UCSD Professor Drops AI Charge After Browser History and Drafts

A UC San Diego student had an AI cheating charge dropped after producing browser history and timestamped drafts showing how the paper was written. The case follows a pattern that has now surfaced across dozens of universities: a detector score triggers the accusation, and process evidence closes it. If you have been accused at UCSD, the sequence of what you preserve, and how quickly, is the difference between a hearing and a withdrawn charge.

What the case shows

The pattern is straightforward. An instructor runs a submission through an AI detector (typically Turnitin's AI writing indicator at UCSD), sees a score above whatever internal threshold they treat as significant, and refers the student to the Academic Integrity Office. The student responds by producing browser history covering the drafting window and Google Docs or Word version history showing the paper being built over time. The instructor, on review, withdraws the allegation before a formal hearing is scheduled.

This is not a legal victory or a policy change. It is what happens when the evidence a student can produce outweighs the evidence a detector score alone provides. The documented limitations of Turnitin's AI indicator mean that once a student produces credible process evidence, the instructor's original claim becomes harder to sustain.

Why browser history and drafts were decisive

A detector score is a probabilistic classification of the finished text. It says nothing about how that text came to exist. Browser history and version history do the opposite: they document the process, not the product. Together they create a timeline that is difficult to fabricate after the fact.

What each artifact proves:

• Browser history shows research activity: database searches, JSTOR access, Wikipedia lookups, source pages visited during the drafting window. It maps intellectual engagement with the topic before and during writing.
• Google Docs or Word version history shows the document being built incrementally: outline, rough paragraphs, revisions, cuts, restructuring. AI-generated text pasted whole into a document produces a fundamentally different version history than text written over hours or days.
• The absence of AI service visits in the browser record during the drafting window is itself evidence, particularly if the record is otherwise complete.

Why the detector flagged the paper in the first place

UCSD serves a large international and multilingual student population, and the research on detector bias here is directly relevant. The 2023 Stanford paper by Liang and colleagues, published in Patterns, found that GPT detectors misclassified non-native English writing at strikingly higher rates than writing by native speakers. Writing that avoids idiom, favors clearer syntax, and uses common vocabulary produces the low-perplexity signal detectors associate with AI output.

Formal academic writing produces similar signals. Careful topic sentences, hedged claims, and disciplined paragraph structure are precisely what instructors ask for and precisely what detectors flag. This is not a bug the vendors have solved. Turnitin itself has publicly acknowledged elevated false positive rates in documents where the AI indicator is below 20%.

What UCSD policy actually requires

UC San Diego's Academic Integrity Office handles AI-related allegations under the Policy on Integrity of Scholarship. The process typically begins when an instructor completes an Academic Integrity Review form and forwards it to AIO. The student is notified, offered a meeting, and given the chance to respond before any finding is entered.

UCSD policy requires a preponderance of the evidence standard: it must be more likely than not that a violation occurred. A raw detector score, without corroborating evidence, generally does not meet this bar on its own, particularly given the published research on detector reliability. The UC system's 2023 guidance to faculty cautioned against relying on AI detectors as sole evidence, and UCSD instructors operate within that broader UC context.

Questions to ask in writing

• Which specific detector produced the score, and what was the exact number?
• What threshold does the instructor or department treat as significant, and is that threshold documented anywhere in policy?
• Did any human review of the flagged passages occur before the referral to AIO?
• What corroborating evidence, apart from the detector score, supports the allegation?

If this is you at UC San Diego

The sequence matters. Preserve evidence before you respond, respond in writing before you meet, and keep the meeting focused on process and policy rather than emotion.

1. Export browser history covering the drafting window as HTML or CSV, from every device you used.
2. Open the flagged document in a browser, without editing, and screenshot the full Google Docs or Word version history with timestamps visible.
3. Save research artifacts: PDFs downloaded, citation manager entries, library database access records, notes.
4. Request in writing the detector name, the exact score, the threshold used, and whether any human review occurred before referral.
5. Review your syllabus and the Policy on Integrity of Scholarship to confirm what was actually prohibited in the course.
6. If suspension is on the table, or if you are an international student on an F-1 visa, consult an education law attorney before your AIO meeting.

If you are preparing a written response to an AIO referral, NotBot generates a personalized defense package that names your detector, cites the research on false positive rates, and organizes your process evidence into a form the reviewer can act on quickly. For post-finding cases, the appeal package covers the specific grounds that matter after a hearing has already produced a finding.