The standard does not change with the tool
The consumer-protection direction in the United States, as the site's disclosure and data privacy guide summarises it from the FTC's business guidance, is that the test for deception does not change with the technology. Claims must be truthful and substantiated whether a person or a model wrote them, endorsements must be genuine, and using AI to deceive is treated as deception rather than as a technology question. In the UK the ASA applies the advertising codes across media, and the CMA addresses hidden advertising under consumer protection law, as the influencer contracts and disclosure guide describes. Other markets have their own regulators with similar expectations.
That has a practical consequence for how a marketing team thinks about AI. The question is never "is this allowed because AI wrote it" or "is this risky because AI wrote it", but the ordinary one: would this mislead a reasonable customer, and can we prove what it says? Generated copy, generated imagery and generated people are judged on what they communicate. The tool only matters in two ways. It makes misleading material quicker to produce at volume, so errors scale. And it introduces new kinds of claim, chiefly claims about AI itself and content that looks like it came from real people, which the rest of this guide takes in turn.
The disclosure and data privacy guide owns the question of when to tell customers that AI was involved and how to handle their data. This guide is about the advertising-standards side: what a claim says, what evidence sits behind it, and whether the voices in an advert are who they appear to be.
Claims that a product uses AI
"AI-powered", "smart", "intelligent", "personalised by AI" and "learns your preferences" are claims. The site's disclosure guide puts it directly: "AI-powered" is itself a claim that must be true. Inflating the role of AI in a product, sometimes called AI washing, misleads in two ways. It can misdescribe how the product works, and it can imply benefits the AI component does not deliver.
Four questions test an AI claim before it runs:
- Is there AI in the product at all, as a customer would understand the word? A rules-based filter or a fixed recommendation list described as AI invites the question of what the word means in your copy.
- Does the AI do what the copy says it does? "Personalised by AI" implies the output differs by person in a meaningful way. If it differs only by a segment chosen manually, the claim overstates.
- Is the benefit attributed to the AI real and shown? "Our AI finds you the best price" is an outcome claim and a superlative at once, and both need evidence.
- Is the comparison fair? "Smarter than" or "more accurate than" a rival or a previous version is a comparative claimA marketing claim comparing a product with an identifiable competitor or earlier version, held to like-for-like, verifiable and fair comparison rules., with the same evidence expectations as any comparison.
The fix for most AI claims is specificity. Describe what the feature does in plain terms ("suggests replies based on your past orders") rather than what it is ("AI-driven"), and claim only the benefits your testing supports. Specific claims are easier to substantiate and more persuasive than vague ones.
Reviews, testimonials and endorsements
This is where AI creates the most obvious risk, because a model can produce text that looks exactly like customer voice. The rules on reviews, as the reviews and social proof guide sets them out from the FTC's business guidance, are already clear: fake reviews are prohibited, incentives cannot be conditional on sentiment, suppressing negative reviews misrepresents the picture in the same way as inventing positive ones, and endorsements from people with a material connection to the business must say so. A generated review is a fake review whatever prompt produced it.
The same principle reaches further than the review section. Social ProofEvidence that other people have chosen, used or approved something, used to reduce a buyer's doubt. in all its forms borrows its force from the belief that real, independent people said it. That covers testimonials on a landing page, quotes in an email, "what our customers say" carousels, star ratings, case-study quotations and comments presented as UGCContent about a brand made by customers or creators rather than by the brand itself.. Each needs a real person who said it, permission to use it and wording that matches what they said. AI can legitimately help around the edges: sorting real reviews by topic, drafting the request email that asks customers for a review, suggesting which genuine quotes fit a page. It cannot supply the words. The tutorial on writing a customer case study shows how to keep a story's quotes in the customer's own words, from the recording through to their written approval.
Endorsements by creators follow the endorsement rules regardless of how the content was made. The FTC's published questions and answers on its Endorsement Guides explain that a material connection between an endorser and a brand should be disclosed clearly and conspicuously, and the site's influencer guide describes the UK position through the ASA and the CMA. If a creator uses AI to produce sponsored content, the relationship still needs disclosing, and the creator should still only say what they believe and what the brand can support.
Three practices cause trouble in AI-assisted work specifically. Polishing a real review until it reads better changes what the customer said. Generating "composite" testimonials "based on real feedback" presents a summary as a person. And seeding comments or replies with generated accounts creates endorsers that do not exist. All three should be off limits in the AI use policy. When a customer asks outright whether the reviews or images are real, the guide to answering customer questions about AI covers how to give an answer that matches what the business actually does.
Synthetic people and virtual creators
Generated faces, voices and full virtual personalities are now easy to produce for adverts, explainer videos and social accounts. The question is not whether they are permitted in principle, but what they communicate. A synthetic presenter reading a product description is a production choice. A synthetic "customer" describing their results is a fabricated endorsement, as the disclosure and data privacy guide puts it. A synthetic "doctor" or "expert" recommending a product implies expertise and experience that no one has.
