There is an ad I have watched a dozen times and never once questioned. A woman in a bright kitchen picks up a jar, tells me it fixed her mornings, and smiles in a way that makes me not mind being sold to. I never wondered whether she was real. I am not sure I registered her as someone the question applied to.
What bothers me more, and it took a new California law to make me see it, is who else was in that shot. Behind her, out of focus, sat four or five people at café tables. I never looked at them, and I have no idea whether any of them existed.
Neither do you, about the ones in yours.
What the law says, and the adverb it hangs on
SB 1050 is not pending: the Legislature's file records it as Chapter 246 of the 2026 statutes, signed September 16 at the headquarters of SAG-AFTRA, its sponsor. Gavin Newsom's line: Californians deserve to know when the person selling them something isn't a person at all. With no urgency clause, counsel read it as taking effect on January 1, 2027. California is the second state to require the label, after New York.
The prohibition is one sentence, and it all rides on one adverb: it is unlawful to publish "an advertisement that prominently includes a synthetic performer" without a clear and conspicuous disclosure. A synthetic performer is a generated figure or voice realistic enough to read as human and "not recognizable as any identifiable natural person."
The statute gives three ways to be prominent: "(A) In the foreground and demonstrating or illustrating the product or service. (B) Providing or voicing the on- or off-camera narration or commercial message. (C) Illustrating or reacting to the on- or off-camera narration or commercial message."
All three describe a figure doing something to the sales message. (A) is the woman with the jar. (B) is the voice. (C) — "illustrating or reacting to" — is the fuzzy one, and where the whole argument lives. No court has construed it.
It even prescribes the wording: "this performance features a synthetic performer," or "no human performer is depicted." Sit with the second. An ad with a live spokesperson and a wholly generated crowd could not honestly say it — my reading, not anybody's holding.
The people who have to comply read it the other way
Now the concession: the strongest objection comes from the people writing the compliance memos.
Four advertising trade bodies asked Newsom on September 4 to veto the bill as "an overbroad mandatory labeling regime." Their complaint is the mirror image of mine: they say it "mandates disclaimers even when a synthetic performer is used for purely illustrative, background or thematic purposes." They read prong (C) as sweeping the café in. Their second argument: disclosure overload — label everything and you have degraded the label. So prong (C) is contested, not settled, and the firms advising advertisers are buying insurance against the broader reading.
One state over, New York's version has been live since June 9 with no prominence threshold at all: there, even background performers are covered. Whether that reaches digital extras is a prediction, not a rule: counsel read New York's undefined "performer" as a word that may be construed expansively. In its first two months, Straight Arrow News reported, only two companies had faced a consumer complaint — allegations, not findings — one an athletic clothing company, the other an eyewear brand, both accused over AI-generated models. Foreground figures again.
Brussels never asks who is standing in front
Europe wrote the same rule from the other end.
Article 50(4) of the EU AI Act binds any deployer of "an AI system that generates or manipulates image, audio or video content constituting a deep fake," and asks nothing about prominence. The Commission's example of a deployer is an advertising company — not the animator who rendered the crowd. The obligation has applied since August 2, 2026. And "deepfake" there does not mean impersonating a celebrity — the subject need only be someone who "can plausibly exist," which is why US counsel tell advertisers that EU compliance means labeling product shots, backgrounds, and human as well as non-human characters.
Europe's rule is vaguer than that sounds. Generated "background scenes, special effects" in standard production are, the Commission says, "not likely" to make content falsely appear authentic — a hedge aimed at movie production, but soft all the same. The test can also turn on expectation: where an audience never expected authenticity, content "may not falsely appear" to be. An artistic carve-out carries a lighter duty, though the guidelines largely shut it to advertisers: exclusively commercial content cannot benefit from the attenuated transparency regime. Two lines, two jurisdictions, both leaving something outside.
The ad nobody has to label yet
Follow the incentive. From January 1, a synthetic star costs a California advertiser a disclosure; a synthetic crowd — on my reading of prong (C), though not the trade bodies' — may cost nothing. Which line does a budget holder cut?
Picture the 2029 spot: one human host, booked cheaply, holding the product and saying the words, because she is the compliance surface. Everything behind her — the café, the eleven people in it, the weather — rendered at the price of compute. Fully compliant, almost entirely unreal. One casting company alone lists 200,000 background actors, by its own count, and its executive does not think synthetic characters will eliminate the need for them. He may be right. But a rule that prices one kind of synthetic person and not the other is not neutral about which kind gets made.
Who else is uneasy, and from which direction
Not a left-right fight, which is the interesting part. The market-liberal Reason Foundation, testifying on this bill, argued for tailoring disclosure to cases where the speaker's identity is likely to matter rather than imposing it in all cases — essentially an argument for the prominence test California adopted. The pro-disclosure Transparency Coalition, from the other corner, counts the states.
Tech Policy Press names four failure modes; two of them are the ones a partial label invites: banner blindness, because a label everywhere is a label nobody reads, and false reassurance, because users may infer that unlabeled content is authentic. A survey experiment with a German sample of 877 found an implied authenticity effect: exposure to labeled content "slightly increased" perceived authenticity in unlabeled images, about one-fifth the size of the direct effect. Small, directional, one study — but it points the wrong way for a rule that labels some synthetic people and not others.
So what do you actually do with this?
Read the label as a claim about one person, not the whole ad. "This performance features a synthetic performer" never says how much of it was. "No human performer is depicted" does.
Treat a missing label as weak evidence, not proof. In California it can mean nobody synthetic carried the sales message; in New York it means more; elsewhere, nothing at all.
If you brief or approve advertising, change the question from "is our spokesperson AI" to what each generated figure does to the commercial message — your own trade bodies have told a governor they read the statute as reaching background uses.
If you perform for a living, the state on the call sheet matters. New York's text is read as covering background performers; California's is keyed to prominence; Europe's to the content.
My vote? Follow the frame, not the face.
Read the law's own stated interest: consumers have an interest in knowing when an ad uses a synthetic performer "in a manner that could reasonably be interpreted as a human performance." Perception, in other words — and perception does not check who is holding the product. The people at the café tables read as a human performance too — the entire reason they are in the shot.
A federal cloud hangs over all of it: a December executive order told Commerce to flag state laws that compel … deployers to disclose information in a manner that would violate the First Amendment. As of September 19, 2026, neither law has been challenged in court — I would not assume that holds.
A good first law, then: it defines its terms, prescribes the wording, exempts trailers and translation, and bans nothing. Aimed at the right liar, drawn at the wrong depth of field. California has until January 1 to learn whether prong (C) reaches further than I think. If not, the fix is one sentence long.
Next time an ad works on you, count the people in it nobody has to disclose — then pass this to whoever writes the briefs. The HAIA Foundation keeps reading the fine print out loud, one adverb at a time.