What a detector is actually measuring
Broadly: how *predictable* the text is. AI-generated prose tends to be fluent, evenly-paced and statistically unsurprising. Detectors look for that signature.
The problem is obvious once stated: plenty of human writing is also fluent, even and unsurprising. Formal academic prose is *supposed* to be. A careful writer producing clean, conventional English looks, to a detector, exactly like a machine.
Who gets falsely accused
- Non-native English speakers, disproportionately and repeatedly. Writing that draws on a narrower, more conventional range of constructions scores as machine-like. This has been documented, and it is a serious equity problem.
- Anyone who writes in a plain, formal register — which is what academic writing training explicitly teaches you to do.
- Students who edit heavily, sanding away the idiosyncrasies that read as human.
- Neurodivergent students whose writing is unusually structured or consistent.
If you're accused and you're innocent
- Ask what the evidence is. A detector percentage is not evidence, it is a probability estimate from an unreliable instrument, and institutions are increasingly being told so.
- Produce your process. Drafts, version history, notes, search history, timestamps. This is the single most effective defence, and it only exists if you kept it.
- Offer to discuss the work. Someone who wrote the essay can talk about it. That conversation is more probative than any tool.
- Ask for the policy in writing and whether the detector's own vendor endorses its use as sole evidence. Several explicitly do not.
This is also, incidentally, why a study tool that refuses to ghostwrite is doing you a favour. Work you actually wrote has a trail behind it.