AI Detection Score

An AI detection score is the number a detector reports — "85% AI," "likely human," a probability, or a highlighted percentage of the document. It's easy to read it as a measurement of fact, but it's an estimate produced by a statistical classifier, and understanding that changes how much weight it deserves.

Different tools express the score differently. Some report the probability that the whole document is AI-generated; others, like Turnitin, report an estimated percentage of the text that appears machine-written; others give a per-sentence heat map. Crucially, these numbers aren't comparable across tools — 60% in one detector is not the same quantity as 60% in another, because they're trained on different data with different thresholds.

That's why detectors so often disagree on the same passage: each is a different model with a different decision boundary. A score is a confidence estimate from one classifier, wrapped in a precise-looking percentage that overstates its certainty. It can be right, wrong, or borderline, and it shifts when the tool is retrained.

The practical takeaway is to treat any score as one uncertain signal, never as proof. If you're humanizing text, moving the underlying perplexity and burstiness will usually move the score — SynthGuard shows a live burstiness reading so you can watch it change — but chasing a specific number in one tool is a mistake, because another tool may score the same text completely differently.