Why AI detectors flag human writing

The two measurements underneath almost every detector

Perplexity asks: given the words so far, how surprising is the next one? A language model is trained to pick likely next words, so its output is unsurprising by construction — low perplexity.

Burstiness asks: how much does sentence length and complexity vary within a passage? Human writing tends to lurch — a long winding sentence, then a short one. Model output is comparatively even.

A detector flags text that is simultaneously unsurprising and even. Nothing about that description is exclusive to machines.

Who writes that way without any machine involved

• Second-language writers, who choose common, safe vocabulary and regular structures — the best-documented false-positive population
• Students taught rigid essay templates, which produce even paragraph rhythm by design
• Technical, medical and legal writers, where conventional phrasing is required for precision
• Anyone who edits heavily, since revision smooths out exactly the irregularities that read as human
• Neurodivergent writers who favour consistent, structured sentence patterns

Why a low advertised error rate still means many flagged students

Suppose a detector has a genuine 1% false-positive rate at the document level. Across a course of 500 submissions, that is five students accused wrongly. Across a university term, it is hundreds.

The rate sounds reassuring on a marketing page and stops sounding reassuring the moment you scale it to the number of documents actually being scanned. This is why the false-positive rate, not the accuracy figure, is the number that matters — and why so few vendors lead with it.

Thresholds are a policy choice, not a fact

Every detector picks a cutoff above which text is called AI. Move it down and you catch more machine text and more innocent students. Move it up and you miss more of both.

Vendors selling to publishers tune aggressively — a false accusation costs them little. Vendors selling to schools have more reason to be cautious. The same underlying model can produce very different verdicts depending on where the line was drawn, and that line is a commercial decision.

What would actually be evidence

Statistical resemblance is not provenance. Provenance signals are, and they exist — they are just mostly limited to images so far.

C2PA Content Credentials attach a cryptographically signed record of how a file was produced. Google's SynthID embeds a watermark in output from Google models. Both give you a fact rather than an estimate. Neither is available for arbitrary text pasted from a chat window, which is why text detection remains probabilistic and will stay that way.