A team generates a character “based on no one.” It still lands on a face a viewer recognises. That is not bad luck — it is how the models work.
The model's prior pulls toward faces it has seen
Generative image and video models learn from enormous datasets full of real, often famous, faces. Ask for “a confident forty-year-old executive” and the model samples from a distribution shaped by the faces it saw most — which skews toward well-photographed public figures. The result can approximate, and occasionally closely reproduce, a real person, with no name anywhere in the prompt.
“Nearest known identity”
Recognition runs the other way too. Ask a vision-language model “who is this?” about an ambiguous synthetic face and it will often answer with the nearest famous person it knows — confidently, and wrong. We have watched a model name a generic animated character as a specific franchise lead it merely resembled. The lesson is blunt: a name from a model is a lead, not a fact.
Why you can't catch it by looking
At one or two assets, a person can eyeball resemblance. At the volume a modern pipeline produces — hundreds of clips, thousands of sampled frames — eyeballing does not scale, and the riskiest matches (a brief background appearance, a stylised near-miss) are exactly the ones a human skim misses.
What good screening looks like
It flags resemblance and copyrighted characters; it ranks findings by how sure it is rather than how loud; and it presents everything as a candidate for review with the evidence attached, so a human makes the call. It never asserts “this is so-and-so” — it says “this resembles so-and-so; verify.” That honesty is the difference between a useful screen and a liability of its own.