What the AI Act actually wants from ad agencies

What agencies have to disclose

Ad agencies keep asking what the AI Act actually wants from them. The short version is this: if you use AI to make image, audio or video that resembles real people, places or events and would pass as authentic, you have to say it's AI-generated or manipulated ^[1]. That includes deepfakes in ads.

For creative, satirical or fictional work, a general disclosure that AI content is present is enough, as long as it doesn't ruin the experience. Think of it like the "contains nuts" line on a packet of biscuits: it tells you AI was in the mix, but it doesn't interrupt the story.

The machine-readable marking of AI output, on the other hand, is the job of whoever provides the AI tool, not the agency using it ^[1]. The tool maker marks the output; the agency decides what to say about it to the audience.

Who the rules apply to

The AI Act applies from 2 August 2026 ^[2]. Article 50(4) covers users, which in this case means agencies, brands and anyone else deploying AI to make content for an audience ^[1].

If you're working on political ads, stricter rules apply. The ad has to be labelled, you need a transparency notice with the sponsor's name and the election it relates to, and if you use targeting with personal data, you need explicit consent and no sensitive data profiling. Service providers keep records for seven years ^[3].

What this means for you

First, decide now who checks whether a piece of work counts as a deepfake under the definition: content that resembles real people, objects, places, entities or events and would erroneously appear to be authentic ^[1]. That's a judgement call, so someone on the creative or legal side needs to make it before the work goes live.

Second, work out how your disclosure fits with the format. A note at the end of a 30-second video spot is fine; a watermark in the corner might not be. For satirical or fictional work, a general statement that AI content is present can be enough, as long as it's clear and doesn't wreck the viewing experience.

Third, if you're working on political ads, the transparency notice and the consent rules for targeting matter as much as the creative. The seven-year record-keeping duty means someone will be keeping your work on file long after the campaign ends ^[3].

When the evidence question turns up

Months after a campaign, a complaint or a rights dispute can turn on what was AI-generated, who approved it and what the audience saw. The client or a platform will ask, and a screenshot from someone's phone won't be enough.

That's where the campaign's paper trail matters. Which version of the brief did the client sign off? What did the AI tool generate, and which model version did it? Who approved the final cut, and is that the version that ran?

How a digital witness fits here

Months after a campaign wraps, someone will ask what the AI tool actually generated, which version the client approved and whether the final cut that ran is the one everyone remembers. A screenshot from someone's phone won't settle it.

The thing to seal is the brief when it's agreed, a record of each AI generation as it's created (the tool, the model version and the settings used), any licences for third-party material, the client's sign-off and the final versions as they're published. Sealing happens as each step occurs, so you're not scrambling to piece it together later.

Your system sends a fingerprint of each record to WitniumChain, through the REST API, the SDK or the MCP server, and gets a witness ID back. The record itself stays with you. When someone questions what ran, anyone you share the record with can check that it's the one sealed on the day by calculating the fingerprint again. The first 2,000 witnesses a month are free, and the documentation shows you how to start.

What it shows is which material existed and was approved at what time, unchanged since it was sealed. What it doesn't show is that the campaign was lawful, that a licence was valid or that a disclosure met the AI Act's standard.

How it would work

1. Seal the brief when it's agreed, each generation when it's created and each approval when it's given.
2. Seal the final cut as it's sent for distribution.
3. When someone questions what ran, share the sealed record and the verification link.

What to watch

2 August 2026 is the date the disclosure rules in Article 50 apply ^[2]. If you're working on political ads, the transparency and targeting rules under Regulation (EU) 2024/900 have applied since 10 October 2025 ^[3].

The AI Office has not yet published detailed guidance on what counts as an adequate disclosure for creative work. When that guidance arrives, it will be worth comparing your current practice against it.