Perplexity and burstiness, explained simply

Perplexity: how surprising is your next word?

Language models predict next words. Perplexity measures how surprised a model is by the word that actually comes next. "The cat sat on the ___" → "mat" barely surprises it; "ledger" does. Average that surprise over a passage and you have its perplexity.

Why it detects AI: generated text was built by choosing likely words, so a model re-reading it is rarely surprised — low perplexity is the residue of generation. Human writing reaches for odd verbs and takes tangents; the surprise score runs higher.

Burstiness: does your rhythm vary?

Burstiness measures variation — mostly in sentence length and structure — across a passage. Humans write in bursts: a long accumulating sentence, then a short one. For emphasis. Models converge on a steady medium: fifteen to twenty-five words, subject-verb-object, again and again.

Plot sentence lengths and human writing looks like a mountain range; model output looks like a picket fence. That flatness is measurable in one pass.

How detectors use them — and what supplanted them

Early GPTZero was nearly this simple: low perplexity + low burstiness → likely AI. Modern detectors use trained classifiers that learn hundreds of subtler features, but these two remain the interpretable core — and the reason formal human prose false-flags: academic and business English is deliberately low-surprise and low-variation. The metrics can't tell disciplined from generated.

Using the concepts on your own writing

This is why synonym swaps fail (they don't change either metric) and why structural rewriting works (it changes both). It's also a practical self-edit: scan your sentence lengths for picket-fence patterns, and ask where a reader could finish your sentence for you. A humanize pass automates exactly this variation — and the detector shows you the before/after instead of leaving it to faith.