Maya Chen spent eight months in the archives. The doctoral candidate in postcolonial literature had traveled to three countries, photographed over two thousand pages of primary source documents, and conducted interviews with scholars on four continents. Her thesis chapter - a 12,000-word analysis of narrative resistance in mid-century Southeast Asian literature - represented the most rigorous scholarly work of her career. When she submitted it to her advisor in January 2026, the department's newly mandatory AI detection scan returned a result that would derail the next four months of her life: 67% AI-generated.
The Writer
Chen had been drawn to postcolonial literature as an undergraduate, when a professor assigned Chinua Achebe's essay "An Image of Africa" and something shifted in how she understood the relationship between language and power. By the time she entered her doctoral program, she had published two peer-reviewed articles and presented at six conferences. Her writing style - dense, precise, heavy with scholarly apparatus - was recognizable to anyone who had read her work.
It was also, she would later learn, precisely the kind of writing most likely to trigger a false positive. Academic prose, with its formal vocabulary, structured argumentation, and predictable paragraph organization, shares statistical characteristics with AI-generated text. For non-native English speakers writing in formal academic registers, the risk is compounded.
The Accusation
The detection report was granular. It highlighted specific passages - some from her close readings of primary texts, others from her theoretical framework - as having "high AI probability." The passages had nothing in common thematically. What they shared was a quality of careful, precise argumentation that the tool's algorithm associated with machine output.
Chen's advisor, to his credit, was skeptical of the result. But the department's policy required a formal review for any submission flagged above 50%. The review committee consisted of three faculty members, only one of whom specialized in Chen's field.
The Evidence
Chen assembled her defense with the meticulousness she brought to her scholarship. She presented: eighteen months of timestamped drafts in Google Docs, showing the chapter's evolution from outline to final text; photographs of her archival research, geotagged and dated; email correspondence with scholars she had consulted; her research notebooks, filled with handwritten notes in three languages; and a detailed process narrative explaining how each flagged passage had been developed.
She also retained an independent AI detection expert, who analyzed the flagged passages and identified the likely cause of the false positives: Chen's use of formal academic English as a non-native speaker, combined with the highly structured nature of literary analysis, produced text with low perplexity scores - the same statistical signature the tool associated with AI generation.
The Resolution
The review took four months. During that time, Chen could not defend her thesis, could not apply for postdoctoral positions, and was technically under investigation for academic misconduct. The committee ultimately found no evidence of AI use and cleared Chen completely. The department chair issued a written statement acknowledging the false positive and apologizing for the delay.
Chen has since resumed her doctoral work. She is expected to defend her thesis later this year. She now saves screen recordings of her writing sessions - hours of footage showing her typing, pausing, deleting, rewriting. "I shouldn't have to do this," she told us. "But I also can't afford not to."
The Lesson
Maya Chen's case illustrates several systemic failures: a detection tool applied without understanding its limitations; a policy that treats algorithmic output as presumptive evidence; an appeal process that took four months to resolve a clear false positive; and the disproportionate impact of detection technology on non-native English speakers and writers in formal academic disciplines.
The tools will improve. But the fundamental question Chen's case raises is not technical - it is ethical. When we give a statistical model the power to override eight months of documented human work, what are we saying about how much we value that work? And when the model is wrong, as it inevitably will be, who bears the cost?