Operation AI Comply and claims about your AI
FTC Section 5 applies to what you say your AI does, not just what you generate with it. Capability claims rewritten so they survive a deception review.
Most compliance conversations about generative AI run in one direction: what happens if the output is a problem. Deepfakes, likeness, copyright, disclosure. All real. All downstream of the model.
The FTC's Operation AI Comply, announced in September 2024, went the other way. It targeted deceptive AI claims and schemes — what companies said their AI could do, sold, or would earn for the buyer. Not the pixels. The pitch.
If you sell anything with "AI" in the description, or run an agency whose deck says AI does part of the work, the exposure is on that side of the sentence too. There is no AI-specific advertising rule in the US. Section 5's prohibition on deception and unfairness covers it, and Section 5 has been catching overstated capability claims for a very long time in every other category.
General information, not legal advice.
The claim, not the technology, is the subject
Section 5 analysis does not care whether a transformer is involved. It asks three things about a claim: what representation does the ad convey to a reasonable consumer, is that representation material to a purchase decision, and did you hold adequate substantiation before you made it.
Every one of those is answerable about "our AI writes ads that convert." None of them is answerable by pointing at a model card.
The specific failure mode with AI claims is that they are unusually easy to make and unusually hard to substantiate. Marketing copy about a physical product is bounded by the product. Marketing copy about a system whose behaviour varies per input has no natural ceiling, and the vocabulary available — intelligent, autonomous, learns your brand — is all implication and no measurement. That combination is what a sweep like Operation AI Comply exists to find.
Four claim shapes that draw attention
Outcome and earnings claims. "Build a passive income business with AI." This is the highest-risk category on the list by a distance, because the claim is about the buyer's results and the substantiation would have to be about typical buyers rather than best cases. If you cannot describe the typical result, do not describe any result.
Capability claims that outrun the system. "Fully automated," "no editing needed," "handles the entire workflow." Each of these is falsifiable in about four minutes by a customer who tries it. If a human reviews outputs before delivery, the workflow is not fully automated, and describing it as such is a false statement about the product rather than an aspirational one.
Comparative and superiority claims. "The most accurate model," "outperforms every alternative." Superiority claims need substantiation of the comparison, on a methodology, against the named alternatives. If a leaderboard is the basis, name the leaderboard, because different arenas measure different things and swapping them is its own accuracy problem. Our note on why brand trust depends on AI transparency covers why this one costs more reputationally than legally.
Claims about what the AI is. Describing a rules engine, a template library, or a human team as AI is a straightforward misrepresentation of the product. So is implying a proprietary model where you are calling somebody's API. Which model actually runs matters to buyers, which is exactly what makes it material.
Rewriting claims so they survive
The productive move is not to strip the marketing. It is to convert implication into a statement you can evidence.
| Claim as written | The problem | A version you can hold |
|---|---|---|
| "Our AI writes ads that convert" | Outcome claim about the buyer's results, unsubstantiated | "Generates ad variants from your brief for you to test" |
| "Fully automated end to end" | False if a human reviews anything | "Automates generation and assembly; you review before publish" |
| "The most accurate AI video model" | Superiority claim with no named comparison or method | "Ranked first on [named leaderboard] as of [date] on [named metric]" |
| "Powered by our proprietary AI" | Misrepresents an API integration | "Runs on [named model] through our pipeline" |
| "Replaces your creative team" | Implied capability well beyond delivery | "Handles first drafts and variant generation your team directs" |
| "Trained on your brand" | Implies fine-tuning that may not exist | "Applies your brand kit to every generation" |
The right column is duller and it is also true, which means nobody has to reconstruct what you meant a year later when a customer complains. The rewrite that matters most is the fourth row. Naming the model you actually run is the single cheapest credibility move available to a reseller, and it removes an entire category of misrepresentation risk. Our own model catalogue names each model, and the agent chat names the models it routes a prompt to when it fans one request across several of them.
Substantiation you should already be holding
For each material claim in your own marketing, you want a file with four things in it.
- The claim as written, in the exact words that shipped, per surface. Website, deck, ad copy and app store listing frequently disagree, and the strongest version is the one that gets tested.
- The evidence, dated. Benchmark run, internal test set, customer cohort, whatever it is. "We believe" is not evidence.
- The scope the evidence actually supports. A result on one content type and one language does not support a general claim.
- A review date. Model capability claims decay fast. A superiority claim that was true against a named competitor in March is a live misrepresentation in August if the competitor shipped.
That is a lightweight version of the governance layer for AI content, pointed inward at your own marketing rather than outward at what you generate.
The adjacent rules that catch the same teams
Section 5 is the general instrument. Two named rules sit next to it and catch a lot of the same conduct more directly.
The Rule on the Use of Consumer Reviews and Testimonials, effective 21 October 2024, prohibits AI-generated fake reviews and fabricated testimonials. Anyone generating social proof for a product page or a launch deck should treat this as the binding constraint. A synthetic customer is not a design element.
The Government and Business Impersonation Rule, effective April 2024, covers impersonation of businesses and government agencies. It bears on AI marketing in a specific way: a generated case study, a fabricated integration badge, or an implied partnership with a named platform is impersonation-adjacent, and none of it becomes acceptable because a model produced it.
If you also run creator partnerships, the disclosure rules for endorsements apply on top of everything above, and they are walked through in our guide to the FTC rules for AI creator sponsorships.
The reseller's specific problem
If your product wraps somebody else's models, three claims deserve extra care.
Ownership. "You own everything you generate" is a claim about legal status, and legal status is not entirely yours to grant. Provider terms can assign whatever interest the provider holds, but they cannot create copyright where the human-authorship test fails, and in the US purely AI-generated material is not registrable. The Supreme Court declined to take Thaler v. Perlmutter on 2 March 2026, which leaves that line where it was. Say what you actually give the customer — the files, unrestricted commercial use, no watermarks — rather than making a statement about copyright you cannot back. The copyright and safety guide covers the distinction in more detail.
Indemnity. If you do not offer one, do not imply one. Provider indemnities are generally tied to enterprise or API tiers and conditioned on using safety filters, and an upstream indemnity you benefit from is not automatically an indemnity your customer benefits from.
Pricing clarity. Usage-based products invite claims about what a given spend buys. Ours are in credits and the plan ladder is on the pricing page, which is a deliberately boring answer, and boring is the correct target for a claim that has to hold up when a customer checks it.
FAQ
Is Operation AI Comply a rule I need to comply with?
No. It was an enforcement sweep announced in September 2024, not a rulemaking. What binds you is Section 5's prohibition on deceptive and unfair practices, which applies to everyone regardless of any sweep. The sweep is useful mainly as a signal of where attention was directed.
Does "AI-powered" on its own create exposure?
On its own, generally not — it is vague enough to convey little. The risk starts where the phrase carries a specific implication a reasonable consumer would draw and you cannot substantiate. "AI-powered scheduling" is fine if scheduling is AI-driven. "AI-powered results" is an outcome claim wearing a technical jacket.
We publish a benchmark. Is that safer than a marketing claim?
Safer if it is reproducible and honestly scoped, riskier if it is selective. Publishing a methodology and a date makes a comparative claim substantiable. Publishing a chart with no method, no date and a favourable prompt set makes the same claim harder to defend, not easier, because now you have documented that you tested.
What about claims in a demo video?
Demos convey claims exactly as copy does. A demo showing output quality or speed not typical of real use is a performance representation. Show real runs, or state the conditions on screen. The same principle governs the disclosure line you attach to generated marketing assets.