The question behind the question
Nobody asking this is really asking whether generative tools are permitted. They are asking whether a campaign can be run without creating a problem — regulatory, platform, or reputational. That is a disclosure and substantiation question rather than a technology question.
This is our operating read of the current landscape, not legal advice. Requirements differ by market and change; confirm your own position with qualified counsel.
The United States: the FTC position
The operative principle is that where AI created or substantially modified advertising content, consumers should be told. The distinction that matters is between tools that assist production and tools that generate what the audience sees: grammar checking and analytics are not the target, while generated imagery, video, voice and ad copy are.
Sponsored content carries two separate obligations. The paid relationship must be disclosed, and the AI involvement must be disclosed. Satisfying one does not satisfy the other, and treating them as interchangeable is a common and avoidable error.
Disclosure must be clear and conspicuous — meaning placed where it is actually encountered rather than technically present. For video, that means on-screen and early, not appended to a caption most viewers never expand.
Critically, AI does not launder a claim. Anything asserted in generated content must be truthful and substantiated exactly as if a person had said it. 'The model wrote it' is not a defence, and penalties are assessed per violation, which is what makes a large campaign an expensive way to learn the rule.
The European Union: AI Act Article 50
Article 50 applies from 2 August 2026, which makes this immediate rather than forthcoming for anyone running campaigns into the EU.
Deployers creating deepfake content must disclose that it was artificially generated or manipulated. The definition is broad — content resembling real persons, objects, places or events that would appear authentic to a viewer.
Two details catch advertisers out. Content that looks or sounds like a real person requires labelling even where there was no intent to deceive. And providers of systems generating synthetic audio, image, video or text must ensure outputs are marked in a machine-readable format, detectable as artificially generated. A visible caption alone does not satisfy the machine-readable requirement; provenance signalling is a separate technical obligation.
Penalties are set in millions of euro or a percentage of worldwide annual turnover, whichever is higher — which puts this outside the range of risks a growth team can quietly absorb.
What is unambiguously not allowed
Fabricated testimonials. Invented reviews. A synthetic person delivering first-person claims about an experience nobody had. Unsubstantiated health or financial claims. None of these become acceptable because a model produced them, and all of them are straightforward to identify after the fact.
This is the part of the landscape with no interpretive grey area, and it is where the serious enforcement risk sits.
A workable operating posture
Disclose by default rather than when asked. Apply the strictest applicable standard across a multi-jurisdiction campaign rather than maintaining variants — segmenting disclosure by market is a category of mistake you do not need. Obtain written consent for any real likeness or voice. Substantiate claims before production rather than after a complaint. Keep a provenance record so a campaign can be reconstructed months later.
The compliance posture and the trust posture are the same posture, which is convenient: doing this properly is also the thing that makes the work defensible to an audience.