The foundational metric: perplexity
A language model assigns a probability to every sequence of tokens. Perplexity is the inverse of those probabilities: a text that the model finds very predictable — where most words are exactly what the model would have chosen — has low perplexity. AI-generated text typically has very low perplexity against the model that generated it, because the model always picks high-probability continuations. Human writing has higher, more variable perplexity because humans make unexpected but coherent word choices.
Burstiness: the human variance signal
Human writers do not maintain a constant complexity level. They alternate between simple, direct sentences and complex, embedded clauses — what researchers call 'burstiness'. AI text tends to maintain a smooth, consistent complexity level across paragraphs, which is statistically unusual for human writing. Detectors measure the variance in sentence complexity to produce a burstiness score: low variance suggests AI generation.
N-gram frequency analysis
Certain phrase patterns are so strongly associated with AI-generation that their presence alone predicts the source with reasonable accuracy. Common examples include 'It is worth noting that', 'In conclusion', and 'It is important to'. N-gram detectors maintain large lookup tables of over-represented phrases and score text based on their density. These detectors are fast but brittle — they fail against models trained after the lookup table was built.
Token probability watermark detection
When a model uses a deterministic watermarking scheme like SynthID, the detector has access to the same secret key used during generation. It re-scores every token against the green list and sums the number of green-list tokens. A score significantly above the expected chance level constitutes a positive detection. This is much more reliable than perplexity scoring but only works for watermarked models whose key the detector holds.
Stylometric feature models
Deep-learning detectors extract dozens of stylometric features — average sentence length, punctuation density, vocabulary richness, passive-voice ratio — and pass them through a classifier trained on large corpora of human and AI text. These models generalise better than n-gram lists but suffer from distributional shift: they misclassify text from new AI models released after training.
Why cleaning character-level artefacts improves detector scores
Several AI detectors use the presence of invisible Unicode characters as a strong positive signal for AI-generated text, because legitimate human text almost never contains zero-width spaces or byte-order marks mid-sentence. Removing these characters eliminates one of the most reliable signals. Separately, normalising non-breaking spaces and smart punctuation reduces the n-gram overlap with AI-training corpora, further lowering the detector confidence score.