Total:
$0.00
Eyewear has a photography problem that no other category has quite as badly. You cannot sell a frame on a white background. The customer is not buying an object, they are buying how the object sits on a face — and not a face, their face, which means you need many faces: narrow and wide, warm and cool, young and old, every bridge height you stock.
Shooting that properly costs a fortune. So over the last two years a large part of this industry quietly stopped shooting it. Generative models produce forty faces in an afternoon for the price of a coffee, in every demographic your catalogue serves, wearing frames composited perfectly on the nose.
Three separate things happened this year to that practice, and none of them was written about for this vertical. A fourth happened on 30 September, and it is the one that makes the other three matter.
What actually changed
9 June 2026 — New York's synthetic performer law came into effect. Signed on 11 December 2025 as S.8420-A / A.8887-B and added to section 396-b of the General Business Law, it is the first law of its kind in the United States. It requires a person who produces an advertisement containing a synthetic performer to “conspicuously disclose in such advertisement that a synthetic performer is in such advertisement, where such person has actual knowledge” of it. It took effect 180 days after signature.
The statutory definition is worth reading slowly, because its edges matter more than its centre. A synthetic performer is:
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.
Penalties are $1,000 for a first violation and $5,000 for each subsequent one. Exempted: audio-only advertising, advertisements for expressive works such as films, television, documentaries and video games where the use is consistent with the work, and use of AI solely to translate a human performer's speech. Media platforms that merely publish an ad are protected unless they have written notice.
22 July 2026 — Amazon started asking for the tag. In Amazon's own words: “You must use an IPTC-compatible metadata editor to add the keyword ‘contains-synthetic-performer’ to the dc:subject (XMP) field of your image or video files before adding them to your listings and A+ content.” It applies across Amazon's worldwide stores, and only to photorealistic AI-generated people — not to real people who have been retouched or AI-edited, not to cartoons or stylised characters, not to images with no people in them.
Two details in that notice deserve more attention than the coverage gave them. First, Amazon says that where the metadata is present it will “add an indicator to inform customers that the image includes AI-generated people”. This is not a silent compliance field. Tagging your imagery puts a label in front of the shopper, which makes it a merchandising decision as well as a legal one — and makes the choice not to tag look considerably worse if it is ever examined. Second, Amazon published no compliance deadline and no enforcement mechanism alongside the requirement. Treat that as unfinished rather than as slack.
2 August 2026 — California's AI Transparency Act became operative. This one is aimed at the AI companies rather than at you. The original date under SB 942 was 1 January 2026; AB 853 moved the whole chapter to 2 August 2026 and added phases behind it. Covered providers of generative-AI systems must maintain a free public tool that tells anyone whether a file was created or altered by that provider's system, and must apply latent disclosures — durable provenance metadata identifying the system, its version and the creation date. Under AB 853, large online platforms pick up detection and display duties on 1 January 2027, and capture device manufacturers follow on 1 January 2028.
30 September 2026 — SB 1000 was signed, and it rewrote that chapter in the middle of its first year. The amendment this article previously described as pending is now law. SB 1000 (Senator Josh Becker) amends sections 22757.1 to 22757.5 of the Business and Professions Code, and it is an urgency statute: section 8 of the bill provides that it “is an urgency statute necessary for the immediate preservation of the public peace, health, or safety” and “shall go into immediate effect”. There was no run-up and there is no transitional period.
Three of its changes matter to anyone publishing synthetic imagery.
The threshold is gone. A covered provider is now, in the statute's words, “a person that creates, codes, or otherwise produces a generative artificial intelligence system that is publicly accessible within the geographic boundaries of the state”. The one-million-user test that used to sit in that definition has been deleted. The practical effect is that the small and mid-sized image generators a marketing team actually reaches for — the ones that were comfortably under the old threshold — are now inside the chapter alongside the large ones.
The disclosure your customer could see has been removed, and the one only a machine can read has been hardened. SB 1000 deletes the requirement that a provider offer the user the option of a manifest disclosure — the visible, on-the-face-of-it label. What it strengthens is the latent disclosure, which must now be “permanent or extraordinarily difficult to remove or tamper with”.
