AI Detection Flags Code-Switching in Multilingual Graduate Theses

International graduate students writing theses in English face a specific risk from AI detectors: the stylistic patterns of multilingual writing, including code-switching between discipline-specific terminology and simplified connective prose, overlap heavily with the signals detectors treat as evidence of AI generation. The research is clear that this overlap produces disproportionate false positives on non-native English writers, and graduate theses sit at the intersection of several of the risk factors involved.

What code-switching looks like in a graduate thesis

In linguistics, code-switching refers to alternating between languages, registers, or styles within a single text or conversation. In a graduate thesis written by a multilingual author, this often shows up as alternation between dense technical terminology (where the vocabulary is fixed by the discipline) and simpler connective prose used to explain, link, and transition between those terms.

A chapter on neural network architectures might contain tightly packed noun phrases drawn from the field, interleaved with short, high-frequency sentences used to summarise or paraphrase. That contrast is a product of how multilingual writers manage cognitive load in a second language: offload complexity to the fixed vocabulary of the discipline, and keep the surrounding prose structurally simple. The research on non-native graduate writing and detector bias documents this pattern across disciplines.

Why GPTZero and Turnitin flag these patterns

Both GPTZero and Turnitin rely heavily on perplexity (how predictable the next word is) and burstiness (how much sentence length and complexity varies). Multilingual thesis writing produces statistical signatures that resemble AI output on both axes.

• Simplified connective prose produces low perplexity because the word choices are high-frequency and predictable.
• Discipline-specific noun phrases are also low-perplexity within the field, because they are formulaic by convention.
• When a writer alternates between short explanatory sentences and longer technical ones, burstiness can look either artificially low (if the simplified passages dominate) or artificially patterned (if the alternation is regular).

None of these features indicate AI authorship. They indicate a multilingual writer managing a second language in a technical register. But detectors cannot distinguish between the two.

What the research actually shows

The foundational study is Liang et al. (2023), published in Patterns. The authors tested seven GPT detectors against essays written by native English-speaking US eighth-graders and essays written by non-native English speakers from a TOEFL dataset. The native-speaker essays were misclassified as AI-generated at low rates. The TOEFL essays were misclassified at rates above 60 percent on average, with one detector flagging nearly 98 percent of non-native essays as AI-generated.

The authors explicitly attributed the gap to linguistic features that overlap between non-native English writing and AI output: lower lexical diversity, more formulaic phrasing, and reduced syntactic variation. The Weber-Wulff et al. (2023) evaluation in the International Journal of Educational Integrity reached a converging conclusion when testing fourteen detectors, finding that none met a reliability threshold suitable for institutional decision-making.

Why graduate theses are especially exposed

Several features of thesis writing compound the detector problem for international graduate students:

• Length and consistency. A thesis chapter is long enough that statistical patterns stabilise. Short false-positive bursts in an undergraduate essay may average out; in a chapter, they can dominate the detector's output.
• Literature review conventions. Paraphrased summaries of prior work use predictable discipline-standard phrasing ("Smith et al. argue that," "building on this framework"), which is exactly what low-perplexity text looks like.
• Editing passes. Supervisors routinely ask graduate students to simplify sentences for clarity. Clearer sentences are more predictable sentences, which drives perplexity further down.
• Translation drafting. Writers who draft in a first language and translate, or who think bilingually while composing, often produce output with reduced idiomatic variation.

What multilingual graduate students should document

If a thesis chapter is flagged, the evidence that tends to resolve the case is process evidence that no AI model can plausibly fabricate after the fact. The evidence gathering guide covers the general procedure; for graduate thesis cases, prioritise:

• Version history from Google Docs, Overleaf, or Word showing incremental drafting over weeks or months.
• Notes and outlines in your first language, if you drafted bilingually.
• Supervisor email threads with comments on earlier drafts, showing the evolution of the argument.
• Reading notes, annotated PDFs, and citation manager entries corresponding to the sources cited in the flagged chapter.
• A written statement describing your language background, drafting process, and any translation or editing steps.

Procedural points to raise

The procedural questions worth asking in writing, before any hearing, are consistent across institutions. Our procedural rights FAQ covers the baseline; for a multilingual thesis case, add:

• Which detector was used, what score was reported, and what threshold the institution treats as significant.
• Whether the detector's documented limitations on non-native English writing were considered before the allegation was filed.
• Whether a human reviewer evaluated the flagged passages against the student's prior written work in the program.
• Whether the institution's policy requires evidence beyond a detector score for a finding.

For international students, the stakes extend beyond the academic record. A finding of misconduct can affect visa status, program funding, and future admissions. If the proposed sanction is dismissal from the program or if visa status is implicated, consulting an education law attorney before the hearing is reasonable. If you are preparing a written response, NotBot generates a personalized defense package that cites the research on non-native English detector bias, documents your writing process, and addresses the specific procedural requirements of your institution.