newer guidance for instructors adds another layer: scores under 20% no longer get a number at all. Since July 2024, anything below that threshold displays only an asterisk, because the company's own testing found a higher incidence of false positives in that range (Turnitin AI Writing Report guide). GPTZero, another detector, also reports its own confidence ranges for the same reason — every detector built this way carries some version of the same uncertainty. Read any score as a range, not a decimal.

Where a legitimate editing tool complicates the picture

The Kato complaint raised a detail that keeps surfacing in disputes like this one: the teacher said the student admitted to using Grammarly for "synonym assistance" (the family disputes that this admission happened). Whether or not that claim is accurate, the underlying mechanism is documented and real. The University of Lethbridge's 2025 study, nicknamed internally "Scarlet Cloak," tested a human-written academic abstract — composed before consumer AI tools existed — edited only with Grammarly's basic suggestion feature, then checked with Turnitin. The result: Turnitin scored the fully human-written, Grammarly-edited abstract at 16% likely AI-generated (International Journal of Educational Technology in Higher Education).

The researchers weren't out to embarrass Turnitin. They'd started the study because their own students, flagged for AI use, kept insisting they'd only used Grammarly — and the researchers didn't believe it was possible until they tested it themselves. Their conclusion: widely used, non-generative writing aids can measurably shift a detector's reading, especially for non-native English writers who lean on tools like Grammarly more heavily for comma placement and word choice. None of this proves the Palo Alto essay was human-written. It proves that "the student used an editing tool" and "the essay was AI-generated" are not the same claim, and a single percentage can't tell them apart on its own.

What this actually changes for a teacher or a student

None of this means detection scores are useless. It means a score is a reason to look closer, not a verdict to act on. Turnitin says this about its own tool in writing: "it should not be used as the sole basis for adverse actions against a student." That's not a caveat buried in fine print — it's the company's stated policy for how its own product should be used.

If you're reading a flagged score, treat the number as a range, not a decimal: ask what the detector's own documented margin of error is, not just what the headline percentage says. Ask what counted as "qualifying text" — short-answer, bulleted, or table content often doesn't get scored the same way prose does. And ask whether anything in the writing process (a grammar tool, a disability accommodation, a non-native-English background) is documented and known to shift detector output before treating a single run as settled. TheChecker.AI's free demo breaks a document down sentence by sentence instead of collapsing everything into one number, so you can see exactly which parts of a submission are driving a score up — the same kind of transparency Turnitin's own documentation says any single percentage needs before anyone acts on it. For what to do if you're the one holding a flagged essay, see our guide to appealing an AI-detector accusation.

FAQ

Does a 76% Turnitin score mean 76% of the essay was AI-written? Not precisely. Turnitin's own documentation describes its percentage as a confidence-weighted floor with roughly a 15-percentage-point range of uncertainty in its published examples, not an exact measurement of how much text is AI-generated. For more on how detection percentages work across different tools, see our breakdown of what a detection score actually means.

Can using Grammarly make an essay look more like AI to a detector? Yes, under documented conditions. A 2025 peer-reviewed study found a fully human-written abstract scored 16% likely AI-generated after only basic Grammarly suggestions were applied — not generative rewriting. See our full look at where Grammarly fits in AI-detection policy for how that distinction plays out in practice.

What happened in the Palo Alto lawsuit? The family sued the district for $150 million after a Turnitin score led to a lowered grade. The case was voluntarily dismissed in September 2026 before any court ruled on whether the essay was AI-written or whether the district's process was fair.