Corrected August 30, 2026 · not an accuracy study
Our current detector data is a development set, not a blind benchmark
This page previously described an 81-text corpus and its train/test split as a publishable held-out accuracy study. That description was too strong. Model choices, feature experiments, the bias, and the operating point were selected after reviewing training, test, invariant, and full-corpus results. The reported test split therefore no longer qualifies as untouched evidence.
What the product actually returns
HumaText computes a hand-weighted 0–100 writing-pattern score from sentence-length statistics, vocabulary richness, formulaic phrases, and contraction rate. It is not a trained authorship classifier, a calibrated probability, or a prediction of any third-party detector. Sentence highlights are separate local heuristics and carry the same limitation.
What the corpus can support
The repository corpus remains useful for regression testing and for finding obvious failure modes. It cannot support population accuracy, precision, recall, false-positive, or calibration claims. Its sample provenance also needs an independent, source-by-source review before it can support a public evaluation. Historical rates are retained in the repository as development evidence, clearly labeled as non-blind, but are no longer advertised here.
What a valid study requires
- Documented, independently reviewable provenance for every label.
- Representative modern human writing, including ESL, technical, and short-form slices.
- Outputs from multiple current models, prompts, domains, lengths, and editing conditions.
- A preregistered scoring rule and decision threshold frozen before evaluation.
- A genuinely untouched test set evaluated once, with uncertainty intervals and per-slice error rates.
- Replication by someone who did not build or tune the detector.
How to use the detector today
Use it only as a private prompt to inspect formulaic writing. Do not use the score to accuse a writer, make an academic-integrity decision, or claim who authored a passage. You can inspect the current behavior on the detector page, which now labels the number as an uncalibrated pattern score.
HumaText is published by HumaText. This correction remains available at the prior URL so existing links resolve to the updated limitation instead of silently disappearing.