What happened
According to the complaint and reporting by the San Francisco Standard and Palo Alto Online, a sophomore submitted an essay on Arthur Miller's The Crucible to his English teacher in October 2025. Turnitin's AI writing detector reported that 76% of the essay was likely AI-generated.
The family disputed the flag. Over the following months they assembled what the complaint describes as a 1,162-page evidentiary packet: drafts, notes, and the document's full revision history. The district did not reverse course. The student was made to rewrite the essay in person, received a D on the rewrite, and saw his class grade drop. In May 2026 his father, Takashi Kato, sued the Palo Alto Unified School District in federal court, alleging discrimination, retaliation, lack of due process, and improper grading.
Nothing here has been decided. These are allegations in an active lawsuit. But the shape of the dispute is worth every teacher's attention, because the same shape can form in any classroom that treats a detector score as proof.
The number at the center: 76%, plus or minus 15
The detector returned 76%. What often goes unsaid is that Turnitin itself acknowledges a variance of plus or minus 15 percentage points in its scores. A "76%" can mean the model's estimate ranges from 61% to 91%. That is a wide band to hang a disciplinary decision on.
Turnitin has also published its own false-positive rates. From a 2023 post by the company's Chief Product Officer:
Our document false positive rate — incorrectly identifying fully human-written text as AI-generated within a document — is less than 1% for documents with 20% or more AI writing. Our sentence-level false positive rate is around 4%. Annie Chechitelli, Chief Product Officer, Turnitin, June 2023
Read that carefully. The "less than 1%" applies only to documents that already scored 20% or higher, and it is a rate, not a guarantee. Across a single school's volume, even a sub-1% rate produces real students every term. And the 4% sentence-level rate means that of the sentences a teacher sees highlighted as AI, roughly one in twenty-five is human-written, with nothing on the screen to say which ones.
A detector score is a probability, not a witness. It can tell you something looks unusual. It cannot tell you who sat at the keyboard. When the score is the only evidence, the student is left trying to prove a negative.
Why 1,162 pages still wasn't enough
The family's packet was built mostly from Google Docs revision history: the trail of saved versions Docs keeps as a student types. It is the most common after-the-fact defense, and it has a real weakness. Docs revision history is coarse. It batches edits, collapses time, and shows you snapshots rather than the actual moment-to-moment act of writing. It can suggest that a document grew over time, but it cannot cleanly distinguish a student who wrote slowly from one who pasted text and then nudged it.
So the family was forced to compensate with volume. Eleven hundred pages is what it takes to argue from revision history when revision history was never designed to be evidence of authorship. And even then, it left enough ambiguity for the district to hold its position.
What real evidence of authorship looks like
The thing a detector score gestures at, and the thing 1,162 pages tried to reconstruct, is the writing process itself. Not the finished text, but how it came to exist: the rhythm of typing, the pauses to think, the sentences revised mid-draft, the moments of pasting and where the pasted text came from.
That record is unambiguous in a way a percentage never is. Human writing has a characteristic shape: uneven pacing, bursts and stalls, deletions and rewrites. Pasted or machine-generated text arrives differently, all at once or at an inhuman cadence. Captured at the keystroke level, the difference is visible at a glance, and it does not depend on a model guessing.
Build the evidence first, not after
The lesson of this case is not that the family did too little. They did an extraordinary amount. The lesson is that the evidence they needed did not exist yet when the accusation landed, so they spent months trying to manufacture it from a tool that was never meant to provide it.
Manupropria captures the writing process as the student writes, inside the assignment, by default. If a question ever comes up, the proof is already there: a replay of how the essay was actually written, not a probability and not a stack of revision-history printouts. The student never has to prove a negative, because the positive record exists from the first keystroke.