AI-Generated Ads: What You Really Need to Disclose by 2026

Running AI-generated ads since August 2, 2026? You're probably wondering if you need to disclose anything, and what happens if you don't. The short answer: yes for some creatives, no for most.

What triggers the obligation is what your content makes the viewer believe, not the tool you used to create it. A fully AI-generated product video doesn't have the same requirements as a synthetic face pretending to be a real customer. It's the line that decides if you're affected, and most articles on the topic never draw it.

I run AI UGC production for e-commerce brands all year round. Here's the operator's version of the topic, platform by platform.

What Article 50 of the AI Act really says, and what it doesn't

Transparency obligations under theEuropean AI regulation, Article 50have been in effect since August 2, 2026. They weren't postponed by the Digital Omnibus, unlike the rules on high-risk systems.

The overlooked point is that the text divides the burden between two different parties.

Article 50(2) mandates technical marking,machine-readable, of generated content. This obligation falls on the model provider, not on you. Practically speaking, it's up to Google, OpenAI, or the tool's publisher to embed metadata in the file. You don't have to tag your exports.

Article 50(4) requires public disclosurewhen the content is a deepfake, meaning an image, sound, or video resembling a real person, object, or event. This obligation falls on the deployer—so you, if you're the one distributing. And it's explicit on one point: disclosure must be perceptible by a human at the first exposure, without any special technical tool. You can't rely solely on the machine marking from your provider.

The grace period you might have heard about, until December 2, 2026, only concerns the machine marking obligation of Article 50(2), and only for systems already on the market before August 2, 2026. A tool released after this date has no delay. And this grace period belongs to your provider, not to you.

Penalties fall under Article 99: up to 15 million euros or 3% of global turnover. For an SME or a startup, the calculation reverses, and the lower amount applies, not the higher. If you're a DTC brand with 3 million in turnover, the real cap isn't 15 million.

Last point nobody mentions: in France, the authorities overseeing the AI Act have yet to be formally designated. The planned structure places the DGCCRF as the operational contact, but the designation awaits the adoption of the adaptation text by Parliament. The obligation exists, but enforcement isn't fully armed yet. That doesn't mean you should ignore it—it means you can comply without panicking.

Platform by platform: what to check and where

A compliant ad on one platform isn't automatically compliant on the next. The scopes are completely different.

Meta Ads.Disclosure is mandatory for political, electoral, or societal ads containing photorealistic images, sounds, or videos generated or modified by AI. The disclosure control appears during ad setup. There's no dedicated box for a standard product ad, but Meta scans C2PA metadata with each submission and can label your ad on its own, which you can't remove.

Instagram, at the profile level.What triggers the label is a profile showcasing an AI-generated person. It's checked in the profile settings and is declarative. Without the label, posts are no longer recommended to non-followers in Reels and Explore, with a possible appeal via Account Status.

TikTok.The broadest scope of all: any realistic content in image or audio, synthetic face, cloned voice, generated backdrop, photorealistic packshot. It's declared with the AI-generated content toggle in the publication flow. Automatic detection via Content Credentials applies the label without you, and the content can be removed.

YouTube.Only deceptive realism: a real person saying what they didn't say, a real place, a real event modified. It's checked at upload, in the box for modified or synthetic content. YouTube can label the content itself when there's a risk of deception.

The takeaway from all this is the gap between TikTok and YouTube. TikTok asks you to declare a generated photorealistic packshot. YouTube doesn't care as long as no one believes a real person said something. If you're distributing the same creative on both, you declare on TikTok and not on YouTube.

And on Meta, the real issue is the automatic label, not the checkbox. Detection runs on the metadata embedded by your generation tool, so the label can appear without you declaring anything. That's the part you can't control, and that's why the question "does the label lower my performance" matters more than compliance.

The French case: the "Virtual Images" label that almost nobody knows about

If you're distributing in France, this text affects you more directly than the AI Act.

TheJune 9, 2023 law on commercial influencerequires the mention"Virtual Images"on content containing AI-generated images "aimed at representing a face or silhouette." The mention must be clear, readable, and identifiable on the image or video, in all formats, throughout the entire viewing. Not at the end of the video, not in the description. During the whole duration.

Failure to comply is punishable by one year of imprisonment and a 4,500 euro fine. It's much less than the 300,000 euros circulating: that amount sanctions the absence of "Advertisement" or "Commercial Collaboration" mentions, not the absence of "Virtual Images." Both can accumulate but don't overlap.

The scope of application is decisive. This obligation targets those engaged in commercial influence, defined as leveraging one's notoriety with an audience, for a fee, to promote goods or services. A brand distributing its own ads on its own account isn't leveraging influencer notoriety. The day you pay a creator or run a persona account, you fall within the scope. And the law explicitly targets legal entities.

On the ethics side, the ARPP published on August 3, 2026 apractical guide on labeling AI-generated or manipulated advertising content. It offers a decision tree built on three criteria: the level of AI intervention, the degree of realism, and the impact on consumer perception of authenticity. Example formulations given are "Image generated by AI" and "Voice generated by AI."

Look closely at these criteria: none measure the percentage of AI in your creative; they all measure what the viewer will believe. The Advertising Ethics Council wrote in black and white in July 2026: advertising transparency is about the product's characteristics and the truthfulness of claims, not the narrative construction methods.

This comes from a French authority, not from me.

Instagram's "AI-generated profile" label, and why it probably doesn't concern you

On August 31, 2026, Instagram renamed its "AI creator" label to "AI-generated profile" and announced it would limit the reach of profiles showcasing an AI-generated person that don't apply it.

It's declarative and done at the profile level, in settings, not creative by creative. The penalty is exclusion from recommendations: posts are no longer pushed to non-followers in Reels and Explore. No account deletion. A correctly labeled profile retains its reach.

