Do You Have to Label AI-Generated Ads? The 2026 Rules

New York started requiring disclosure on AI-generated ads on June 9. The EU joins on Sunday. Google shipped the controls to label AI-generated ads across its entire ad stack this month. Three jurisdictions, three different triggers, one ver

New York started requiring disclosure on AI-generated ads on June 9. The EU joins on Sunday. Google shipped the controls to label AI-generated ads across its entire ad stack this month.

Three jurisdictions, three different triggers, one very short runway.

Here is the part nobody leads with: most of your AI ad creative needs no label at all. The rules are narrower than the panic around them. The exceptions just happen to land on exactly the creative that performs best right now.

New York moved first and almost nobody noticed

S.8420-A/A.8887-B was signed on December 11, 2025 and took effect on June 9, 2026. It is the first state law in the US requiring advertisers to disclose AI-generated synthetic performers in advertising.

The trigger is specific. A visual or audiovisual ad, distributed to a New York audience, that features a synthetic performer: a digitally created asset intended to create the impression that it is giving an audiovisual performance, and not recognisable as any identifiable real performer.

If that is in your ad, you conspicuously disclose it.

Four details that matter more than the headline. It applies to any advertiser whose ads reach New York consumers, regardless of where you are based. The duty sits with whoever produces or creates the ad, not the platform that distributes it. It only bites where you have actual knowledge a synthetic performer is in there. And it is not limited to AI at all — a digital human built with conventional VFX counts.

Audio-only ads are carved out. So is AI used purely to translate a real performer into another language.

The statute does not define "conspicuous." That gets settled by enforcement practice and whatever the market converges on, which means the safe move today is a visible on-creative label rather than a line buried in your landing page footer.

The EU's August 2 deadline covers less ad creative than the panic suggests

Article 50 of the AI Act applies from August 2, 2026. It is not a high-risk provision, so the Digital Omnibus delay everyone reported does not save you. We wrote about why the AI Act delay does not cover the transparency rules when the confusion started.

Two of its four limbs touch advertisers.

Article 50(2) makes providers of generative AI systems mark their outputs in a machine-readable way. That is the model vendor's obligation, not yours. Systems already on the market before August 2 get until December 2, 2026 to comply.

Article 50(4) is the one with your name on it. Deployers who create deepfakes must disclose that the content is artificially generated or manipulated. The Act defines a deepfake in Article 3(60) as AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear authentic or truthful.

Read that definition twice, because it does a lot of work. The Commission's own guidance confirms that clearly fantastical or physically impossible content sits outside it. Dragons are fine. A generated photo of a shopfront that does not exist is a different conversation.

Google shipped the label controls this month

From July 2026, Google allows text or visual AI labels directly inside image and video ad creative that was generated or modified with AI. There is also an AI label setting across Google Ads, Display & Video 360, Campaign Manager 360, Merchant Center and Google Ads Editor.

The genuinely useful part is buried in the policy: those labels do not violate Google's own rules banning text overlays and watermarks. That prohibition was the practical blocker. You could not comply with a disclosure law and a platform creative policy at the same time. Now you can.

Google also says it may automatically apply labels to assets built with its own AI tools.

And then it says the quiet part out loud: using the AI label setting does not guarantee compliance with any specific regulation. That is not boilerplate. The platform is handing you the mechanism and keeping the legal judgment on your side of the table.

Google gave you the label. It did not give you the judgment about when to use it.

— Arthur, Bunny Honey Club

Most AI ad creative triggers nothing

This is where the honest answer is more useful than the cautious one.

No label needed for a generated product still on a clean background. None for generated graphics, patterns, or abstract backdrops. None for AI used to upscale, retouch, relight, or extend a real photo of your real product. None for AI-written headlines and body copy in a commercial ad, because Article 50(4)'s text limb covers material published to inform the public on matters of public interest, and your carousel ad for a sofa is not that.

Three things do trigger a disclosure.

A synthetic human in a visual ad reaching New York. The cleanest of the three rules. UGC-style AI spokespeople, the format half the market piled into this year, is the textbook case.

Photoreal imagery that would pass as an authentic record. A generated "photo" of your premises, your team, or a customer using your product. That is the Article 3(60) definition doing its job. The tell is whether a normal viewer would think they were looking at something that actually happened.

Any recreation of a real person's likeness or voice. Never a grey area anywhere, and never worth the argument.

We moved the disclosure call from review to generation

Our AI ad creative pipeline ships around 50 variants a week across the portfolio, at €4 to €8 all-in for a still. Roughly 60% of raw generations come out brand-usable, and about 95% after the QA loop, which is the number we track in the brand imagery workflow as well.

The change we made is small and boring. The disclosure decision now happens at generation, not at review.

Every variant gets flagged the moment it is created: synthetic human, photoreal scene, or neither. Same for video variants coming out of Kling, where a synthetic presenter is the whole point of the format. The flag travels with the asset into the ad account.

That is the actual operator lesson here, and it generalises past this deadline. Compliance that lives in a checklist gets skipped the week a launch runs late. Compliance that lives in the asset's own metadata survives contact with a deadline. If you would rather not build that layer yourself, running AI ad creative end to end is one of the things we do, and the disclosure call is part of the job rather than an extra.

The label is cheaper than the argument

The pushback is always the same: a disclosure will hurt the ad's performance.

Maybe. Then test it properly, as a creative variable with its own split, rather than treating it as settled. A four-word disclosure is one of the cheapest things you will ever run an experiment on.

Now weigh it against the other side. $5,000 for each New York violation after the first, a duty that attaches wherever your ads land rather than where your office is, and an EU regime arriving on Sunday with a definition broad enough that "we thought it was fine" is a weak position.

The creative most likely to need a label is the creative you are scaling. That is not a coincidence, and it is the reason to decide this now rather than in a support ticket later.