She writes like a researcher, gets flagged like a bot

Sumaiya, a technical program manager leading data and AI implementation at a public research university — 200-plus ADF pipelines, more than ten departments, 530-plus Power BI users — sat down one July evening and pasted a draft of her own into an AI detector. The result came back: entirely model-generated content. She closed the tab and never ran her own writing through a detector again. The gap between that instant reaction and the months of silence that followed is the real story.

Not just a statistical error

The detectors Substack adopted via Pangram and LinkedIn's "looks like AI slop" button don't make random mistakes. They target patterns that read as non-spontaneous: orderly structure, dated evidence, cross-references, phrasing that resembles a report built to withstand skeptical stakeholders. That is exactly how Sumaiya has written for years, long before ChatGPT launched. The professional habit of showing your work has become a tell that trips the alarm.

Second language, first suspicion

She arrived in the United States at 19 and learned to write professionally in a second language while her peers did it in their first. Throughout her career the accent drew a second look, the syntax earned corrections no one else received, the proof of competence required a longer runway. A detector that decides in half a second that her writing style "isn't hers" adds an automated layer to bias that already exists in meeting rooms.

Assistive tool turned offense

Long before large language models she used Grammarly and Google Translate to verify correctness before an email, a publication, or meeting prep. No one ever called it suspicious; it was simply what someone building a career in a non-native language does. The tools didn't change, the process didn't change; what changed is the public category that conflates "assisted" with "generated" and folds a daily habit into a panic that was never aimed at it.

Rigor looks like a pattern

Evidence-heavy, dated, cross-referenced writing built like a finding that must be defended — that isn't a red flag, that is what rigor looks like on the page. A statistical detector hunting for deviation from spontaneity cannot tell the difference between a hollow post tuned to sound confident and a researcher showing her work. Both produce the same symptom; only one is an actual problem.

The practical takeaway

The problem isn't "detectors are inaccurate" — that's an abstraction. The problem is that this specific inaccuracy lands on people who sound like her, on people who write the way rigor actually looks. Until the public conversation separates assistance from generation, and structure from fabrication, professionals whose only offense is competence will keep receiving failing grades from a machine that doesn't understand context.