And most importantly: using AI tools for production, editing, or assembling doesn't trigger anything. The label targets profiles showcasing an AI-generated person presented as real. If you're distributing UGC-type videos from your brand's account, you're not in this case. If you're running a synthetic persona account with a name and a life, you are.

In case of a wrongly applied label, recourse is through Account Status, and a successful appeal restores recommendation eligibility.

Where's the real risk: faces and voices

What increases risk is the presence of a credible human making a claim. The AI percentage in your creative doesn't matter. Here's the scale, from the safest creative to the most exposed.

Product b-roll, packshot, generated texture, decor.Low risk under both the AI Act and French law. Declare on TikTok if photorealistic.

Synthetic voice-over on product images.Low risk, unless the voice imitates an identifiable person. ARPP suggests the mention "Voice generated by AI."

Generic avatar giving a testimonial on camera.This is where it shifts. A synthetic face giving a customer review is a face under French law.

Lookalike or voice of a real person.Red zone. Deepfake under Article 50(4), plus image rights, plus deceptive commercial practice.

No one believes a bottle filmed in macro has an opinion on your product. An avatar saying "I've been using it for three months," yes. That's where declaration becomes a real obligation, and omission becomes deceptive commercial practice, which is a completely different sanction regime than the AI Act.

In practice, this means compliance is settled at the briefing stage, not at the ad submission. You decide upfront if your creative needs a synthetic face giving a testimonial, or if a product b-roll with a voice-over does the same job. In the latter case, you have almost nothing to declare.

What you don't have to declare

Using AI in your production process without embodying a fictional human doesn't trigger anything on Instagram or YouTube. Script written with a model, assisted editing, upscaling, color grading, thumbnail generation, translation: nothing to declare.

Content that's clearly artistic, satirical, or fictional falls under a lighter regime under Article 50(4). The obligation then reduces to not hindering the work.

Content that's clearly unrealistic or stylized, that no one can mistake for real, is outside TikTok's and YouTube's scope.

And a text reviewed, validated, and editorially endorsed by a human is outside the obligation of Article 50(4) for general interest publications.

Translated for an e-commerce brand: if your creatives are product videos with voice-over, you're pretty safe everywhere except on TikTok's toggle. If your creatives rely on faces giving testimonials, you have a real issue to address at the briefing stage.

Why labeling works in your favor

The spontaneous reading is that the label is a penalty. I don't believe it, and not out of optimism.

We produced nine UGC-type videos for Cheerz, across three markets—France, Spain, and Italy—delivered in five business days. All content was labeled as AI-generated.

Zero negative comments on all nine. Not a single complaint about it being AI, even though it was displayed.

And in Italy, where we compared with traditional creators on the same market and Meta account, AI creatives took the lead: 0.49% outbound CTR versus 0.27%, and €2.19 cost per install versus €2.92. That's 81% more CTR and 25% less CPI, label displayed.

Nine creatives don't make a sample, and I don't draw a general law from it. Spain, on the same wave, came out at €8.62 CPI, so not everything worked everywhere. But on the label question, the answer is clear: transparency cost nothing. What triggers rejection is a failed creative trying to pass as real.

A label makes the content's origin readable. Today, AI UGC is weighed down by suspicion, that viewer wondering if they're being fed nonsense. When the origin is displayed, the suspicion falls away, leaving only the creative's quality to judge. It's the normal advertising situation.

The other, more immediate effect: labeling dries up the gray area. Those selling fake testimonials as real lose their edge.

If you want to understand what the format truly allows before worrying about what to declare, check outwhat AI UGC isand theperformance comparison with creator UGC. And if budget is the issue, thereal cost of a UGC videoand thecomplete price listanswer faster than I do.

FAQ

Do I need to write "AI-generated" on an ad?

Not always. The obligation arises when the content can be mistaken for real, particularly when featuring a face or voice. For an AI-produced image or video representing a face or silhouette in commercial influence content, French law requires the mention "Virtual Images," visible throughout the viewing. ARPP recommends the formulations "Image generated by AI" and "Voice generated by AI."

What are the risks of not declaring?

Three regimes overlap. The AI Act provides for up to 15 million euros or 3% of global turnover, with the lower amount applied to SMEs. French law on commercial influence foresees one year of imprisonment and 4,500 euros for missing the "Virtual Images" mention. Platforms impose their own sanctions, from automatic labeling to content removal.

Does the AI label reduce performance?

On Instagram, a properly labeled profile retains its full reach. It's the absence of a label that costs recommendations to non-followers. On Meta Ads, the automatic label can appear without your request, so the creative's quality matters more than the label's presence. In a campaign of nine contents all labeled AI, we saw no negative comments, and in Italy, these creatives beat traditional creators by 81% on CTR.

Does an AI-generated voice-over need to be declared?

Generally no, as long as it doesn't imitate an identifiable person. A cloned voice reproducing a real person is a completely different issue: it's a deepfake under Article 50(4) and also involves image rights. On TikTok, voice cloning is explicitly part of the scope to declare.

Who's responsible, the agency or the advertiser?

The technical marking of the file falls to the model provider. Public disclosure falls to the deployer and distributor, so the advertiser in most cases. A contract with an agency or studio doesn't transfer this obligation; it organizes who does the work best.

What to remember

The right question to ask yourself is: can my content make someone believe a real person said or did something? Not: did I use AI. If the answer is no, you have a toggle to check on TikTok and nothing else. If the answer is yes, you declare, and you decide this at the briefing stage rather than at the ad submission.

Inventory your active creatives according to the faces and voices grid. You'll likely find that most of your production triggers no obligation, and the real issue lies with three or four creatives.

If you want to produce declarable UGC-type videos without spending weeks on it,test a first batch at Hoox studio. Delivered in 5 days.