A marketing agent generates ad copy for a fitness app. It produces compelling claims like "Lose 20 pounds in 30 days, guaranteed!" and "Scientifically proven to triple your metabolism." The product makes no such guarantees and has no published studies. Legal flagged the campaign as both deceptive advertising and FTC-actionable.
What's the architectural fix?
Why did you pick that answer? Two or three sentences. The act of articulating it is what builds the judgment — not the click that follows.
Marketing claims have to be tied to verifiable facts — not just because of FTC liability, but because deceptive advertising erodes brand trust independently of legal consequences. The architectural fix is to ground generation in a structured product-truth document: the actual capabilities, the actual studies, the actual claims the company can substantiate. Generation can't produce claims that aren't in that document. This is RAG applied to marketing — the source of truth grounds what the agent is allowed to say.
Marketing copy LLMs are trained to be compelling, which is structurally in tension with being scrupulously accurate. The model isn't maliciously deceptive — it's producing language that fits the marketing genre, which often includes hyperbolic claims. Truthfulness has to be enforced architecturally.
Human review catches some issues but doesn't scale — and reviewers can miss subtle claims (the "scientifically proven" phrasing in particular is easy to overlook). Structured grounding catches claims that human review wouldn't consistently flag.
Asking another Claude call "is this misleading" without giving it the actual product truth produces guesses about misleadingness — the second LLM doesn't know what claims the company can substantiate either. Without grounding, both passes produce plausible-sounding outputs unconstrained by reality.