New York's Synthetic Performer Law and Amazon's AI Tag

Quick answer: since 9 June 2026, New York General Business Law 396-b requires anyone who produces or creates an advertisement for a commercial purpose to conspicuously disclose in that ad that a "synthetic performer" is in it, where they have actual knowledge, on penalty of $1,000 for a first violation and $5,000 for any subsequent one. A synthetic performer is an AI-generated human performer who is not recognisable as any identifiable real person, which is exactly what an AI UGC actor is. Amazon then turned that statute into a file requirement: sellers were told to embed the keyword contains-synthetic-performer in the dc:subject XMP metadata field of qualifying images and videos before upload. Two things follow. Disclosure lives in the creative for the law, and in the file's metadata for Amazon, and those are different jobs.

The statute, in its own words

Governor Hochul signed S8420-A/A8887-B at the SAG-AFTRA New York office on 11 December 2025, describing the pair of bills as ensuring "we are fully transparent when using images generated by artificial intelligence". The synthetic performer bill took effect 180 days later, on 9 June 2026.

The operative sentence is short. From the bill text:

"Any person engaged in the business of dealing in any property or service who for any commercial purpose produces or creates an advertisement ... shall conspicuously disclose in such advertisement that a synthetic performer is in such advertisement, where such person has actual knowledge."

And the definition that decides whether you are in scope:

"A digitally created asset created, reproduced, or modified by computer, using generative artificial intelligence or a software algorithm, that is intended to create the impression that the asset is engaging in an audiovisual and/or visual performance of a human performer who is not recognizable as any identifiable natural performer."

Read that emphasis twice, because most summaries skate over it. The law is not about deepfakes of real people. It is about invented people. A face swap of a named creator is governed by other law entirely. A photorealistic presenter who exists nowhere, which is the standard output of every text-to-video model on the market, is the thing this statute was written for.

Who it hits, and who it does not

The bill carries three carve-outs and one important allocation of responsibility. Taken straight from the text:

The medium exemption is why the burden lands on you rather than on Meta or TikTok. Those platforms label AI content under their own policies, and Meta applies its AI label itself for ordinary ecom ads, but a platform label is not a disclosure you made in your advertisement. Do not treat one as satisfying the other.

The bigger practical gap is that the statute never defines "conspicuously". Cooley's analysis of the law flags the same thing, noting the statute leaves the content, placement and formatting of the disclosure to be worked out in practice. Until that settles, the defensible version is the boring one: a legible on-screen line in the creative itself, present long enough to read, not buried in a caption that travels separately from the file.

Amazon turned the law into a metadata field

In late July 2026 Amazon sent sellers a notice. Per Forbes, Amazon told sellers that "recent legislation requires disclosure when images or videos in advertisements contain photorealistic AI-generated people", directed them to embed an IPTC-compatible metadata keyword in qualifying assets, and said it plans to add a shopper-facing indicator on listings where applicable.

The specifics, as reported by eWeek and by the Amazon Sellers Attorney breakdown:

One caveat you will not see stated elsewhere, so state it yourself before you act: Amazon has published no public page for this. We looked. The notice is behind the Seller Central login, so every account of the keyword and the field, including this one, is secondhand. Verify the exact string against your own Seller Central notice before you write it into a batch process. And note the practical point the Amazon Sellers Attorney piece makes plainly: nothing in Seller Central asks you the question. The disclosure travels inside the file, which means nobody will prompt you and nothing will fail loudly if you skip it.

Which of your renders actually trips this

Here is where our own product data is useful, because the answer depends on what the model produced, not on which tool you used. Across the models we run, three cases behave differently:

The middle row is the one that will cost people time. Our own terms state that where a likeness derives from a real performer it is used under licence, and we do not provide likenesses of real, identifiable people who have not licensed them. That is a good position on rights and a genuinely ambiguous one on disclosure, because a licensed likeness is AI-rendered and based on a real person at the same time. Neither the statute nor the reported Amazon guidance answers it. If your creative depends on that case, ask a lawyer rather than a blog. If you want the simple path, an invented presenter is unambiguous: it needs the disclosure, and you can ship it knowing which rule applies. Rights are a separate chain from disclosure either way, and who actually owns an AI-generated video adfollows that one provider by provider, because no tool's terms can override the model's.

The metadata problem nobody warns you about

An image editor writing dc:subject into a JPEG is easy. Video is where this breaks, and we can be specific because we can read our own pipeline.

