Methodology

Updated July 27, 2026.

Stylome treats an AI-text detector as a measurement system, not an authorship oracle. A public result must be tied to a named model, documented test population, fixed decision threshold, and known limitations. The classifier output is model evidence and must not be used as proof of who wrote a passage.

Freeze the evidence before tuning

Measurement holdouts are frozen before they are used to judge a model. Their normalized content hashes are audited against training and calibration inputs before a result is reported. If overlap is found, the affected measurement is not treated as independent evidence.

Compare at the same false-positive rate

Raw deployed accuracy numbers are not comparable when detectors use different thresholds. Stylome compares models at matched false-positive rates and reports the threshold, sample counts, and uncertainty needed to interpret the comparison.

Do not hide register shortcuts inside an aggregate

Results are decomposed by source and writing register—including news, reviews, educational writing, creative prose, email, and ESL-focused sets where available. Tail errors matter: an attractive overall score does not excuse concentrated false positives on one kind of writer.

Interpret the three authorship classes

Human written

The classifier found no material AI contribution.

AI edited

The classifier found retained human authorship with an AI contribution. This does not estimate the percentage written by AI.

AI generated

The classifier found fully machine-derived text, including machine-generated text rewritten by another model.

Confidence is separate from class. Low, medium, or high confidence describes relative KNN support from neighbor agreement, proximity, and binary decision margin compared with calibration examples. It is not a probability of correctness or proof of authorship. AI edited does not mean an inconclusive result.

Current deployed contract

The paste box and browser extension use Stylome 1.2 Large at the 0.01% calibration FPR target. Inputs contain 200–10,000 characters after whitespace-only normalization, with Unicode distinctions preserved and no silent truncation. The developer API accepts stylome-1.2-large or stylome-1.0-large. Only 1.2 accepts a 0.01% (default) or 0.1% operating point and provides relative KNN confidence. Large 1.0 retains its separate score-based response. Each model has a frozen scoring and input contract. A calibration FPR target is not a guaranteed error rate on new text. Historical benchmark reports retain the labels and thresholds of the model they evaluated.

The methodology and published measurements serve different jobs: this page documents the rules for earning a claim; the transparency report shows the attributable results that currently meet those rules.