If you have replaced a lifestyle photoshoot with an AI-generated model this year, you are now carrying two disclosure obligations that ask for opposite things, and doing the Amazon one perfectly satisfies exactly none of the other one. That is the whole post, and it took me two months and a public correction to work out why.
Here is what happened. New York enacted S.8420-A/A.8887-B, amending General Business Law § 396-b, signed by Governor Hochul on December 11, 2025, and it took effect on June 9, 2026. It requires a conspicuous disclosure in any advertisement featuring a "synthetic performer" — an AI-generated human likeness that is not a real, identifiable person. Civil penalties are $1,000 for a first violation and $5,000 for each one after that, and because the statute contains no express private right of action, enforcement sits with the state, not with a plaintiff's bar. Cooley, Crowell & Moring, Skadden, Reed Smith and McDermott have all published client alerts on it. It is real, it is dated, and it has been live for nearly three months.
Separately, on July 27, 2026, Amazon began requiring sellers to add the keyword contains-synthetic-performer to the XMP metadata of listing images and A+ Content that feature photorealistic AI-generated people.
Two rules, three months apart, about the same pixels. One of them wants a field the shopper can never see. The other one wants a label the shopper cannot miss.
I am not a lawyer and this is not legal advice. Take the scope question to counsel. Take the creative question to whoever briefs your images, which is the part I can actually help with.
The correction I owe first
On August 4 and again on August 17 I wrote that the "label all your AI content by August 15" claim circulating in seller newsletters had no Amazon policy page behind it, and that anyone quoting a deadline was quoting something they could not produce. That was right about Amazon and it was the wrong shape of answer.
I kept treating the claim as noise because I kept looking for it in the wrong jurisdiction. There is a real, dated, enforceable disclosure obligation on AI-generated people in advertising. It is not Amazon's. It is New York's, it went live on June 9, and it does not care whether Amazon has a policy page.
That is why the claim would not die. Half the people repeating it had heard something true, garbled it into an Amazon deadline, and attached the wrong month. I was checking Seller Central. The thing was in the New York statute book.
The dumb take, and the real signal
The dumb take: "New York state law, I'm not in New York, next." The statute reaches any advertiser whose advertising reaches consumers in New York regardless of where the company is headquartered. Your detail page is served to New York. If your reasoning for ignoring this is geography, that reasoning does not survive one sentence of the statute.
The second dumb take, which is the expensive one: "we already did the metadata thing, we're covered." You are not, and the reason is structural rather than sloppy. Amazon's requirement is that a keyword sits in an XMP field inside the file. It is a machine-readable provenance signal. Its entire design intent is that the shopper never encounters it. New York's requirement is a conspicuous disclosure — the statute does not mandate wording, placement or size, but the whole point of the word conspicuous is that a consumer sees it. Those are not two versions of the same job. One is a hidden field, one is visible copy, and the first cannot satisfy the second under any reading.
I wrote in August that the metadata block is a compliance artifact living in a field any Save-for-Web preset silently strips. That is still true and it is now the smaller of the two problems, because at least that one is a field Amazon told you about.
The real signal: the compliance layer on your creative just split into two, and the second half has to live in the frame.
What is actually in scope, because most coverage gets this wrong
The narrowing matters more than the rule, so be precise about it.
A synthetic performer, per the statute, is a digitally created asset — created, reproduced or modified by computer using generative AI or a software algorithm — intended to create the impression of a visual performance by a human performer who is not recognizable as any identifiable natural person.
That means:
- An AI-generated human model in a lifestyle frame is in scope. This is the one brands are actually doing.
- AI used to enhance or retouch a real human model is not. You photographed a person, you cleaned it up. Different thing.
- AI-generated imagery with no human in it is not in scope. Your AI background fill, your generated countertop, your synthetic studio sweep, your infographic — none of that is a synthetic performer. There is no person in the frame.
- Hands count. If the only human element in your scale-reference frame is a generated hand holding the product, you have a generated human likeness in an image. I would not want to be the one arguing that a hand is not a visual performance, and I would not want a client of mine arguing it either.
- There are carve-outs for audio-only ads, for marketing of expressive works, and for platforms disseminating third-party advertising, and the obligation attaches where the advertiser has actual knowledge. On a listing you commissioned, actual knowledge is not a defence you are going to reach for.
The genuinely unsettled part, and I am going to hold it as unsettled rather than pretend otherwise: whether an Amazon product detail page image is an "advertisement" under § 396-b has not been tested. The statute reaches advertising content broadly and excludes audio-only. Several of the law firm alerts advise retailers and marketplace sellers to audit product and lifestyle imagery in listings accessible to New York consumers, which tells you where sophisticated counsel is landing. Nobody has litigated it. Anyone telling you it is definitively covered or definitively not is telling you what they think, not what a court has said.
What actually changes for a $200K/mo brand
Four things, in the order they will bite.
