Similarity is a lead, not a lineage.

A technical match can identify something worth reviewing. It cannot, by itself, establish derivation, access, permission, authorship or ownership.

Original editorial illustration of several similar blue crystalline artifacts surrounding a central artifact, while only one candidate is connected to it by a precise amber chain of transparent transformation stages.
Original conceptual illustration.
Watch the 48-second attribution explanation

Resemblance starts an investigation

Two passages share unusual language. Two code fragments have the same structure. An image embedding places two assets close together. These signals can be valuable: they can narrow a large search space and identify a relationship worth examining. But the match is not the history.

Similarity does not show that one person accessed the other asset. It does not identify which version existed first, what transformations occurred, whether the shared material came from a common source, or whether a use was authorized. A high score can still be evidence of independent creation, a shared template or a method that is conventional in the field. The score is a lead—not a conclusion about lineage.

Provenance records relationships, not resemblance alone

The W3C PROV data model offers a useful vocabulary for describing entities, activities, agents, usage, generation, derivation and association. The point is not that every creative process must adopt one technical format. It is that a lineage claim needs relationships across a process: what existed, what happened, who or what participated, and what resulted.

C2PA’s Content Credentials provide another implementation example for digital media. A C2PA manifest can bind signed claims and assertions to an asset and can reference other manifests. Those structures can preserve statements about provenance and editing history. They are trust signals, not automatic proof that every assertion is true, complete or legally sufficient.

Neither framework turns a similarity score into a recorded chain of events. If the chain was never observed or captured, a later match may justify review—but it should not be rewritten as provenance.

Keep the technical finding separate from the rights decision

Attribution systems also need a boundary between technical evidence and legal or organizational authority. The U.S. Copyright Office’s January 29, 2025 copyrightability report emphasizes that human contribution to AI-assisted work is case-specific. A record of model use, a prompt or a similar output does not by itself answer which expressive elements a person controlled.

The same separation applies to permission. A provenance record may identify a source and a transformation without granting a license. A license record may authorize a use without proving that one asset was derived from another. Ownership, authorship, contractual permission and technical lineage are related questions, but they are not interchangeable.

Build an attribution dossier that can be challenged

I would treat similarity as the first field in an attribution dossier, not the final verdict. The dossier should include the candidate match and method; precise evidence spans; the versions compared; documented access or custody; the transformation sequence; human and machine contributions; relevant permission or license records; uncertainty; independent review; and a challenge path.

This is an editorial synthesis, not a W3C, C2PA or Copyright Office standard. It is also not a report of measured performance. A proposed evaluation could give reviewers the same candidate relationships in two forms: similarity scores alone and scores accompanied by the dossier. Researchers could measure false lineage claims, missed relationships, reviewer agreement, review time and legitimate reuse decisions.

The practical rule is available before that study: let similarity decide what deserves investigation. Require evidence of the process before calling it lineage.

More from the blog RSS feed