How the detector works
humaniser.eu does not score text on a single 0–100 axis. Each passage is tokenised, segmented, and evaluated against a calibrated classifier per language. The classifier emits a posterior probability, which is then mapped through a per-locale calibration curve — the same curves published on the calibration page — to produce a verdict: human, AI, or manual review.
The manual-review state is not a hedge. It is the model declining to make a call when the posterior sits inside the abstain band defined by the language's calibration card. Critics dispute whether detectors should ever abstain; we think the alternative — pretending to know — is what produced the false-accusation incidents that academic integrity offices are still cleaning up from 2023 onward.
Inputs accepted
- Plain text, up to 2,000 characters per scan in the web UI.
- 24 EU languages plus English. Auto-detection is on by default; override per scan if needed.
- Citation-rich academic text is preferred input — the calibration set includes a 4,500-sample arXiv subset.
What the receipt contains
The signed receipt is a JSON object with a SHA-256 hash of the input text, the
model SHA, the calibration-card reference, the verdict, the abstain-band
boundaries used, and a timestamp. The receipt is signed with the
humaniser.eu Ed25519 key whose public half is published at
/.well-known/humaniser-receipts.pub. The format and verification
steps are documented on the receipts page.