Faut-il signaler une publicité faite avec l'IA? Les règles en octobre 2026

Sometimes. It depends on two things: what the AI made, and where the ad runs. A generated person who looks real now triggers a legal duty in New York ads and can trigger automatic labels on Meta and YouTube. A generated shot of your own product carries lighter duties, though not none, and the old rule still applies: an ad must be truthful and not misleading. Between those two cases sits a patchwork of laws and platform policies, and most brands are guessing. This is the map we work from, checked on October 3, 2026, at the source wherever we could reach it, and through a named law firm where we could not.

This is general information, not legal advice. Rules change, and your counsel has the final word. We are a film studio in Ahmedabad that has made ads for clients in 50+ countries since 2017, and we read these rules for a practical reason: our films have to run under them.

First, sort out what the AI made

• A synthetic person: a generated human who looks real but is nobody in particular. Most of the new rules aim here.
• A real person, recreated or altered: someone's face or voice doing or saying something they did not. Consent comes first, and a label sits on top of consent, never in place of it.
• No realistic person at all: a product, a place, an invented world. Duties are lighter but not zero: Europe's deepfake rule also covers existing places, objects and events shown in a way that could pass as authentic, and YouTube's automatic label looks for significant photorealistic AI.

New York: synthetic performers, since June 9, 2026

New York's synthetic performer law, an amendment to section 396-b of the General Business Law signed on December 11, 2025, took effect 180 days later. Anyone in business who produces or creates an ad for a commercial purpose, and knows it contains a synthetic performer, must disclose that conspicuously in the ad. A synthetic performer is a digitally created asset meant to look like a human performance by someone who is not recognisable as any real, identifiable person. Civil penalties are $1,000 for a first violation and $5,000 for each one after.

The exemptions are narrow: ads for films, shows and games that use the synthetic performer the same way the work does, audio-only ads, and ads where AI only translates a human performer's language. Publishers and carriers generally have an exception, but it can end after written notice giving them actual knowledge and the law's removal period. The law firm Crowell & Moring reads it as reaching any company whose ads reach New York consumers, wherever it is based. Its wording, produces or creates, can reach the studio that makes an ad as well as the brand that runs it, so where an ad will run is a question for the brief, not for launch day.

European Union: deepfakes, since August 2, 2026

Article 50 of the EU AI Act has applied since August 2, 2026. Providers of AI systems that generate audio, images, video or text must mark the output in a machine-readable way. And whoever uses a system to make a deepfake must disclose that the content was artificially generated or manipulated. The Act defines a deepfake broadly: AI-generated or manipulated image, audio or video that resembles existing people, objects, places, entities or events and would falsely appear authentic.

For evidently artistic, creative, satirical or fictional work, the duty shrinks to a disclosure that does not spoil the work. Whether a stylised ad qualifies is a question for counsel. The European Commission puts fines at up to 15 million euros or 3 percent of worldwide annual turnover. Its current guidance describes a machine-readable marking grace period until December 2026 for generative tools already on the market before August 2; confirm the applicable deadline with counsel rather than assuming a date for every tool. And, as the law firm Baker Botts notes, the rules reach companies outside the EU when the output is used inside it.

India: labels on realistic synthetic media, since February 20, 2026

India's amended IT Rules came into force on February 20, 2026, according to MeitY's official FAQ. The duties sit mainly on intermediaries, including platforms and services that offer AI tools: realistic synthetic audio, images and video must be clearly and prominently labelled, with permanent metadata or an identifier where technically feasible. Large social platforms must ask users to declare synthetic content and take the verification steps set out in the Rules. Routine, good-faith editing such as colour correction is excluded. This summary follows MeitY's consolidated Rules and FAQ, not the earlier consultation draft. Our studio is in India, so this one is home ground.

California: the label moves into the file

California's AI Transparency Act became operative on August 2, 2026, after AB 853 moved the date from January 1. Its core duties fall on covered providers of publicly accessible generative AI systems with more than one million monthly visitors or users, not on an advertiser merely using a tool: a free detection tool and hidden disclosures in generated image, video or audio content, to the extent technically feasible. AB 853 separately schedules large-platform provenance duties for January 1, 2027 and capture-device duties for January 1, 2028. Those later duties are not already in force. Preserve provenance data in your workflow; a hidden marker is not a substitute for a visible disclosure where one is required.

