Skip to main content

Comparison · authorship · AI detection

An AI detector asks how the text was made. Kaneme asks whose voice survived.

Both may return a probability. That doesn’t make the probabilities interchangeable. Start with the evidence each tool sees.

Three questions that should stay separate

Universal detector

Was this produced by a machine?

A broad inference made without knowing the author.

Kaneme Q1

Does it carry your voice rather than a generic AI voice?

Measured against imitations built from the target writer’s material. Published AUC: 0,992.

Kaneme Q2

Was it written by you rather than another person?

The harder comparison between human writers. Published AUC: 0,887.

Reliability needs a named opponent.

“Accurate” means little until you say what the system had to tell apart. Kaneme publishes human-versus-human and voice-versus-AI results separately.

The benchmark holds authors back before tuning, tests real writing, and is rerun on the engine users actually receive. Read the corpus, protocol and failure modes on the method page.

What actually differs

CriterionAI detectorKaneme verdictLocal pattern review
InputThe text aloneThe text plus the owner’s own writingOne text, reviewed locally
OutputUsually an AI probabilityA calibrated authorship probability or a named abstentionVisible writing patterns, with no author verdict
OriginInferred from the textReported by the product when known; never guessedOutside scope
Best usePolicy, moderation, volume filteringProtecting and proving your own writing voiceRevising generic passages before publication

Put the limit next to the number.

A probability isn’t proof

Don’t turn a probabilistic signal into an accusation.

Short writing weakens the signal

A high match may still help. A low match on a short text is an absence of evidence.

French is the validated domain

An English interface doesn’t extend the production engine’s language coverage.

Frequently asked questions

Is Kaneme an AI detector?+

No. A detector looks at a text and guesses how it was produced. Kaneme starts with writing supplied by its owner, then asks whether a French candidate carries that same voice.

Can the result prove misconduct?+

No. Treat a probability as evidence to review, never as an accusation. Kaneme also needs the voice owner’s reference writing, which limits who can run the comparison.

Why does Kaneme sometimes return no number?+

Because guessing would be worse. The engine abstains when the language or register falls outside its validated ground, the candidate overlaps the reference, or a required layer is unavailable.

When should I use an AI detector?+

Use one for policy, high-volume moderation, or an initial filter when you have no author reference. Keep a person in the loop: its reliability for one particular writer is usually unknown.