Primary detector
desklib/ai-text-detector-v1.01
A verdict, not a score.
EU-resident AI-text detection with per-locale false-positive rates and signed review-ready scan details. For journals, HR, and academic integrity.
For academic integrity officers, journal editors, and HR teams under EU AI Act Article 22.
Humaniser engine / live
Analyze, understand, and revise
The same Humaniser workflow runs on every edition. The surrounding template changes; detector, editor, profiles, models, and API contracts do not.
Detect AI-generated text
Paste up to 2,000 characters. We return a verdict, not just a score.
Humanisation workbench / beta
Build a treatment profile, not a magic rewrite
Combine local rewrites and NLLB round trips, inspect every intermediate, then run the current detector on the result. Translation can change meaning: citations, facts and terminology still require author review.
Execution chain
Ordered transformations
One Turnstile check authorises exactly one chain run. Profile definitions stay in this browser; the complete chain is executed as one EU API job.
Runtime transparency
Model routes available to this workbench
The scan and run receipts are authoritative: they name the model actually used for that request.
Balanced local
humarin / T5-base route
Structural variation
dipper / optional 11B route
NLLB round trips
translation-*
Provider-resolved rewrite
claude-rewrite
A listed route is not a promise that it is loaded or enabled. Each result shows actual availability, model name, route, and missing evidence.
Common questions
Frequently asked questions
What is humaniser.eu?
humaniser.eu is an EU-resident AI-text detector and editorial humaniser operated by Ariada (Agonist Development AB, Sweden). It returns a detector state — low signal, review, or high signal — rather than proof of authorship, with review-ready scan details and published evaluation limits.
Where is data stored?
The submitted-text path is intended for EU processing. Until the public EU gateway and retention controls are independently verified, do not submit confidential, unpublished, or personal data. Scan details are not persisted unless a future authenticated retention mode is explicitly selected.
Is the verdict legally usable?
Each scan returns unsigned JSON details with the detector output and available model metadata. They support human review but do not prove authorship, originality, identity, or detector correctness. A human reviewer must make the final decision.
What's the false-positive rate?
False-positive rates are published per language on the calibration page. English text scored on the v0.3 model holds AUROC near 0.94 at FPR 5% on the RAID-style benchmark; rates for Mandarin-L1 and lower-resource European languages run higher and are disclosed honestly, not averaged away.
How is this different from Originality.ai?
Originality.ai publishes a percentage score and English-centric accuracy claims. Humaniser is being designed around explicit uncertainty, sentence evidence, EU processing, and published evaluation limits. Current scan details are unsigned; Merkle anchoring is not implemented or claimed.