1. The disclosure has nowhere to go except into the image. Amazon does not give you a caption field, an overlay, or a disclosure line under the carousel. There is no UI for this. If you conclude you need a conspicuous disclosure on a frame containing a generated model, the only surface available is the pixels. That is a creative decision, it consumes frame area, and it lands on the exact images that were generated to save money on a photoshoot.
2. It cannot go on your main image at all. Main image rules prohibit text, graphics and overlays. So a synthetic performer cannot appear on a hero without creating a conflict you have no way to resolve inside Amazon's own rules. The practical answer for most brands is the simple one: do not put a generated human on your main image. You should not have one there anyway — a hero with a person in it is usually failing the identification job at 280px before any of this comes up.
3. The frames most likely to carry a generated person are frames doing merchandising work. In most stacks the lifestyle frame is also, whether anyone intended it or not, the first scale reference the shopper meets. Adding a disclosure line to it is not free — it is type competing with the one message that frame is responsible for. This is not a reason to skip the disclosure. It is a reason to decide where the line goes on purpose, at a size that is legible at thumbnail without eating the callout, instead of having a designer drop 9pt grey text in a corner where it is neither conspicuous to New York nor invisible to your composition.
4. Your exposure is not one listing, it is a template. If a generated model appears in an A+ module chassis applied across 40, 80 or 140 ASINs, one decision multiplies. Penalties escalate per violation. I am not going to compute a number for you because I do not know how a regulator counts a violation across a catalogue, and neither does anyone else writing about this. What I will say is that the shape of the risk is the same shape as every other template problem: a single over-reaching decision that was cheap to make once and expensive to have made eighty times.
None of this is a Q4 emergency. It is a catalogue audit with a real statute behind it and no deadline pressure, which is exactly the kind of work that never gets done.
What I'd do this week
1. Find out whether you have any generated humans at all. Most brands do not, and the audit ends in twenty minutes. Ask your creative vendor a single written question: has any image or A+ module you have delivered to us in the last twelve months contained an AI-generated human figure, including hands? A vendor with a workflow answers with a yes or no and a list. A vendor who answers with a paragraph about their commitment to authentic imagery has told you something too.
2. Separate the three questions everyone is merging. Was it generated. Is there a person in it. Is it substantiated. Only the second one triggers New York. Only the third one gets your listing suppressed. A generated background is not a synthetic performer and an unsupported "clinically tested" is a problem regardless of who typed it.
3. If you have generated people, make the disclosure a briefed element, not an afterthought. Decide the wording once, decide the placement once, and put it in the creative spec so it renders at a legible size and does not get dropped by whoever exports the next batch. Get the wording from counsel — the statute does not prescribe it, which sounds like freedom and means the burden of judgment is yours.
4. Fix the metadata separately and stop assuming it travels. The contains-synthetic-performer XMP keyword survives your editor and frequently does not survive an export preset. Add one line to your vendor question set: does your delivery process preserve XMP metadata. It is a different obligation with a different failure mode and it takes one email to close.
5. Add a provenance column to the creative change log. Per asset: who made it, what model if any, does it contain a human figure, is the metadata present, is a disclosure rendered. The pattern across every one of these stories this year is identical — the fix is never clever, it is find the artifact, write it down, put a date on it.
What I'd ignore
The "AI content penalty" discourse. Two years old, never once traced to an Amazon policy page. Nothing about a disclosure obligation makes your listing rank worse.
Anyone selling an AI content compliance audit. It is a spreadsheet with five columns and three emails to your creative vendor. I just published it free.
The urge to strip AI out of your creative pipeline entirely. The rule is narrow. It is about generated people. Background generation, concept iteration, infographic layout, upscaling, retouching a real model — none of that is in scope, and pulling AI out of your production workflow because of a statute about synthetic performers is an overcorrection that costs you real speed for zero risk reduction.
The federal-preemption argument. There is a genuine live debate about state versus federal AI regulation and it is interesting and it changes nothing you ship in September. If it resolves, it resolves after Q4.
Panic. This has been live since June 9. If you were going to get a letter for the imagery you have been running all summer, the shape of that risk did not change today because I wrote about it. Fix it deliberately in the next few weeks.
The part worth keeping
Every governance practice I have published this year assumes an artifact you can go and inspect. Pin the model string, because a string is a contract you can read. Diary the price, because pricing lives on a page. Add a retirement column, because shutdowns have dates. Check the XMP field, because metadata is something you can open a file and look at.
This one is the first where the artifact is a statute, and the interesting thing about statutes is that they do not get announced to you in a changelog, they do not appear in Seller Central, and nobody in your vendor chain has any obligation to mention them. A law took effect on June 9 that governs a creative decision a lot of brands made in April to save four thousand dollars on a photoshoot, and the first most of them will hear about it is a blog post.
I spent two months telling people the labelling claim was unsourced. I was checking the wrong filing cabinet.