AI ad compliance for supplements and health brands

20 September 2026

Supplements and health products attract more advertising scrutiny than almost any other category, and generated creative adds a second layer on top of the first. This is a practical summary of what applies and how to write within it. It is not legal advice, and it is not a substitute for your own counsel or your regulatory team.

Two separate sets of rules

It helps to keep them apart, because they fail differently.

Claim rules govern what you say about the product: structure/function claims, disease claims, substantiation, testimonials. These apply whether your ad was shot or generated, and they are enforced by regulators.

Synthetic media rules govern how you present generated people and content. These are mostly platform policy, enforced by ad review and account-level enforcement, and they are changing faster than the claim rules.

An ad can pass one and fail the other. Most rejections in this category are claim problems that people misattribute to the AI.

Claims: the line that matters most

The distinction that causes the most trouble is between a structure/function claim and a disease claim. "Supports healthy sleep" describes an effect on normal function. "Treats insomnia" claims to treat a condition, which for a supplement puts you in drug territory.

The line is crossed more often by implication than by wording. A script that never says "treats anxiety" but shows someone visibly distressed, then calm, with the product in between, can be read as making the claim. Implied claims count. So do claims made by captions, by on-screen text, and by the name of the product.

Practical approach: get the approved claim language in writing before anything is written, and treat it as the vocabulary rather than the ceiling. Writing around approved language produces better ads than writing freely and then cutting back.

Substantiation

Every objective claim needs support you hold before the ad runs. This includes the ones people forget: "clinically studied" (of what, at what dose, in whom), "#1 selling" (by what measure, over what period), "works in 30 days" (in what proportion of users).

Specific numbers in a hook are effective and they raise the substantiation bar. If your hook is "most people guess two hours, it's nine", you need a source for nine. Keep the source with the asset, not in someone's memory.

Testimonials and generated people

This is the sharpest issue where the two rule sets meet.

A testimonial is a statement of personal experience, and it must reflect the genuine experience of a real person. A generated person cannot give one. Writing a script in which a generated creator says "I've taken this for three months and I sleep through the night" is a fabricated testimonial. That it is obviously synthetic to you does not help; the question is what the viewer reasonably understands.

This is the most common serious mistake in the category, and it is entirely avoidable. Generated creators can:

• explain how something is supposed to work;
• ask questions, including in a street-interview format;
• react to a fact or a number;
• demonstrate use;
• voice a general statement that is not a personal claim.

They cannot report personal results. If a first-person testimonial is the mechanism your ad depends on, you need a real customer who genuinely had that experience, with a disclosure that results are not typical where that applies.

The same applies to anyone presented as an expert. A generated person in a white coat implying clinical authority is a problem regardless of what they say.

Disclosure of synthetic media

The major platforms require disclosure when realistic content is AI-generated, particularly where it depicts people. Meta, TikTok and YouTube each have their own labelling requirements, some applied automatically from metadata, some requiring you to declare it at upload.

Practical position: declare it. Use the platform's own toggle rather than relying on detection, and where the ad depicts a photorealistic person delivering information, an on-screen note that the presenter is AI-generated is cheap insurance. Undisclosed synthetic people in a health context is the combination most likely to attract enforcement, and account-level enforcement is expensive in a way a rejected ad is not.

Things that reliably get rejected

• Before-and-after body imagery, generated or not.
• Implying a personal health condition in the viewer ("your gut is inflamed").
• Specific weight-loss numbers or timeframes.
• Anything implying diagnosis, cure, prevention or treatment of a disease.
• Fabricated urgency about a health risk.
• Generated people presented as doctors, pharmacists or researchers.
• Reviews or endorsements that were not given.

Building review into production

The cheapest place to catch a compliance problem is in the script, before anything is rendered. A workable process:

1. Approved claim language agreed before writing.
2. Scripts reviewed as text — including captions and on-screen words, which are frequently missed.
3. Substantiation filed against each specific claim.
4. Synthetic disclosure decided at the same time as the script, not at upload.
5. A final check of the finished cut, because editing can create an implication the script did not contain.

Twenty hook variants make this more important, not less: it is easy for one variant in twenty to drift across the line while nineteen are fine. Review the variants, not just the concept.

The useful reframe

Compliance constraints in this category tend to push creative toward mechanism, curiosity and specificity rather than promises — and those generally perform better anyway. An ad that explains something surprising and true usually outperforms one that promises a result, and it is the one that survives review.