How AI Detectors Actually Work — and Why Careful Writing Gets Flagged

Large language models write by choosing the most likely next word. As a result, AI-generated text tends to be statistically predictable — and that is exactly what perplexity measures. The lower a text’s perplexity, the more machine-like it scores.

Human writing is messier: unexpected word choices, personal habits, occasional awkwardness. That irregularity is the statistical fingerprint of a person.

The second cue is burstiness — variation in sentence length and structure. Human drafts swing: a short sentence, then a long one, a sudden aside. AI output tends to be evenly paced and uniformly structured.

This creates an unfair edge: the more carefully you edit, the smoother your burstiness becomes — and the more machine-like your writing looks to this family of detectors.

GLTR by Gehrmann, Strobelt & Rush (ACL 2019 Demo, arXiv 1906.04043) colors every word by how predictable it was in context — the clearest public demonstration of what statistical detectors see.

In that study, showing people a GLTR-style view improved their own detection of fake text from 54% to 72%. Teaching people how detectors see works — which is part of why this page exists.

An inference from the mechanism: any writing style that trains predictability — formal templates, fixed transitions, perfectly parallel structure — sits closer to what these detectors flag. Non-native academic English is hit especially hard: a 2023 study (Liang et al., Patterns) found commercial detectors misclassified about 61% of non-native English essays, versus near zero for native ones. That figure is specific to non-native English writing, not a general false-positive rate.

Being flagged is not evidence about who you are. It is evidence about how your writing scores on two statistical habits.