And cropping no longer defeats it. The free tool is now called a disclosure verification tool, it must be available “at no cost to the user”, and it must output “any system provenance data that is detected in the content”. Critically, the chapter now carries a definition of minor modifications — and the listed examples include file resizing, scaling, cropping and file format conversions. Those are not incidental examples. They are the exact four things a merchant's image pipeline does to a generated asset on its way to a product page.
Why the California limb is the one that changes the risk
New York's duty is yours. Amazon's tag is yours. California's is not — and that is precisely why it is the one to understand.
Until this year, “is that model real?” was an unanswerable question in practice. A regulator or a plaintiff's firm looking at your hero image had a suspicion and no way to prove it, and proving it was the entire barrier to any claim.
Now the generators in scope are required to sign their output durably, and the same statute requires the provider to run a public tool that will tell anyone who uploads a file what made it and when. Anyone means a competitor, a consumer group, a state AG's office, or a firm that has decided AI models in advertising is its next docket.
Read that as an evidence rule, not a compliance task. Nothing in the chapter tells a merchant to do anything. You are not a covered provider — you do not create, code or produce a generative AI system, you buy pictures from one. You are not a “large online platform” either; that term is defined as a public-facing social media platform, file-sharing platform, mass messaging platform or stand-alone search engine, and a Shopify storefront is none of those. The $5,000-per-violation penalty in section 22757.4 is collected in a civil action brought by “the Attorney General, a city attorney, or a county counsel”, and it runs against providers. There is no private right of action.
What California has done is make a third party manufacture durable, free, publicly checkable proof about the images on your product pages. The obligation is somebody else's; the evidence is about you.
Two honest qualifications, because this is the strongest claim in this article and it should not be oversold. First, the latent-disclosure duty has been expressed as applying where it is technically feasible and reasonable, which is a real limit rather than a formality; we have not re-verified against the amended text whether that qualifier survives SB 1000 in the same form, and you should not assume either way. Second, provenance metadata is still only as durable as the pipeline it passes through, and a screenshot is not a file format conversion. What has changed is the floor: the four ordinary operations that used to strip provenance by accident — resize, scale, crop, convert — are now the operations the verification tool is defined to see through. Concealment used to be free and is now something you have to do on purpose, which is exactly the kind of change that turns an unprovable suspicion into a viable complaint.
One more thing we looked for and did not find: a provision making it unlawful for anyone else to strip a latent disclosure. The chapter requires providers to make the mark hard to remove; we found nothing in it that forbids a third party from removing one. So the merchant's risk here is not a CATA penalty. It is what the stripped file, or the intact one, says about you under New York's statute and the FTC's rules — and it is worse to be caught having removed a mark than never to have had one.
So the position is now this: New York makes concealment an offence, California has made it substantially more detectable and just closed the easiest accidental escape route, and Amazon has already built the field where the answer goes — and will show the answer to your customer. A practice that was invisible in January is broadly machine-readable now. That is the shift, and it happened without a single eyewear trade outlet mentioning it.
Is a product listing an “advertisement”?
This is the honest ambiguity in the New York law, and it deserves stating plainly rather than being papered over.
The statute speaks of advertisements. It does not resolve whether a product detail page — the image carousel on a Shopify storefront — is an advertisement or merely a product description. A paid social creative clearly is. A Meta ad running to New York consumers clearly is. The seventh lifestyle shot on your product page is a harder question, and no court has answered it.
What is not ambiguous is where the largest marketplace in the country landed. Amazon read the law, pointed at it by name, and decided that listing images, videos and A+ Content are within its blast radius. When the biggest retailer in your channel treats product imagery as advertising, arguing the other way in front of a regulator is not a comfortable place to be.
There is also a plainer route to the same conclusion that does not depend on the New York statute at all. The FTC's Rule on the Use of Consumer Reviews and Testimonials, in force since October 2024, makes it unlawful to materially misrepresent “that the reviewer or testimonialist exists” (16 CFR § 465.2(a)(1) and (b)(1)). If your AI-generated woman in the tortoiseshell round frames is captioned as a customer, or sits inside a review carousel, or is presented as anyone's endorsement, that is a federal problem and it has nothing to do with which state your buyer lives in.
And if the synthetic person moves, a different California statute is waiting. SB 1050 regulates synthetic performers in video advertising on terms that do not match New York's rule for stills — two statutes, two media, neither one covering the other.
