AI Detection False Positives on Non-Native Graduate Students

Graduate students writing in a second language sit at the intersection of two well-documented AI detector failure modes: technical, formal prose and non-native English phrasing. A UC Santa Barbara case, in which a professor publicly withdrew an AI cheating accusation after the student produced browser history and draft evidence, illustrates what happens when process evidence meets a detector score in the record.

What happened at UC Santa Barbara

The pattern reported in the UCSB case is now familiar: a written assignment was flagged by an AI detector, an accusation was issued, and the student responded with two categories of evidence. The first was a full browser history covering the drafting window, showing library database access, research reads, and the times spent on source pages. The second was a document version history and earlier draft files, showing the paper being written incrementally rather than pasted in.

After reviewing the material, the instructor retracted the allegation. For a broader look at how the same category of evidence has closed similar cases, see our writeup of the UCSB browser history retraction.

Why graduate non-native English writing triggers detectors

Graduate-level writing tends to be structured, careful, and vocabulary-controlled. Non-native English writing tends to favor high-frequency vocabulary and regular syntax. Both properties reduce perplexity, the statistical signal AI detectors rely on to distinguish human from generated text. Low perplexity is one of the main things that trips an AI flag.

The most-cited piece of research on this problem is Liang et al. (2023), published in Patterns (Cell Press). The Stanford team tested detectors against essays written by non-native English writers and found sharply higher false positive rates on that population than on native English writing.

Weber-Wulff et al. (2023), in the International Journal of Educational Integrity, tested fourteen AI detection tools across a range of texts and concluded that none met a threshold of reliability appropriate for institutional decision-making. Vocabulary control, edited or paraphrased text, and formal register all degraded accuracy.

What the UCSB process evidence actually proved

Browser history and draft versions do not directly prove that a document is not AI-generated. What they prove is the presence of a normal human writing process during the window the paper was written: source pages opened and read for realistic durations, incremental edits, revised sentences, and cited works actually accessed.

In practice, this shifts the question the instructor has to answer. A detector score alone asks the student to disprove a probabilistic flag. Process evidence forces the accuser to explain how an AI-generated paper could have produced hours of matching library database access, revisions, and reading history. That explanation is difficult to construct, which is why retractions in these cases tend to follow.

What graduate students should preserve

• Full browser history covering the drafting window, exported before the retention window closes
• Document version history (Google Docs, Word AutoSave, Overleaf history for LaTeX)
• Reference manager records (Zotero, Mendeley, EndNote) with import timestamps
• Library database access logs where available through your institution
• Advisor emails, meeting notes, and any feedback exchanges tied to the draft
• For coded or quantitative work: Git history, notebook checkpoints, dataset access logs

Using the research in a graduate-level response

A graduate response should name both problems: the general accuracy limits of detectors and the specific bias against non-native English writers. Both are peer-reviewed and citable. The Liang paper is directly on point for anyone whose first language is not English. The Weber-Wulff paper is the standard citation for arguing that no detector meets the reliability threshold expected of institutional evidence.

Pair the research with process evidence and a specific request: ask, in writing, which detector produced the flag, what the score was, what the institution's stated threshold is, and whether a human review preceded the accusation. Our procedural rights FAQ covers what you are entitled to request before a hearing.

If you are a graduate student under investigation

The steps that closed the UCSB case are the same steps that close most detector-only cases. Stop editing the flagged document. Export browser history, document version history, and reference manager data immediately. Write a clear timeline of how the paper was researched and drafted, mapping specific sources to specific sections. Cite Liang et al. (2023) and Weber-Wulff et al. (2023) by name when addressing the detector score. If the finding has already been issued, the grounds and format for challenging it are covered in our guide to the appeal package.

If you are preparing a written response, NotBot generates a personalized defense package that names the detector that flagged you, cites the peer-reviewed research on non-native English bias, and organizes your draft, browser history, and reference evidence into a document the reviewer can actually read.