canary trap

A canary trap is a mark planted in a work before it leaves your hands — a fictitious detail, a unique spelling, a planted phrase — so that if it appears somewhere else, it proves the copying happened.

A child who tells a secret to five friends, each with one tiny detail changed, knows who leaked when the changed detail surfaces. The trap does not stop the thief; it rings the bell and names the footprint. In publishing, the technique goes back to mapmakers' fictitious streets and dictionaries' invented words; in software, to plus-addressed email and security canarytokens; in the age of language models, to trap sentences planted in text before training, detectable by membership inference.

The honest limit the castle holds: detection and entitlement come apart by design. The trap proves the copying but does not prove the right — facts belong to no one under US copyright, so a fabricated fact presented as a fact is still a fact, and the copier walks free. The trap is a fact-maker, not a right; it spends wherever any law cooperates (contract, database right, reputation) and nowhere when none does. And as models grow and data is deduplicated, the ordinary author's planted trap becomes less detectable, not more — scale dilutes the footprint, and the larger choir drowns out the canary.

The castle's rooms that lean on it: seeded-fingerprint (the trap's legal anatomy — perfect detection, zero entitlement), what-the-seed-is-for (the six purposes the trap serves beyond the lawsuit it cannot win), and the-scaling-canary (the scaling question — does the trap survive as models grow?).