Listings and ads
AI-Generated Product Descriptions: Compliance Risks for CPG Brands
A product description written with a model often reads cleanly. The sentences are grammatical, the bullets are parallel, and the tone matches other brands in the category. That fluency is the risk. The draft can shift positioning: a therapeutic implication, an overstated performance claim, a government-endorsement claim, or a comparison the brand never substantiated, written so it looks like finished copy.
The brand that publishes the text is the advertiser. The Federal Trade Commission’s substantiation standard does not have a separate, lower bar for copy a model drafted. Objective claims still need a reasonable basis before they run. The FTC Policy Statement Regarding Advertising Substantiation and the Health Products Compliance Guidance are the references for that expectation.
What a draft tends to add
Models learn from text that is already online, and category pages repeat the same kinds of promises. A prompt that says “write a premium skincare listing” will often reproduce those patterns because they are common, not because they fit this product. Review language that could imply treatment, prevention, clinical validation, government endorsement, or effects beyond the product’s intended category.
Consider a fictional body lotion, Harbor Daily, positioned as an everyday moisturizer. This is not a real product. A prompt asks for bullets “in the style of top-selling creams.” A typical draft can introduce clinical-validation language, condition-oriented claims, a government status the product does not have, an absolute absence claim, or a professional endorsement nobody documented. None of that came from Harbor Daily’s claim file. Condition-oriented language can change how the product is interpreted, which is discussed in when cosmetic marketing changes the interpretation and cosmetic vs. drug positioning. FDA explains cosmetic oversight, including what it does not do, in its cosmetics labeling and regulation materials, starting with Cosmetics Labeling and Is It a Cosmetic, a Drug, or Both?.
The prompt does not review the claim
Teams sometimes add “make it compliant” to the prompt. That instruction does not check the formula, the evidence, or the intended category. The model does not have the brand’s substantiation file unless someone pasted it in, and even then it can misread the file. Cautious wording can still be an objective claim. Hedging does not replace evidence, and a single keyword is not the whole problem. Surrounding copy, imagery, and the product category decide what a shopper thinks was promised.
The same problem appears when the model is asked to rewrite a careful paragraph so it ranks for a condition people search. The search demand becomes the positioning. Search volume is not a regulatory category. If the brand did not plan to sell a drug, the listing should not be optimized into that story.
Images and alt text get less attention. A model that describes a before-and-after of a condition has created a claim. So has a caption generated for a retailer module. Review every field that will be public, including titles, bullets, enhanced content, ads, and image text, as one presentation.
Who is responsible
Agencies, freelancers, and internal marketers who use a drafting tool are still writing the brand’s advertising. A contract that says the vendor “used AI” does not move the claim onto the tool. Someone who knows the product should compare the draft with the positioning the brand actually intends. A short claim sheet — what the product is for, and which evidence supports which kind of statement — belongs with that person before publication. Pasting the sheet into the model is not the same as a review.
If the copy will also be spoken by a paid creator, the claim and the disclosure are both in scope. A model-written caption that omits a material connection is a disclosure gap on top of a substantiation problem. See FTC influencer disclosure basics.
A workable review, not a ban
Drafting tools are useful for structure: shortening a paragraph, proposing a headline the team then edits, or turning an approved idea into plain language. The control is the edit. Before a model sentence ships, the team should be able to say what a shopper will think the product does, and whether that reading matches the category and the evidence. ProductGuard reviews the full presentation rather than isolated words.
Strong proof language still needs evidence that fits this product, even when a study exists somewhere in the category. That question is covered in what “clinically proven” is asking you to show.
Keep version history. If a listing is later questioned, the useful record is the claim, the evidence, and who approved the public text. A chat transcript that says the model was told to “be accurate” is not that record.
What deserves attention
- Drafts that imply treatment, prevention, clinical validation, government endorsement, or effects beyond the intended category.
- Category-positioning shifts caused by surrounding copy, not by one word in isolation.
- Titles, image text, and captions, reviewed with the long description.
- Evidence-sensitive claims compared with this product’s support, not with competitor bullets.
- A named person who approved the final text. The drafting tool is not the approver.