A student accused of AI use on a written assignment produced browser history and draft timestamps as evidence. The professor withdrew the accusation. The pattern is now common enough that it has a shape: detector flag, process evidence, retraction. Understanding why it works, and what the evidence actually proves, matters more than the name of the institution involved.
The pattern: detector flag, process evidence, retraction
Publicly reported cases at institutions including UC Santa Barbara, UC Davis, Vanderbilt, and others follow a recognizable sequence. A written assignment is run through an AI detector. The detector returns a high AI-likelihood score. The instructor issues an accusation based primarily on that score. The student then produces evidence of the writing process: browser history, saved drafts with timestamps, research notes, or version history. The instructor withdraws the accusation, often within days.
The retractions are not concessions that the detector was wrong in principle. They are acknowledgments that the score, on its own, cannot establish authorship when the student produces contemporaneous evidence of writing. For context on how these dismissals happen at UC campuses, see our coverage of the UCSB retraction case.
Why browser history and draft timestamps work
A detector score is a probability estimate based on statistical features of the finished text. It says nothing about how the text was produced. Browser history and draft timestamps do the opposite: they show the process, and they are hard to fabricate after the fact.
Specifically, this kind of evidence tends to show:
- Time spent on library databases, JSTOR, Google Scholar, or specific source pages relevant to the assignment
- Timestamps on saved drafts that align with the research activity
- Incremental revision patterns consistent with human writing: rewritten sentences, moved paragraphs, deleted false starts
- The absence of pasted blocks of text appearing suddenly and without prior drafting
When a student can show, minute by minute, that they were reading source material and drafting the paper across multiple sessions, the detector score becomes an unexplained anomaly rather than proof of misconduct.
What the research says about detector accuracy
The reason retractions like this keep happening is that the underlying tools have documented reliability problems. In a 2023 study published in the International Journal of Educational Integrity, Weber-Wulff and colleagues tested fourteen AI detection tools and concluded that none performed consistently well enough for institutional decision-making. A separate 2023 Stanford paper (Liang et al.), published in Patterns, found that detectors flagged writing by non-native English speakers as AI-generated at strikingly high rates even when the text was fully human-written.
Neither study says detectors never work. Both say the outputs are not the kind of evidence that alone should support a finding of misconduct. That is the same conclusion instructors have reached, case by case, when confronted with process evidence.
What most academic integrity policies actually require
Most institutional policies require a specific standard of proof, commonly a preponderance of the evidence or, at some institutions, clear and convincing evidence. In either case, the question is whether the totality of the evidence supports the allegation, not whether a single automated tool produced a score above a threshold.
When a student submits contemporaneous process evidence, the instructor or panel has to weigh that evidence against the detector output. A number produced by a tool with known false positive rates rarely outweighs timestamped drafts and browser history showing hours of research and revision. Our procedural rights FAQ covers what information you can request before a hearing.
If you have been accused
The first hours matter. Retention windows on browser history and cloud document version history are limited, and once they expire, the evidence is gone.
- Stop editing the flagged document. Every save can overwrite the version history that supports you.
- Export browser history for the drafting window as HTML or CSV before the retention period ends.
- Screenshot and download the full Google Docs or Word version history, including timestamps.
- Preserve secondary evidence: library database access logs, citation manager entries, messages with classmates, and any handwritten notes.
- Request in writing the specific detector used, the score, and any human review notes that led to the accusation.
A response letter that pairs process evidence with the peer-reviewed research on detector reliability is the format that has moved these cases at multiple institutions. If you are preparing a written response, NotBot generates a personalized defense package that addresses the specific detector, cites the relevant research, and structures your process evidence for the reader. If you are past the initial finding and need to file an appeal, the appeal package covers the procedural grounds that matter at that stage.
If the proposed sanction is suspension, expulsion, or has visa consequences, consult an education law attorney before your hearing. The research and the process evidence support your argument, but severe sanctions warrant qualified counsel.
Build your defense package
A personalized response that documents your writing process and addresses the detector that flagged your work.
Get your defense package$49 one-time · Generated in 60 seconds