General practice that follows from the standards above:
- Never present a synthetic person as a real customer, employee or expert, or as someone who has used the product.
- Label synthetic people and voices visibly where they could be taken for real, and follow each platform's documented labelling rules for realistic generated media at the time of upload.
- Do not generate a likeness or voice of a real person without their documented agreement; that is an impersonation question as well as a consent one.
- Treat a brand-owned virtual creator as advertising throughout: its posts are the brand's own marketing, and its "opinions" are the brand's claims, with the same evidence expectations.
Substantiation before publication
Substantiation means having evidence for a claim before it is made, not finding it after someone asks. The FTC's business guidance and the UK advertising codes the ASA applies both expect objective claims to be backed by evidence, and the accuracy and review workflows guide builds that expectation into the claim check. AI raises the stakes because a model writes claims freely and has no idea which of yours are supported.
A substantiation file is the working answer. For every objective claim a team makes regularly, one entry records the exact wording approved, what it means in plain terms, the evidence (test results, data extracts, certifications, the documented behaviour of a feature), the date the evidence was last checked, the markets where the claim is approved and the person who approved it. Generated copy is then checked against the file: a claim that matches an entry passes; a claim that is new, stronger or broader goes back for evidence or is softened. The file also protects the team when a model rephrases an approved claim into something it is not: "helps reduce" becomes "eliminates" in one rewrite, and the check catches it. The tutorial on building a claims library builds the file across all marketing, not only AI-assisted copy, and connects it to intake. The academy exercise on an AI-drafted email with an unproven claim at an opticians chain practises writing the file entry a claim should have had once it has already reached customers.
Outcome claims deserve special care. "Customers save hours each week" or "see results in days" are the claims models generate most readily and teams can support least often. If a claim describes typical results, the evidence has to show they are typical, not just possible. Treat any figure or outcome in generated copy as unverified until the file says otherwise.
Thistlecove Skincare's claim check
Thistlecove Skincare, an invented online skincare brand, uses an assistant to draft 30 product-page and email variants for a new moisturiser and a skin-quiz feature. The claim check against its substantiation file flags 14 claims. "AI-powered skin analysis" is softened to "a quiz that recommends products from your answers", because the feature applies 12 fixed rules rather than a model. "Clinically proven to hydrate for 48 hours" matches an entry with a test report, so it passes in the approved wording only. "Loved by 10,000 customers" has no source and is cut. 3 drafts include testimonial-style quotes that no customer wrote; they are removed and replaced with 2 genuine reviews used with permission. The launch creative's presenter is a synthetic face, so the team adds a visible label and removes the line "I've used it for a month". The review takes the brand manager and a compliance reviewer about 3 hours, logged against the launch.
Who checks what
Claims and endorsements need more than one pair of eyes, and each pair needs a defined job. The table sets out one division of labour; it fits inside the sign-off routes in approvals and compliance reviews.
| Item | Prompter or writer | Editor | Specialist (legal, compliance, subject expert) |
|---|---|---|---|
| Claims that the product uses AI | Describes the feature in plain terms from the product brief | Checks the description against the substantiation file | Approves new or stronger wording |
| Objective and comparative claims | Uses only approved wording from the file | Flags anything new, stronger or broader | Approves evidence and markets |
| Outcome and typical-result claims | Avoids them unless the file has an entry | Removes figures with no source | Decides whether evidence shows typicality |
| Reviews and testimonials | Uses only real customers' words with permission | Checks quotes word for word against the source | Reviews incentive and connection disclosures |
| Creator and partner content | Briefs the disclosure wording | Checks the label is clear and upfront | Reviews contracts and disclosure terms |
| Synthetic people and voices | Records that the asset is synthetic | Checks labels and that no real person is implied | Approves any use of a real likeness or voice |
| Regulated categories (health, finance, environment) | Escalates at the brief | Holds publication until specialist sign-off | Approves claims and disclaimers |
Track the check like any other quality measure. The number of claims flagged per batch of generated copy is a useful KPIA metric tied directly to an objective. for the prompt and source pack: a high rate means the model is being asked to write claims rather than to express approved ones. When generated pieces are among the work your own legal or compliance reviewer keeps sending back, the diagnosis for legal rejecting marketing work separates them from the other reasons for returns, such as claims missed at intake or edited approved wording.
Common mistakes
- Treating "AI-powered" as decoration. It is a claim about how the product works and needs to be true and shown.
- Letting a model write testimonials. Generated customer voice is a fabricated endorsement, however closely it echoes real feedback.
- Polishing real reviews. Tidied wording is no longer what the customer said.
- Synthetic experts and customers. A generated person who appears to have used or recommended the product implies experience nobody has.
- Finding evidence after publication. Substantiation comes before the claim runs.
- Approving the claim, not the wording. A model rewrites approved claims into stronger ones unless the check compares wording against the file.
- Assuming one market's rules cover all. The FTC, the ASA and the CMA each set out their own guidance, and other markets have their own regulators.
- Reading this page as legal advice. It describes general direction; your adviser decides what applies.
