EU AI Act Labels Are Live: A Shot-by-Shot Checklist for Producers
Since 2 August 2026, some AI shots in EU ads need a visible label. A four-question test, edit examples and who should own the label.
Cover illustration generated with AI for Fix It In Post.
Since 2 August 2026, some of the AI work in your EU deliverables needs a visible label and some doesn't. An AI-generated video of mice arguing about cheese is fine as it is. An AI-polished packshot that makes the product look better than it really is may not be. The rule is short; applying it shot by shot is the hard part.
Our view is that most production companies are asking the wrong question. They ask whether they used AI. The law asks whether the result could fool someone into thinking it is real. A spot can be heavily AI-assisted and need no label, or use one generated shot and need one. Here is how to tell the difference in the edit, and who should take responsibility for it.
What the law actually requires
Article 50(4) of the AI Act requires deployers of AI systems to disclose content that has been artificially generated or manipulated when it counts as a deepfake. The Act defines a deepfake as image, audio or video that resembles existing people, objects, places, entities or events and would falsely appear authentic. Under paragraph 5, the disclosure must be clear and distinguishable no later than the viewer's first exposure. It cannot sit only in metadata.
There is an important carve-out. Where the work is evidently artistic, creative, satirical or fictional, the duty shrinks to disclosing in a way that does not get in the way of the work. Advertising does not automatically qualify. A realistic product demo is not "evidently fictional" just because it comes from a creative agency.
The European Commission published guidelines and a voluntary Code of Practice on 20 July 2026, according to Paul Weiss. Under the Digital Omnibus, the separate provider duty to machine-mark output is deferred to 2 December 2026 for generative systems already on the market, and watermark interoperability to 2 February 2027. Those deferrals apply to the tool vendors. The deployer labelling duty, which is yours, applies now.
Penalties under Article 99(4) reach €15 million or 3% of worldwide annual turnover, whichever is higher. For SMEs, Article 99(6) flips that to whichever is lower, so the ceiling for a small production company is far below €15 million. It is still not a risk worth taking for the sake of a two-second icon.
Who is the deployer?
This is not just the brand's problem. In Lewis Silkin's reading of the guidance, the advertiser and the agency can each be a deployer: the brand because it approves the use case and signs off the creative, the agency because it chooses and runs the tools day to day. A production company generating shots in-house is in the same position as the agency. The firm's phrase is that the problem cannot be "quietly outsourced down the supply chain."
So write it into the contract. We suggest three clauses:
- AI log as a deliverable. The production company delivers a shot list showing which shots used generative or manipulative AI, with which tool, and what it changed.
- Label decision owned by the advertiser. The brand makes the final call on whether a label is needed and where it goes, based on that log. The brand controls media placement and sees every version of the spot.
- Label rendered by post. Post produces labelled and clean masters for every version, so nobody adds a label in the platform UI that misses the first frame.
The shot-by-shot checklist
For each shot in the AI log, ask these four questions in order. Stop at the first answer that settles it.
- Is the shot obviously unreal? Talking animals, dragons, cartoon versions of historical events. The Commission's examples, as reported by Davis+Gilbert, include the mice arguing about cheese as content that needs no label. If yes: no label.
- Is the AI change minor and cosmetic? Colour correction, lighting, background tidy-up, retouching, rescaling a product within the frame. Lewis Silkin reports that the guidance treats these as having only a minor effect on authenticity. If yes: no label.
- Does the AI change how the viewer judges the product, person or event being promoted? A product that looks better than the real one, a texture or finish that is not real, a realistic person who does not exist demonstrating the product. If yes: label.
- Does it show a real, recognisable person, place or event doing or saying something that did not happen? A digital double of talent, a cloned voice reading new lines, a real location changed in a way that matters to the claim. If yes: label, and check that you have the consent paperwork, which is a separate legal question.
Common edit situations
| Situation | Likely outcome | Why |
|---|---|---|
| Sky replacement on a car spot | Usually no label | Aesthetic background change; the car is real and unaltered |
| Sky replacement on a holiday resort ad | Label | The weather and view are part of what is being sold |
| AI crowd fill in a stadium wide | Grey area; label if the crowd supports a claim (e.g. "sold out") | Resembles a real event; context decides |
| Digital double of the lead actor for a pickup line | Label | A real person appearing to say something they did not film |
| Fully generated packshot of a real product | Label, unless it matches the product exactly | The guidance flags AI product images that can mislead about appearance or quality |
| Synthetic influencer reviewing a real product | Label | Named in the Commission's examples |
| Voice clone of the voice-over artist for a re-record | Label if the words are new; probably none for a pure audio clean-up | Technical audio fixes that leave the words unchanged are listed as exempt |
| AI denoise, upscale, regrade | No label | Assistive editing; no change in what is depicted |
The grey rows are our interpretation of guidance that Lewis Silkin itself calls extensive but not always clear. When the call is close, label. A small "AI-generated" mark costs very little, and arguing about it with a regulator costs a lot.
What a compliant label looks like
The Commission recommends an optional EU "AI" icon marked as generated or modified. Using it is not proof of compliance, and leaving it out is not a breach. What matters is that the label is legible, on screen at first exposure (so in the first frames of a pre-roll, not the end card) and readable on a phone. For social cut-downs, build the label into each aspect ratio's safe area rather than relying on a platform's AI tag. A platform tag may appear late or not at all in some placements.
Content published before 2 August 2026 does not need relabelling, according to both Paul Weiss and Lewis Silkin. A recut of an older spot is new content, though, so check it again.
The takeaway
Label for deception risk, not AI use: the test is whether a viewer could take the shot as real and be misled about what is being sold. Keep a per-shot AI log, put the label decision with the advertiser in writing, and deliver labelled masters from post. When a shot is borderline, label it.
This article is editorial guidance, not legal advice. Check specific campaigns with counsel.