Three kinds of face, three sets of rules
The useful way to hold all of this is to stop thinking about “AI” and start thinking about whose face is in the frame. An eyewear funnel has three, and 2026 attached a different regime to each.
The synthetic face. A person who does not exist, generated to model your product. Disclose in New York. Tag it on Amazon, and expect the customer to see a badge. Assume the file itself may testify against you, and that resizing or cropping it will not change that. And never, ever let it appear as a customer or a testimonial.
The real face. A hired model or an influencer. The old rules still apply — a licence, a release, a scope, an expiry — and a new one has arrived beside them: if you generate new poses or new frames onto a real, recognisable person, you have left synthetic performer territory (they are recognisable as an identifiable natural performer) and entered digital replica and right-of-publicity territory, which is a considerably more expensive place to be wrong. California signed SB 1111 on digital replicas on the same day as SB 1000; we have not yet read its operative text and are not describing its effect here.
The customer's own face. Your virtual try-on widget. This is the most expensive of the three, because the exposure under state biometric privacy statutes is per-customer statutory damages rather than a $1,000 fine, and a Seventh Circuit decision this summer narrowed the health care exemption that a lot of eyewear try-on had been quietly relying on. We cover what that decision means for eyewear stores separately — including the line the courts have drawn, which turns out to be the prescription. California also took one tracking claim away from private plaintiffs on 30 September, and left standing the two that fit a camera best.
Every step in an eyewear purchase that involves a human face is now a compliance object. That is not a rhetorical flourish; it is a list of three things a merchant now has to manage that they did not have to manage eighteen months ago.
California, incidentally, is having a year on this front generally: the same state is in court over what you may print on your packaging while its provenance rules come into force over your imagery.
What to do this week
Find out which images are synthetic. Most merchants genuinely do not know, because the answer sits with an agency, a freelancer, or a marketplace-listing tool that generated lifestyle shots automatically. Ask, in writing, per asset. This is the whole task and almost nobody has done it.
Then check the files you already published. This step is new, and SB 1000 is what makes it possible. The verification tools are free and public, by statute. Run your own hero images through the tool belonging to whichever generator your agency used, and find out what a stranger would find out. It is better to learn that from your own laptop than from a letter.
Add the Amazon tag if you sell there. It is an XMP metadata keyword written before upload with any IPTC-compatible editor. Half an hour of work, once, plus a step in your asset pipeline.
Put a caption on the synthetic ones. “Images depict AI-generated models” near the imagery. Nobody has told us what “conspicuous” means in this statute and it may be years before anyone does, so the safe answer is legible, adjacent to the image, and not hidden in a footer link.
Change the contract, not just the file. Your agreement with the agency or freelancer should say who is responsible for disclosure and who eats the fine. Do this at renewal, not after a letter arrives.
Purge synthetic people from anything that reads as a testimonial. Reviews, “loved by”, UGC walls, before-and-after. This is the one with federal teeth and it is a delete-key fix.
Do not overcorrect. A rendered frame on a white background contains no person and is outside all of this. A retouched photo of a real model is outside the New York definition. Product photography does not need to become a legal project; three questions, a metadata field and one pass through a free verification tool will cover it.
The part of the funnel with no face in it
Step back and look at where this year's rules keep landing.
The face in your ad now needs a disclosure. The face in your lookbook needs a licence. The face in your virtual try-on collects biometric identifiers and needs consent, a retention schedule and a much more careful legal read than it was getting. Every one of them attaches to the image layer of an eyewear store — the part that gets a customer interested.
The step that converts that interest into revenue involves no face at all. A prescription capture — sphere, cylinder, axis, add, PD — plus a lens configuration. No synthetic person, no model release, no biometric identifier, no metadata field, nothing to disclose. It is the least regulated surface in an eyewear store and it is also the step that determines the size of the order, because that is where single vision becomes progressive, where an index gets upgraded, and where a coating gets added.
There is a version of this business that spends its next quarter re-auditing every image it has ever published. There is another version that spends it making the unregulated, high-margin, conversion-critical step of the funnel work properly, and handles the images with a caption, a metadata tag and an afternoon with a verification tool. The second one is both cheaper and better.
Frequently asked questions
Does the New York law apply to me if I'm not in New York?
