Cobalynx answers with evidence: paste text for a calibrated probability with published error rates — never a guess — or check a file's provenance credentials. No signup, no interstitials.
Scan result
Calibrated AI probability
probability the text is AI‑generated
What this verdict means
Measured on a frozen held-out evaluation set — see the full evidence.
Sample measured
Text inspector
Sections of your text are banded by AI‑likelihood. Select a signal to highlight the sections where it fires.
low elevated high
A “likely AI” verdict is probabilistic evidence, not proof. False positives happen; never treat this result alone as grounds to accuse a writer.
AI-text detectors as a class are more likely to false-flag writing by non-native English speakers. Cobalynx measures and publishes its error rate on non-native (ESL) writing and holds it to the same gate as every other register.
No AI-text detector is fully reliable, including this one. In line with EU AI Act transparency guidance, Cobalynx publishes its measured error rates and reports calibrated probabilities, never certainty.
File provenance result
File checked
What the credentials say
Metadata signals — weak evidence
Document metadata is typed by software and trivially editable. These signals can hint at the tool that touched a file; they can never prove how its words were written.
Provenance evidence is asymmetric: verified Content Credentials are strong evidence about a file's origin, but missing credentials or clean metadata prove nothing — most files have neither, however they were made. Never treat an empty result as proof of human authorship.
No AI-text detector is fully reliable, including this one. In line with EU AI Act transparency guidance, Cobalynx publishes its measured error rates and reports calibrated probabilities, never certainty.
What our verdicts mean
Measured in one sanctioned run on a frozen held-out evaluation set — clean, non-native and adversarial slices together, so paraphrase-attacked and older AI text weighs on these rates (August 16, 2026; every run of the measurement is disclosed in the ledger), abstentions included:
We evaluated the detector in one sanctioned run on a frozen held-out set and committed the results — error rates with confidence intervals, not a marketing number. Every figure on this site traces back to that measurement, and every execution of the measurement command is disclosed in an append-only run ledger.
See the evidenceSome providers watermark their text; most do not. We track who actually does — and we will never claim to detect a watermark that was never embedded. Nobody can.
How watermarks workA wrong “AI” verdict can cost a student or writer dearly. That is why we refuse texts too short to judge, show a calibrated probability instead of a confident label, and abstain when the evidence is thin.
Our false-positive protectionsProvider states as reviewed on August 16, 2026 (served live by the detection service):
