Amazon has added separate "AI-generated" and "AI-generated people" checkboxes to the upload flow sellers use for enhanced product content. The trigger is a New York law, S8420-A, that took effect on June 9 and requires the creator of a commercial advertisement to conspicuously disclose an AI-generated synthetic performer, with civil penalties of $1,000 for a first violation and $5,000 for later ones. One state rule, one global intake change.
Key facts
- The penalty: $1,000 for a first violation, $5,000 for each subsequent one, under New York General Business Law as amended.
- When: the law was signed December 11, 2025 and took effect June 9, 2026; Amazon's seller-side guidance surfaced publicly in July 2026.
- Who: New York State, with Governor Hochul signing; Amazon implementing on the seller side.
- Primary source: the enacted bill text and the Governor's announcement.
The law is narrower and better-drafted than the coverage suggests. It targets a "synthetic performer" - a non-identifiable, human-looking visual or audiovisual performance generated by AI or an algorithm - in a commercial advertisement, where the creator has actual knowledge. That is three limiting conditions stacked before the obligation attaches. It expressly excludes audio-only advertisements, excludes AI used purely to translate a real human performer into another language, and excludes advertising for expressive works where the synthetic performer is consistent with the underlying work, so a game trailer featuring the game's own AI-rendered character does not need a disclosure. It leaves Section 230 protections intact and exempts distribution media from liability when a third party fails to disclose. The obligation sits with whoever made the ad.
That last design choice is what makes the Amazon side of the story interesting. New York did not order Amazon to change anything globally. Amazon is a distributor, and the statute deliberately keeps distributors off the hook for a seller's omission. But a platform facing a state disclosure rule has two options: build machinery to identify which of hundreds of millions of assets might be seen in New York and chase down their provenance after the fact, or capture the seller's declaration once, at upload, everywhere. Amazon appears to have picked the second. It is cheaper, it is auditable, and it scales to whatever the next jurisdiction requires.
The line Amazon has drawn is more specific than the statute's language, and it is the part sellers actually need. In a public Seller Central response, an Amazon representative said photorealistic AI-generated people must be disclosed, while ordinary AI retouching - background removal, colour correction, lighting adjustment - is treated as conventional editing so long as the product remains accurately represented. The representative also described the requirement as prospective rather than retroactive to already-published material. In other words: the boundary is a fabricated human, not a fabricated pixel. Everyone has been using algorithmic enhancement in product photography for years, and no disclosure regime survives contact with reality if it tries to flag all of it.
The honest limit on what can be verified here: Amazon's own public policy page for this is JavaScript-gated in a plain viewer, so the directly readable primary confirmation is the staff response in the Seller Central forum, not a formal policy document. Claims circulating about the exact metadata field, automatic buyer-facing badges, enforcement mechanics, or worldwide coverage go beyond what the readable primary material supports. And Amazon has not said it was legally compelled to standardise globally - that is a reasonable inference from how the statute allocates liability, not a company statement.
Support and opposition both landed roughly where you would expect, and both were narrower than the rhetoric. SAG-AFTRA celebrated the signing, framing it as a narrowly tailored way for viewers to know whether a spokesperson is a real person - the performers' union's interest here is as much about work being substituted as about consumer confusion. Governor Hochul framed it as protecting both consumers and the creative workforce. The New York State Broadcasters Association's objection was procedural rather than anti-disclosure in principle: it noted that amendments removed audio ads, shifted responsibility away from distributors, and added the expressive-work carveout, while still calling the definition broad enough to potentially sweep in many computer-generated human images.
That definitional worry is the real open question. "Non-identifiable human-looking performance generated by AI" has fuzzy edges in an industry where the same asset might be composited, retouched, upscaled and partially regenerated. Amazon's photorealistic-person test is a workable practical line, but it is a platform's interpretation, not a court's.
The broader pattern is worth watching regardless of how this specific statute is enforced. Disclosure obligations are landing at the point of upload rather than the point of display, which pushes the labelling burden upstream into creation tools and marketplace intake forms. The federal AI Labeling Act now before Senate Commerce would extend the same logic nationally, requiring both visible and machine-readable disclosure and obliging platforms to preserve provenance signals rather than strip them. Whether the honest answer to "was a human in this ad?" ends up as metadata that survives a re-encode is an engineering question that the law has not yet had to confront.
Originally published on Ground Truth, where every claim is checked against the primary source.
Top comments (0)