It reaches advertising directed at New York consumers. For an online store shipping nationally, assume it reaches you. This is the ordinary pattern for state consumer-protection statutes and it is why one state's rule effectively becomes the national floor.
Does it apply to AI-edited photos of real models?
On the face of the definition, no — a synthetic performer is one who is not recognizable as any identifiable natural performer. But if the person is recognisable, you need their permission for what you generated, which is a different and often larger problem.
What counts as a conspicuous disclosure?
The statute does not say, and no guidance has defined it. Until it does, treat it the way you would treat a material advertising disclosure generally: legible, in the same visual field as the image, not behind a link or buried in a footer.
Is there an Amazon deadline I have missed?
Amazon announced the requirement on 22 July 2026 in the present tense and published no separate compliance deadline and no enforcement mechanism with it. That is not the same as it being optional, and marketplace policies of this kind usually acquire teeth later rather than never.
Do I need to tag images on my own Shopify store, like Amazon requires?
There is no equivalent metadata requirement outside marketplaces that impose one. But the underlying disclosure obligation is about what the consumer can see, not what is in the file — so a visible caption matters more for your own storefront than the metadata does.
Does California's AI Transparency Act impose any duty on my store?
We have not found one. The chapter's duties fall on covered providers — persons who create, code or otherwise produce a generative AI system — and on large online platforms as defined, which does not describe a Shopify storefront. The $5,000 civil penalty is brought by the Attorney General, a city attorney or a county counsel, and there is no private right of action. Treat the statute as changing what can be proved about your images, not as adding a box for you to tick. Take advice on your own position rather than relying on that summary.
Will the California rules catch every AI image I have used?
No, and it is worth being precise. SB 1000 was signed on 30 September 2026 as an urgency statute and took effect immediately, and it removed the one-million-user threshold from the definition of a covered provider — so far more generators are in scope than were in August. It also deleted the optional visible manifest disclosure and required the latent one to be permanent or extraordinarily difficult to remove, with the verification tool defined to see through resizing, scaling, cropping and format conversion. None of that is retroactive to files generated before a provider complied, and a screenshot is still a screenshot. The realistic effect is a much higher chance of detection going forward, not a certainty about your archive.
What about the EU?
Different regime, same direction of travel. The AI Act's Article 50 transparency obligations began applying in August 2026 and cover disclosure of certain AI-generated content; whether ordinary synthetic advertising imagery falls inside them is a live question, and the marking duty there sits with the provider too.
Is a rendered image of the frame itself affected?
No. All of this is about depictions of people. A CGI or AI render of the product with nobody in it is outside every rule discussed here.
This article is general information about regulatory developments, not legal advice. These rules apply differently depending on your channels, your imagery and where your customers are; take advice on your own position before acting.
VisioncarePro adds structured prescription capture and lens selection to any Shopify product page. It is the step of an eyewear purchase that has no face in it, no disclosure attached to it, and the largest effect on order value. Free to install.
Install VisioncarePro free on the Shopify App Store →
Sources: NY S.8420-A / A.8887-B (2025), amending New York General Business Law § 396-b — bill text read directly, including the definition of “synthetic performer”, the disclosure duty, the $1,000/$5,000 penalties, the exemptions and the 180-day effective date · Amazon seller notice of 22 July 2026 on tagging AI-generated people, as reported with Amazon's wording · California AI Transparency Act, Business and Professions Code §§ 22757.1–22757.5, as enacted by SB 942, amended by AB 853 (Stats. 2025, Ch. 674) operative 2 August 2026, and as amended by SB 1000 (Becker), signed 30 September 2026 — enactment taken from the Governor of California's announcement of 30 September 2026; operative text of the covered provider definition, the latent disclosure standard, the disclosure verification tool, the definition of minor modifications, the deletion of the manifest disclosure, the § 22757.4 penalty and enforcement provision and the § 8 urgency clause read via LegiScan · California SB 1111 (Ashby), digital replicas, signed 30 September 2026 — enactment only; operative text not read · FTC Rule on the Use of Consumer Reviews and Testimonials, 16 CFR Part 465, § 465.2 · Morrison Foerster commentary on the SB 1000 amendments, 1 October 2026 · Cooley, Crowell & Moring, Kelley Drye, Honigman and Morgan Lewis commentary on the New York and California statutes.
