A Turnitin score, a rewrite, and a typed-up handwritten exam

The Palo Alto case is the latest, and one of the more revealing, examples of what happens when a school district treats an AI-detection score as proof rather than as a noisy signal.

In October 2025, a Palo Alto High School sophomore submitted an essay on The Crucible for an English class. Turnitin's AI-writing detector tagged the essay as 76% likely AI-generated. The teacher and an assistant principal interpreted the score as a strong indicator of cheating and required the student to rewrite the essay in class by hand. They also required the student to take the final exam by hand. According to the complaint, the assistant principal, Jerry Berkson, then had a school secretary type both the handwritten rewrite and the handwritten final exam into a digital document and ran those typed versions through Turnitin a second time, without notifying the family or getting consent.

The original 76% AI score knocked the student's grade in the class from a low A or high B down to a C. In a competitive Bay Area college-admissions environment, a one-letter drop in a humanities class is the kind of detail that follows a transcript into application reading rooms. The family put together roughly 1,200 pages of evidence to refute the AI accusation: process drafts, handwritten notes, and the document's revision history showing the essay being composed line by line.

None of that was enough at the school level to undo the grade. The family filed a federal civil rights complaint in the Northern District of California on May 5, 2026.

What the complaint alleges

Multiple outlets, including Palo Alto Online, the SF Standard, and Hoodline, have covered the complaint and its core claims. The thread that runs through all of them is that the district treated Turnitin's output as dispositive without a meaningful chance for the student to refute it.

The lawsuit alleges that the Turnitin tool's output was used as the basis for sanctions without educator-driven evaluation. It also alleges that at least one other Asian student in the same teacher's class was subjected to the same treatment, with a similar grade-replacement outcome. A second allegation is about gender disparity: the complaint claims that in the relevant teacher's class, male students were four to five times more likely than female students to be subjected to the Turnitin assessment and the subsequent punitive grade replacement.

Those are allegations at the complaint stage, not findings. The pattern they describe matters because it points to the larger structural issue with AI-detection tools in classrooms. The tools produce a single confidence number with no internal mechanism for evaluating fairness across student groups. Whether that number gets applied uniformly across a class, or selectively against students the instructor has independent reason to suspect, is entirely a function of how the human reviewer chooses to use it. The detector cannot tell whether it is being deployed evenly or as a confirmation engine for a prior suspicion.

Why the "typed-up handwritten exam" step is the load-bearing problem

Most coverage of the case focuses on the original 76% AI score and the rewrite consequence. The detail that should worry districts more is the second Turnitin run.

Once the student had rewritten the essay by hand in class and taken the final exam by hand, the school had paper artifacts that were necessarily written by the student. A handwritten essay produced in a supervised classroom is about as close to a verifiable human writing sample as a school can get. Running that sample through Turnitin's AI detector by typing it up and submitting the typed version is treating the AI detector as an oracle that can override the physical evidence in front of you. If the typed version came back as a high AI score, what would the district have done - reject the handwritten work the student produced in class because the tool said it looked AI-ish.

This is the core epistemological problem with AI-writing detectors. The tools do not measure whether the text was generated by AI. They measure whether the text resembles statistical patterns the tool's training data associates with AI-generated text. Those patterns include things like sentence length consistency, vocabulary distribution, and certain syntactic regularities, which can also describe a careful student writer trying to sound formal. The 76% score in the original essay was a probability estimate from a model that gets these wrong at meaningful rates. Re-running the supervised handwritten work through the same model and treating the second score as additional evidence is methodologically circular.

Turnitin itself has acknowledged uncertainty in its AI-detection product, and Rolling Stone's broader feature on AI-cheating false accusations documented previous students who were wrongly flagged and went through similar grade-replacement processes before clearing their names. The Palo Alto case is the version that ended in federal civil rights litigation.

The due-process gap

The mechanical question raised by the complaint is whether the district has a procedure for a student to challenge an AI-detection-driven sanction before it lands on the transcript. The implicit answer in the lawsuit is no. The student was given the rewrite as the sanction, not as an opportunity to clear the record. The family's 1,200 pages of evidence had to be marshaled retroactively, after the grade had already moved, and they were directed at a district administration rather than at a formal evidence-weighing process. The federal complaint frames that as a due-process failure with a civil rights overlay because of the alleged gender and racial disparity in who got the Turnitin treatment in the first place.

The Baltimore case on this site - a school district that called police on a student because an AI gun-detection system mistook a bag of chips for a firearm - is the same shape of incident in a different domain. In both cases, the school's AI-driven workflow had no built-in friction between the AI's confidence signal and a real human consequence. In Baltimore, the consequence was an armed police response. In Palo Alto, the consequence was a grade reduction with college-admissions implications. The connecting tissue is that the AI got the call wrong and the institution had no internal step that would have caught the error before the harm was inflicted.

What this means for the broader use of AI-writing detectors

Several districts have already pulled AI-detection tools out of student-discipline workflows, citing exactly the unreliability issues the Palo Alto complaint surfaces. Turnitin and other vendors have argued that the scores should not be used as standalone evidence in academic-integrity proceedings. That position is consistent with what the underlying research says: AI-writing detectors are noisy, have measurable false-positive rates that vary across demographic groups, and tend to produce higher false-positive rates for non-native English writers.

The Palo Alto Unified district's position in the suit is not yet on the public record beyond standard denials. Whatever happens to the federal complaint, the underlying pattern - school treats AI detector as proof, student is harmed, family sues - has now been documented enough times to be predictable. Districts that have not removed AI-detection scores from their discipline pipelines are running on borrowed time.

The graveyard lesson

Educators have a real problem. Some students do use AI tools to cheat. The temptation to reach for a tool that promises to identify them is understandable. The lesson the Palo Alto case underscores is that the current generation of AI-detection tools is not that tool. They produce confidence scores, not verdicts. The structural fix is to either stop using them in discipline workflows entirely, or to require that any detector-driven flag be paired with at least one independent piece of evidence - draft history, in-person interview, a comparison sample of supervised writing - before any sanction lands on the student.

Until that becomes standard practice, every district running Turnitin AI scores into transcript decisions is one motivated family away from the same federal complaint. The student's grade and college prospects are the visible damage. The deeper damage is to the trust between the school and the families who send their kids there expecting that a number from a model will not be the only thing standing between their child's essay and their academic record.