Google Ads and YouTube

Since July 2026, Google's ad products, from Google Ads and Display & Video 360 to Merchant Center, have been rolling out an AI label setting that adds a text or visual label to image and video ads. Google introduced it because rules in the EU, India and New York require disclosures on some AI ads, and it says plainly that using the setting does not guarantee compliance. Election ads have their own rule, in the US and wherever else Google verifies election advertisers: if an ad inauthentically shows real or realistic-looking people or events, the advertiser must tick the altered or synthetic content box, and add a prominent disclosure of their own in formats where Google does not generate one.

YouTube announced updated labels on May 27, 2026. If a creator does not disclose and YouTube detects significant photorealistic AI, its announced approach is to add the label itself: directly below the player on long videos, as an overlay on Shorts. YouTube says a label alone does not change how a video is recommended or whether it can earn money. For videos whose C2PA metadata shows they were fully generated, the announced label is permanent. Check the current upload and ad requirements rather than assuming every account has the same rollout.

Meta

Meta applies an AI info label to ads made or materially edited with its own generative tools. When the result includes a photorealistic AI-generated person, the label sits next to Sponsored rather than behind the three-dot menu. Meta has also said it will automatically label ads made or edited with third-party AI tools, with an AI info label inside About this ad, detected through industry-standard signals such as C2PA. Minor changes such as resizing or colour correction do not trigger the label.

TikTok: check it at the source

TikTok publishes its own rules for AI-generated content, in its Community Guidelines and its ad policies. TikTok is blocked in India, where our studio is, so we could not read the current wording at the source, and we would rather leave it out than guess. Read both directly before you run AI content there.

The IAB's voluntary standard

In January 2026 the IAB published a voluntary disclosure framework for AI in advertising. Its test is materiality: disclose when AI materially affects authenticity, identity or representation in a way that could mislead. It has two layers: a label people can see, and machine-readable metadata built on C2PA, the open standard for recording where content came from and how it was edited. Its examples include synthetic avatars that simulate human interaction, digital twins of real people, and generated scenes of events that never happened. Routine editing needs no label.

Does a label hurt the ad?

The evidence is thin, but it leans one way. In IAB research published in January 2026, 73 percent of Gen Z and Millennial consumers said knowing an ad was made with AI would increase or not change their likelihood to buy. The same study found that 82 percent of ad executives believe these consumers feel positive about AI ads, while 45 percent of the consumers say they do. It surveyed 505 US consumers aged 16 to 43 and 104 ad executives: a signal, not a law of nature.

Our reading, and it is an opinion: people object to being fooled, not to being told. An ad that hides a generated face risks being caught. An ad that says so in one quiet line has nothing to be caught at.

Decide it at the brief: a checklist

• List every market and platform before anyone is cast. New York, the EU and India each change the answer.
• Name the person on screen: a real person with written consent, a synthetic person labelled where required, or nobody.
• Ask whether anything could pass for a real place, product, person or event. That is the EU deepfake question.
• If it is an election or political ad, stop and bring in counsel.
• Write the disclosure into the script, with its wording and its seconds on screen, so it is approved with everything else.
• Keep a record of which shots were generated, with which tool, and when. It answers the first question anyone will ask.
• Check every edition. A vertical cut or a language version is a separate ad and needs its own label where the rules apply.

How a label goes into one of our films

Generated footage garbles text, so every word on screen in our films is set by a person in the edit, never generated inside the footage. A disclosure line is handled like a price card or a legal super: set, timed and checked with the cut. A New York edition can carry it and a version for another market can carry different wording, without regenerating a frame. For our script-led ads, you approve the script and the sound before any footage is made, so that first approval is the natural place to settle the label. Editions multiply quickly: across our archive, 454 delivered files are vertical or short-form editions, and 33 projects shipped in more than one language.

Our safeguards for election work are at filmito.io/en/political-ad-policy.

Planning an ad with a generated person in it? Tell us where it will run when you send the brief through filmito.io/en/see-your-ad, and we will write the label into the script from the start. Prices are the same for every client, at filmito.io/en/pricing.