Labelling AI
Lesson 21 of 22 in the course “AI Video, No Flops”. Free, with animations and a quiz.
A new rule in Europe
An “AI-made” label is no longer just a matter of taste. On 2 August 2026, Article 50 of the AI Act, the European Union’s artificial intelligence law, took effect. It requires those who publish deepfakes to state clearly that the image, audio or video was created or altered by AI. The law isn’t only for big studios: if Mira shows an ad to viewers in the EU, the rule applies to her too. In this lesson we’ll see what counts as a deepfake, which labels platforms have and how it all works technically. This isn’t legal advice: for a large campaign, check the details with a lawyer.
What a deepfake is under the law
Many people think a deepfake is just a fake celebrity. Under the AI Act the idea is much broader. It’s AI-generated or AI-altered image, audio or video resembling real people, objects, places or events that could seem authentic to a person. So it includes realistic invented people, like a photoreal Mira, existing buildings or products, and events that look true. An important detail: no intent to deceive is needed. It’s enough that viewers could take the video for reality. That’s why a realistic AI ad almost always needs a label.
Art, satire and advertising
For clearly artistic, satirical or fictional works the law is lighter but doesn’t drop the label. A note is still needed; it can just be gentle so it doesn’t spoil the work, for example mentioning AI in the end credits. It’s easy to slip here. An ad set in an invented world doesn’t automatically become art: if the main goal is to sell rather than to express something, it’s usually advertising, and the label is needed in the normal way. Breaching Article 50 can mean fines of up to €15 million or 3% of the company’s worldwide annual turnover, whichever is higher.
Labels on platforms
Even where the law is silent, platforms have their own rules. TikTok requires realistic AI content to be labelled: there’s an “AI-generated” toggle when posting, and an unlabelled realistic video can be removed. Since March 2024 YouTube has asked creators to flag “altered or synthetic” content in YouTube Studio if it looks realistic, and shows viewers a note. Meta puts an “AI info” label on Instagram and Facebook and requires disclosure of realistic AI-made video and audio. Interfaces change, so check the platform’s current help pages before posting.
How it works inside the file
Besides the visible label, AI video can carry an invisible one. The first is C2PA, or Content Credentials: an open standard that writes a “passport” into the file, saying who made it and with which tool. Since 2024 TikTok reads these passports and can apply an AI label by itself. The second is a watermark inside the pixels. For example, Google embeds SynthID in Veo videos. But there’s a catch: when a video is re-encoded in an editor or converter, the C2PA metadata can be lost. So don’t count on the platform detecting it all by itself: switch the AI toggle on manually when posting.
Rules by country
Mira’s audience isn’t only in Europe, so here’s a quick look at the main regions. EU: since 2 August 2026 deepfakes must be labelled, with fines up to €15 million or 3% of turnover. China: since 1 September 2025 AI-content labelling rules apply, requiring a visible label for viewers and a hidden label in the file’s metadata. USA: there’s no single federal labelling law, but platform rules apply, and the FTC punishes consumer deception, including fake reviews. Elsewhere rules change fast, so check local requirements before launching in a new country. The simplest strategy: always label, and none of these rules will touch you.
Hide it or label it
Sometimes it seems an AI label will scare viewers off and it’s better to keep quiet. Compare the outcomes. If you hide it: the platform may remove the video or restrict the account, in the EU there’s a risk of fines, and when the truth surfaces in the comments the brand loses trust, as we saw in the last lesson. If you label it: the video lives in peace, “were we fooled?” never comes up, and the AI style becomes your signature over time. There’s no reliable public data that the label itself cuts reach. But the damage from being exposed is plain in real examples. That’s why Mira always labels.
Check yourself
- Mira makes a photoreal ad with an invented girl and shows it in Germany in October 2026. Is an AI label needed?
Correct answer: Yes: a realistic invented person also falls under the deepfake rule. The law looks at whether viewers could take the video for reality, not at whether the person exists. - Mira made an ad in a fantasy world with flying sneakers. Can she call it “art” and skip the label?
Correct answer: No: the main goal is to sell, so it’s advertising and needs a label. The relief covers clear art and satire. A fictional world in an ad doesn’t change its commercial goal. - Mira edited the video and uploaded it to TikTok. No AI label appeared by itself. What should she do?
Correct answer: Turn on the “AI-generated” toggle manually. The editor may have stripped the C2PA metadata on re-encoding. The label is the creator’s responsibility. - Which is the most reliable strategy for a brand that works in several countries?
Correct answer: Always label AI videos and check the rules when entering a new country. Labelling everywhere covers the law, the platforms and the trust question at once.
← What lands and what annoys Other people’s faces, rights and the final project →