Not legal advice, and anything in a regulated category belongs in front of counsel. What this is: a plain account of where AI-generated advertising creates real exposure for a US business, written because the tools produce precisely the copy patterns that draw scrutiny and will never warn you they have done it.
The principle
The advertiser is responsible for the advertisement. Not the tool, not the agency, not the freelancer. The FTC has been explicit that using AI does not change who is accountable for a claim, and "the model wrote it" has no standing.
💡 What makes this genuinely risky is that AI copy reads with total confidence. A model asked for a high-converting ad will produce outcome guarantees and specific-sounding statistics by default, because that is what high-converting ad copy looks like in its training data. It has no concept that those sentences carry legal weight.
Substantiation: the rule that catches most of it
US advertising law requires that you have a reasonable basis for a claim *before* you make it. Not that you could find support later — that you held it at the time.
Ask an AI tool for ad copy and you will reliably get: "trusted by 5,000+ businesses", "customers save an average of 40%", "rated #1 in customer satisfaction". None of these are retrieved from your records. They are generated to fit the rhythm of the sentence.
💡 One habit removes most of this exposure: no number, superlative or comparative claim goes into an ad unless you could produce the documentation within an hour. Applied consistently, it is close to a complete answer, and it costs you nothing — specific verifiable claims outperform vague superlatives anyway.
The Endorsement Guides and synthetic people
This is the area AI video tools walk straight into. The FTC's Endorsement Guides require that an endorsement reflect the honest opinion of a real endorser, and that material connections be disclosed.
An AI avatar speaking to camera about how well a product worked is, to a reasonable consumer, a testimonial. If no such customer exists, that is a fabricated endorsement regardless of how the tool markets the feature. The same applies to AI-generated images of "customers", generated review text, and star ratings that were not collected.
- 1An avatar presenting product information as a spokesperson is fine. An avatar recounting a personal result is not, unless a real person had that result and the presentation makes the illustration clear.
- 2Generated before-and-after imagery is a fabricated result, not a stylistic choice.
- 3Review and rating copy must come from reviews and ratings you actually received.
There is a simpler line worth adopting than any legal test: never publish a review, testimonial, customer name or rating that is not real. It removes the entire category of risk and protects something more valuable than a campaign.
The categories that draw the most attention
Health and wellness
Outcome claims about the body require competent and reliable scientific evidence. AI writes "clears acne", "reverses hair loss", "permanent results" freely. This is the most-enforced area there is.
Earnings and business opportunity
Income claims carry specific requirements. "Our clients make $10k a month" is a claim about typical results and needs to be true of typical results.
Financial services and credit
Highly regulated, with disclosure requirements AI copy will never include because it does not know they exist.
Pricing and "free"
"Free" must be free. "From $99" requires that a meaningful number of buyers get it at $99. Both are stock AI phrases.
State-level AI disclosure
A moving target worth tracking rather than assuming. Several states have enacted or proposed rules touching AI-generated content, particularly around synthetic likenesses and political advertising, and requirements differ by state. If you advertise nationally with AI-generated people in your creative, this is worth a conversation with counsel rather than a blog post.
Platform enforcement is faster than any regulator
Meta and Google run their own automated review, and three patterns get rejected constantly — all three of which AI produces by default.
- 1Implied personal attributes. "Are you struggling with debt?" is rejected; "Our service helps people manage debt" is not. AI addresses the reader directly by default, which is exactly the trigger.
- 2Before-and-after imagery in health and cosmetic categories, regardless of authenticity.
- 3Unrealistic outcome claims, at a threshold neither platform publishes, which makes appeals slow.
Rejections are not legal incidents, but a pattern of them affects account standing — and a restricted ad account is a far more immediate business problem than a regulatory one.
A workable process
The one-hour rule
Every claim evidenced within an hour, or removed. Keep the evidence in a folder organised by claim.
No synthetic endorsers
Avatars present, they do not testify. No generated reviews, names or ratings, ever.
Preview everything
A tool that auto-publishes has removed your only review step. Seconds of human attention prevents most of this.
Specific over superlative
"Same-day appointments, board-certified" beats "the best in the state" both legally and commercially.
The review step stays yours
Adyft shows you every headline, image and audience before anything publishes, and does not generate reviews, testimonials or customer names. Fourteen-day free trial.
Start free — no card needed