When our editor renders a finished clip, it re-encodes to H.264 with -movflags +faststart so the file starts playing immediately. It does not pass -map_metadata, which means source metadata is not carried into the output. That is normal. Most render and edit paths behave the same way, because re-encoding rebuilds the container from scratch. The consequence for you is concrete: do not assume a tag written earlier in your workflow survives to the file Amazon receives. Write the keyword onto the final export, after the last re-encode, and check it with a metadata tool before upload. The same instinct applies to any provenance metadata a generator attaches, and we measured that case on this very pipeline: a C2PA manifest injected at the position the spec mandates did not survive even the lossless stream copy we run on ingest, let alone the download re-encode. The full test, and what did survive, is worth ten minutes if you are relying on any embedded marker.

New York and Brussels overlap, and Brussels reaches further

If you sell into both markets, do not assume one rule picks up where the other stops. They overlap, and the invented presenter is caught twice. Article 50(4) of the EU AI Actsays deployers of an AI system that generates or manipulates image, audio or video content "constituting a deep fake" must disclose that the content has been artificially generated or manipulated, and it has applied since 2 August 2026. The European Commission's FAQ on Article 50 states the resemblance test in one sentence, and the whole sentence matters: simulated persons, objects, places, entities or events "need to resemble someone or something that exists, can plausibly exist or could have plausibly existed in reality". Cut that quote at the first comma, as a lot of coverage does, and you get the opposite answer to the one the Commission gave.

The Commission's Guidelines on Article 50(C(2026) 5054 final, 20 July 2026) then define the term the whole question turns on. At paragraph 113, "persons" means "realistic, human beings (including digital replicas of real persons, realistic AI-generated human avatars or personas, and personal characteristics or expressions, such as image, voice, behaviour, performances etc.)". At paragraph 114 the Guidelines list "fully AI-generated actors" among the uses of AI that are likely to make content falsely appear authentic, adding that a high degree of photorealism "renders it more likely that such content should be considered a deep fake" without being determinative on its own. A photorealistic presenter invented by a model is a realistic AI-generated human persona. It sits inside the EU deep fake definition.

So the two regimes are not mirror images. New York's trigger is the narrower one: it reaches the invented performer and expressly stops at anyone "recognizable as any identifiable natural performer". The EU's trigger reaches both, the invented persona and the digital replica of a real person alike. An invented AI presenter therefore owes a disclosure under each. What differs is what you have to say and where it has to appear, so draft the two disclosures separately rather than assuming one line clears both. Our guide to the EU AI Act ad disclosure duty works through what a compliant EU label has to look like.

What to actually do this week

• Decide, per creative, whether a human appears and whether that human was invented by a model. That single question drives everything below.
• For invented presenters, put a visible disclosure in the creative itself. It travels with the file, survives re-uploads, and is the only version that plausibly meets "conspicuously disclose in such advertisement".
• For Amazon uploads, write contains-synthetic-performer into dc:subject on the final exported file, and verify it after your last re-encode rather than before.
• Check the exact keyword against your own Seller Central notice. There is no public Amazon page to check it against.
• Keep the platform disclosures separate in your head. TikTok has its own AIGC label and a second commercial-content toggle, Meta labels detected AI itself, and none of that is a statutory disclosure you made.
• Log which ads used a synthetic performer and which did not. The New York duty is keyed to actual knowledge, so a record of what you knew and when is the artefact worth having.

None of this is legal advice, and the two most consequential questions here, what "conspicuously" requires and how a licensed likeness is treated, are open on the face of both documents. What is not open is the easy part: if your ad features a photorealistic person who does not exist, disclose it.

Sources

• New York State Senate, Bill S8420A (2025), amending General Business Law 396-b. Signed 11 December 2025, effective 9 June 2026.
• Office of the Governor of New York, "Governor Hochul Signs Legislation to Protect Consumers and Boost AI Transparency in the Film Industry" (11 December 2025)
• Cooley, "New York Enacts 'Synthetic Performer' Disclosure Law for Advertisements" (29 January 2026)
• Forbes, "Amazon Requires Sellers To Label AI-Generated People In Listing Images" (25 July 2026)
• eWeek, "Amazon Requires Sellers to Disclose AI-Generated People in Product Images"
• Amazon Sellers Attorney, "What the contains-synthetic-performer tag actually requires"
• EU AI Act, Article 50, and the European Commission's transparency obligations FAQ
• European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50, C(2026) 5054 final, 20 July 2026, paragraphs 113 and 114 (read as the published PDF)

All read directly in August 2026. Amazon's own notice was not readable: it is gated behind the Seller Central login and Amazon has published no equivalent public page.

Selling on Amazon with AI creative? The disclosure is only half the upload. AI video for Amazon ads: the 16:9 problem covers the encode Amazon actually accepts, which is where most rejections come from, and if you are choosing between an invented presenter and an audio-driven avatar, the best AI video model for lip sync splits the models by how they handle a real voice track. UGC Vids AI runs 12 models from one credit balance, from $49/month, free for 3 days.