← All field notes

When your AI character looks too much like a real person

Right of publicity, defamation, and the “reasonable person” test — the three risks every team shipping generative video should understand.

Generative tools make it trivial to spin up a “fictional” character. But fictional is a claim about your intent, not about what lands on screen — and a synthetic face can resolve onto a real person's likeness by accident. When it does, three legal doctrines come into play.

1. Right of publicity

Real people — and especially public figures — control the commercial use of their likeness. The exposure is highest when two things are true at once: the character carries a recognisable likeness, and you are monetising it — selling a product, running ads, or building a brand around the character. Unauthorised commercial appropriation is the core claim.

Two things surprise teams here. First, in the U.S. there is no single federal right of publicity; it is state law, ranging from robust (California, New York, Tennessee) to nonexistent. Second, “likeness” is broader than a photo-match. Courts have found a person's identity evoked through a distinctive voice (Midler v. Ford) or through unmistakable association alone, with no name and no face (White v. Samsung). A character that “just happens to look like” someone can still implicate it.

2. Defamation

If the character is portrayed as malicious, incompetent, criminal, or otherwise disreputable, a real person who is identifiable in it can claim the portrayal harms their real-world reputation. Defamation generally needs a false statement of fact, communicated to others, that is “of and concerning” the plaintiff. Labelling something fiction is not an automatic shield: the question is whether reasonable viewers would understand the character as referring to the real person.

3. The “reasonable person” test

Across both doctrines the pivotal question is the same — would a reasonable audience conclude the character actually references a real person, or read it as coincidence? Context decides it. A generic resemblance is one thing; a resemblance plus the person's profession, setting, mannerisms, or storyline is another. The more corroborating signals stack up, the more a coincidence reads as a reference.

The AI-specific wrinkle: the law is catching up fast

A wave of new law targets synthetic likeness directly. Tennessee's ELVIS Act (2024) extends protection to voice and likeness with AI explicitly in scope; California's 2024 digital-replica laws regulate AI replicas of performers; and the proposed federal NO FAKES Act would create a nationwide digital-replication right. The direction of travel is clear: “the model generated it” is not shaping up to be a defence.

What to actually do

You cannot eyeball this at the scale modern pipelines produce, and “we didn't intend it” does not help much after publication. Three practical moves:

  • Screen before you publish. Check generated and user-generated content for real-person resemblance — and for copyrighted characters — as a step in your pipeline, not a fire drill after a complaint.
  • Treat a match as a lead, not a verdict. The technology is probabilistic and over-eager to put a name to a face. A human should make the call, with the evidence in front of them.
  • Keep the receipt. A timestamped record of what you screened, and what you found, is itself part of a diligence defence.
Screen before you publish

Catch it before it ships.

Face Check screens user-generated and AI-generated video for exactly this — faces that resemble real public figures, and copyrighted characters — as candidates for review, never a determination, with the frame attached.

See how it works