Advertising on Meta
AI Labels on Ads: What Meta and the EU AI Act Require (2026)
Making more of your ads with AI? Then labels are on your mind. What the EU AI Act requires since August 2, 2026, and when an AI label is needed.
Written by Genly's AI team, reviewed by Marlon Bonink
· 7 min read
More and more of your ads are made with AI: an AI presenter explaining your product, a product photo in a stylish setting, a voice-over. Then you read about the AI Act and suddenly notice "AI info" under ads on Instagram. Do you need to label too? And will it cost you reach?
Short answer: Since August 2, 2026, the EU AI Act requires you to clearly disclose that images, video or audio were made with AI when they look realistic and could pass as real, such as an AI presenter or a cloned voice. A visible line ("made with AI"), a spoken disclosure at the start or the platform's label is enough. Clearly illustrated or animated content does not need it. New since October 7, 2026: with Google's SynthID Detector, anyone can check files for an AI watermark from Google, OpenAI and others.
This article explains the rules in plain language and is not legal advice. If you are unsure about your situation, ask a lawyer.
When do you need to label?
The AI Act calls this "deep fakes": AI-generated or manipulated images, video or audio that resemble existing people, objects, places or events and could falsely appear authentic. A realistic person who does not exist counts too. Whoever publishes the content, which in an ad means the advertiser, must disclose it.
| Type of content | Label needed? |
|---|---|
| AI presenter or AI person who looks real | Yes |
| Cloned voice of an existing person | Yes, and only with consent |
| Realistic AI scene that could pass as real (say, a product in a fake living room) | Generally yes; when in doubt, label |
| Clearly illustrated, animated or fantastical video | No |
| Minor edits, such as removing a background or color correction | Generally not |
| Ad copy written with AI | Not for regular advertising |
What should the label look like?
The law sets no fixed format, only a result: clear and noticeable at the first moment of exposure. In practice:
- A visible line on screen, such as "Made with AI", in a corner of the video from the first second.
- A spoken disclosure at the start, for example the AI presenter saying it is AI.
- The platform's label, such as Meta's or TikTok's AI marking.
Do not hide it in hashtags or at the bottom of the caption. In June 2026 the European Commission published a final Code of Practice on AI-generated content plus official EU icons. The icons are voluntary: a simple line of text is enough.
What Meta and TikTok do
Meta shows "AI info" on ads that were created or significantly edited with generative AI. Since an update on June 1, 2026, that label sits in the "About this ad" menu behind the three dots. It applies to Meta's own AI tools, and to other AI tools when Meta detects technical signals in the file, such as C2PA or IPTC metadata.
TikTok asks creators to label realistic AI content and labels content itself: according to TikTok, more than 3 billion videos so far, using Content Credentials, creator labels and its own invisible watermark. Ads face a stricter rule: in TikTok Ads Manager, the AI disclaimer is mandatory for images, video or audio that are fully AI-generated or significantly modified by AI. Viewers then see a text label at the bottom of the ad.
Note: an automatic label is no guarantee. Metadata often gets lost in editing and export, so the platform sees nothing. You stay responsible.
What changed in October 2026
The rules have not changed since August 2, and the December 2, 2026 deadline for existing AI tools still stands. But AI content is becoming much easier to detect, even without a label.
| When | What happened |
|---|---|
| Since May 2026 | OpenAI adds an invisible SynthID watermark to images from ChatGPT, Codex and the API |
| October 7, 2026 | Google opens the SynthID Detector to everyone worldwide, in English for now |
| October 30, 2026 | Meta's generative AI rules for ads become part of its standard ad terms |
Google. With the SynthID Detector, you upload an image, video or audio file and see whether it carries a watermark from AI made by Google or partners such as OpenAI, NVIDIA and Kakao. The same check is built into Google Search, the Gemini app and Chrome.
OpenAI. Since May 2026, images from ChatGPT, Codex and the API carry a Google DeepMind SynthID watermark on top of C2PA metadata. The watermark lives in the image itself and can survive a screenshot, when metadata is often already gone.
Meta. The "AI info" label is unchanged since June. But on October 30, 2026, Meta's Ad Creative Generative AI Terms are folded into its standard Self-Serve Ad Terms. They say you may not present AI output as human-made or alter or remove watermarks or metadata that verify AI content.
What does this mean for you? Viewers cannot see an invisible watermark, so your AI Act duty still stands. But passing off an AI video as real gets caught faster and faster.
What the AI Act requires, in plain language
Article 50 of the AI Act covers transparency. Four points matter for advertisers:
| Who | What | From |
|---|---|---|
| Whoever publishes realistic AI content (the advertiser) | Clearly disclose it was made with AI | August 2, 2026 |
| Providers of AI tools | Mark output in a machine-readable way, such as metadata or an invisible watermark | For existing systems, by December 2, 2026 |
| Providers of chatbots | Tell people they are talking to AI | August 2, 2026 |
| Whoever publishes AI text on matters of public interest | Disclose it, unless a human reviews it editorially | August 2, 2026 |
Maximum fines are high: up to 15 million euros or 3 percent of global annual turnover. For small businesses the focus is simply on being clear: label honestly and you are on the right side.
Does an AI label cost reach or trust?
There are no reliable public figures showing that an AI label lowers reach. What is clear: viewers are getting better at spotting AI. An AI video that pretends to be real and gets caught costs more trust than an honest label. In our own ads, the AI presenter simply says it: I am AI.
More important than the label: never let an AI person pretend to be a real customer who used your product. That is misleading, label or not. For the bigger picture, read the future of marketing.
Frequently asked questions
Can I remove a watermark or AI marking?
Do not. You lose proof that you were transparent, tool and platform terms often forbid it (Meta's say so explicitly), and it does not change your duty to disclose realistic AI content.
Can people tell my ad was made with AI?
Increasingly, yes. Since October 7, 2026, anyone can check a file for a Google or OpenAI watermark with Google's SynthID Detector, and Meta and TikTok read C2PA metadata. But a missing watermark does not prove something is real: not every AI tool adds one.
Can I use an AI persona in an ad?
Yes, as long as you make clear it is AI and the persona does not claim a real customer experience. An AI presenter explaining your product is fine; an AI person saying "I have used this for years" is not.
Will I be fined if I do not label?
The AI Act has high maximum fines, and national enforcement is still being set up in several EU countries. The sensible move is to label: it costs you little and prevents trouble with platforms and customers.
Do I need to label AI-written text?
Not for regular ad copy. The text disclosure rule applies to AI text on matters of public interest without human editorial review, such as news articles.
The bottom line
Labeling takes seconds. In Genly's video editor you add an AI label with one click, using the official EU icon, in the bottom left for the whole video. Your part: turn the label on (or have it said in the video) for realistic AI people, and fill in the platform's AI disclosure on upload (mandatory for TikTok ads). Our AI transparency page lists the rules and a checklist for advertisers, and our AI UGC page shows videos with AI presenters.
Editor
Co-founder of Genly
This article was written by Genly's AI team and reviewed by Marlon before it went live. How we work with AI: AI transparency.
