The EU AI Act for marketing teams.

Marketing teams are the largest group of people using generative AI daily and the least likely to have read the EU AI Act. The good news: almost nothing marketing does is high-risk. The bad news: the obligations that do apply - Article 50 content marking and deepfake disclosure - became enforceable on 2 August 2026, and they attach to exactly the workflows agencies run every day.

If you generate images, video, audio or synthetic voices for campaigns, this is live law now, not a 2027 problem.

The four things that actually apply

The deepfake rule, plainly

Article 50(4) requires deployers who generate or manipulate image, audio or video constituting a deepfake to disclose that it is artificially generated. There are narrow carve-outs for evidently artistic, creative, satirical or fictional work - but the disclosure obligation is adapted rather than removed, and "it's an ad" is not one of the exemptions.

The practical cases that catch agencies:

• A synthetic presenter or AI-generated spokesperson. If it reads as a real person, disclose.
• Voice cloning of a founder, a customer or talent - including "we cloned our own CEO for the intro". Consent covers the personality-rights problem, not the AI Act disclosure.
• AI-generated testimonials or reviews. Beyond Article 50, these run straight into unfair-commercial-practices law. Fabricating customer voices is a different and worse problem than failing to label.
• Restored or "enhanced" footage of a real person that changes what they appear to say or do.

The one to stop today: AI-generated reviews, endorsements or testimonials presented as genuine. That is not an AI Act labelling question - it is deceptive commercial practice, and disclosure does not cure it.

Who owes the marking - you or your tool?

Article 50(2) is a provider duty: whoever puts the generative system on the market marks its output machine-readably. If you generate through Midjourney, OpenAI or Adobe, the marking duty for the raw output generally sits with them, and most now embed provenance metadata.

What you should not assume is that this survives your pipeline. Re-encoding, cropping, compositing in a video editor, or pushing through a social platform routinely strips provenance metadata. Article 50(4) - the deepfake disclosure - is a deployer duty and lands on you regardless of what your tool did. The four duties in full.

If you're an agency, this is a client question too

Agencies producing AI content for clients sit awkwardly: you generate, the client publishes. In practice the client is usually the deployer making the disclosure, and you are the one who has to tell them there is something to disclose. Getting that into the statement of work - who marks, who discloses, who keeps the record - is cheaper than arguing about it after a complaint.

It is also becoming a pitch asset. "We label AI-generated assets and can evidence it" is starting to appear in RFPs, particularly from regulated clients in financial services and health.

The 30-minute version

• List where AI generates published media, and check whether provenance metadata survives your export pipeline
• Add a visible label to synthetic media in-channel - not only metadata
• Add the AI disclosure line to any site chatbot
• Stop any AI-generated testimonial or review presented as real
• Run and minute one AI literacy session
• Write down what you checked, and date it