EU Deepfake Labelling: 9 Things Production Companies Need to Know
Since 2 August 2026, anyone using AI to make realistic content that reaches EU audiences has to label it. The obligation sits with whoever used the AI system — which, on most jobs, is the production company.
Here are the nine things that actually matter when you're making work.
1. "Deepfake" means far more than synthetic humans
The legal definition covers AI-generated or manipulated content depicting people, objects, places or events that could appear authentic. In production terms: generated product shots, generated locations, synthetic voiceover and AI influencers are all in scope, not just face swaps.
2. The test is perception, not intention
You can have no interest in deceiving anyone and still need a label. What matters is whether a reasonable viewer would take the content as truthful. A generated product shot that looks exactly like the real product still counts, precisely because it looks real.
3. UK and US companies are caught too
The AI Act has extraterritorial reach. If your content is disseminated in the EU — and anything on social, on the open web, or in non-geoblocked media effectively is — you're in scope regardless of where you're incorporated. A London shop making an AI-augmented spot that runs on Meta in Germany is squarely covered.
4. Obviously unreal content doesn't need labelling
Fantasy, surrealism, talking animals, anything that can't be mistaken for reality. The further your work sits from photoreal, the safer it is. Animation and stylised treatments sit comfortably outside the definition.
5. Traditional-equivalent editing is exempt
Colour, grade, clean-up, retouching, relighting, lip-sync for language versioning — the things you'd have done in Photoshop or Flame anyway. Using an AI tool to do them doesn't convert them into deepfakes. The same applies to technical processing on a genuine audio recording.
6. Background elements generally sit outside the rules
If it isn't the focus of the frame, it typically doesn't cross the threshold. The grey zone is AI-assisted CGI — where you're doing something you'd previously have done in comp, there's a defensible argument, but it needs to be a decision you've made deliberately rather than a gap you've ignored.
7. Placement is the part people get wrong
Three requirements: at first interaction, clearly visible, not buried. For film, that means the head of the video, on screen long enough to be read — not the end card. For stills, in or immediately adjacent to the frame, in the viewer's eyeline. Not in T&Cs, not in a footer, not behind a click. Platform-native AI tags on social generally do the job, and you don't need to double up.
8. The Commission has done some of the work for you
There's a free icon set available, distinguishing AI-generated from AI-modified content. Use of it is voluntary, but it's tested and it's recognisable. The Code of Practice on Transparency of AI-Generated Content is also confirmed as an adequate route to demonstrating compliance. For scale: fines run to €15 million or 3% of worldwide turnover.
9. Write a policy, then fix your contracts
The single most useful thing you can do is write a practical AI content policy — not a legal document. Map the types of content you actually make, decide once whether each type needs labelling, specify how, then hand it to the people generating assets. It gives your team a consistent answer and, if a regulator ever asks, demonstrates genuine effort. In the early enforcement phase, that matters enormously.
Then sort out your contracts. There's real ambiguity right now about who owns the labelling decision between advertiser, agency and production company. That conversation is much easier at brief stage than at delivery.
FAQ
Does the label have to stay on screen for the whole film? No. It needs to appear at first exposure and remain long enough to be genuinely perceivable.
Our video has AI voiceover but live-action picture. Do we need an audio label? A visual label is generally sufficient. You can also be more specific — stating that only the audio is AI-generated is permitted and often preferable.
What about an animated character composited into a real environment? The animation itself signals unreality. Unless the environment is exceptionally photoreal, this typically falls outside the definition.
Does AI-generated music need labelling? If it uses a synthetic voice that sounds like a real singer, yes. Instrumental or clearly stylised work is a different question.
Does content made before 2 August 2026 need relabelling? No. The date of generation is what counts.
Who is liable — us, the agency, or the client? Legally, the deployer: whoever used the AI system to create the content. In practice this is unsettled across the supply chain, which is exactly why it belongs in the contract.
We're building AI into commercial work every week, and we've had to answer these questions on live jobs rather than in theory. If you're weighing up an AI-augmented production and want a straight conversation about what needs labelling and what doesn't — get in touch.
