[demark]
PRIVACY
TEXT FORENSICS / AI DETECTION

Machine text leaves traces. This looks for them.

Paste text. Every registered detector runs over it and reports its evidence — how the text is written, AI-disclosure labels it carries, hidden characters embedded in it — annotated in place. Each signal is shown with what produced it, and none of them is proof.

─ EVIDENCE SHOWN INLINE ─ MEASURED ON A PUBLIC DATASET ─ TEXT NEVER STORED WITHOUT CONSENT
0 chars
or press Ctrl↵
WHAT THE ANALYSIS LOOKS AT
01 How the text is written

A learned stylometric score, plus word-frequency shape and excess-vocabulary markers. Probabilistic and wrong at a measured rate; the learned score abstains under 50 words, the other two under 30 — below that the app says it cannot judge, which is not the same as clearing the text.

02 Disclosure labels

Mandated AI-disclosure labels and model self-disclosure boilerplate. Declared provenance, not a covert mark, and scored below the strong-signal threshold on purpose.

03 Hidden characters

Zero-width characters, bidi controls, exotic spaces, homoglyphs — carrier channels for hidden payloads. Deterministic: a hit is a hit, but it shows something was embedded, not that a machine wrote it.

Detection is measured against a research pipeline: public corpora of AI-generated, watermarked and human text, generation harnesses for SynthID, KGW and related schemes, and user-contributed labels. The families are reported side by side, never merged into a single claim about who wrote the text.

DEWATERMARK / REMOVE CARRIER CHARACTERS

Strip the invisible layer.

Removes zero-width characters, normalizes exotic spaces, and maps homoglyphs back to ASCII. That clears the character-level carriers and nothing else: the wording is untouched, so stylometric, statistical and disclosure signals survive cleaning.

0 chars