AI-Generated Ads Need a Claim Evidence Gate, Not Just a Label

AI Trust & Governance

14 October 2026 | By Ashley Marshall

A label saying “AI-generated” will not rescue an advert that makes your product look better than it is. The real approval question is what the customer believes.

The misleading impression matters more than the production method

An AI-generated advert is not automatically unlawful or automatically in need of an AI label in the UK. That is the wrong starting question for a marketing team. The useful question is what a reasonable customer will believe after seeing the finished creative. Does the image imply that a product has a feature it lacks? Does a synthetic voice sound like a real customer testimonial? Does a generated scene suggest a result your service cannot reliably deliver? If so, a small disclosure in the caption is unlikely to repair the underlying claim.

Reed Smith explained in September 2026 that the UK has no single AI Act and advertising remains governed through existing advertising and consumer protection rules. Its review says there is no general UK duty to label AI-generated advertising, but omission of AI use may be misleading where it changes the impression made. Browne Jacobson gives a particularly useful example: an AI visual making a toy look more realistic than the product. Marking that visual as AI-generated would not make the product depiction accurate.

What this means in practice is that the approval workflow must review impressions, not only literal text. A legal reviewer can sign off a perfectly true sentence while the image beside it makes an unsupported promise. Ask a person unfamiliar with the campaign what they think the product does, how quickly it works and what result they expect. Record their answer before the team explains the creative intent. If that expectation exceeds the evidence, change the image or claim. Do not ask the AI tool for another version of the same misleading idea.

Build a claim register before making variants

Generative tools make it cheap to produce dozens of variations. That is precisely why claim control needs to happen before production. Start with a one-page register containing each material claim, the evidence supporting it, the date the evidence was checked, its owner and the channels where the claim may appear. Include implied claims: product appearance, customer satisfaction, safety, performance and availability. A picture can make a claim without putting it in words.

The number of variants should not be confused with the number of approved claims. A single verified proposition may appear in ten layouts, but every layout must preserve the qualification that makes it true. A tool that shortens copy may drop “up to”, remove an eligibility condition or turn a customer scenario into a guaranteed outcome. A generated testimonial can sound authentic even when no customer said it. Treat these as new claims, not harmless edits. Baker McKenzie notes that background AI imagery generally presents a different disclosure risk from AI elements that shape product claims or consumer expectations.

What this means in practice: give the model approved source material and prohibited claim patterns, then have a human compare the output with the register. For a finance product, the reviewer might check rates, eligibility and representative examples. For a software offer, check whether a generated dashboard depicts live features rather than a concept. For a professional service, verify the quoted result and the context in which it was achieved. Keep a copy of the approved creative and its evidence. If the evidence expires, pause the variants that depend on it. This is less glamorous than prompt engineering, but it is the control that stops a rapid content workflow becoming a rapid misstatement workflow.

Separate disclosure from substantiation

Disclosure answers “how was this made?” Substantiation answers “is this claim true?” A campaign can pass one test and fail the other. A clearly labelled synthetic product demonstration can still show an impossible result. An unlabelled AI-generated background can be unobjectionable if it does not alter a buyer's understanding. Conflating the two leads to a box-ticking process: the team adds a label and assumes the job is finished.

Set two explicit decisions for every high-impact asset. First, identify whether a consumer could reasonably regard a person, scene, endorsement or product result as authentic and whether that perception is material. Second, ask whether every express and implied product claim is supported by evidence. Baker McKenzie explains that the UK disclosure question depends on whether AI use could mislead consumers. The same analysis says that EU AI Act Article 50 adds a separate transparency question for certain realistic synthetic content. UK businesses advertising to EU audiences should therefore assess the market and placement rather than applying one domestic checklist globally.

A useful practical example is an estate agent generating a furnished room image. A clear “computer-generated furnishing” label tells viewers what the image is. It does not excuse moving a wall, enlarging a window or implying a property has fixtures it does not include. Another example is a recruitment firm using a synthetic staff portrait: if the image implies an actual employee is giving a personal account, the implied endorsement deserves scrutiny even if the written copy is restrained. Approvers should annotate the final asset with both decisions. If either answer is uncertain, test the consumer impression with someone outside the creative team and seek specialist advice for the relevant market before launch.

Give your approval gate three people and a stop rule

Approval does not have to become a committee that delays every social post. It does need named responsibility. The campaign owner knows the intended message. A product or service owner knows what can actually be delivered. A final reviewer can challenge the consumer impression and the quality of evidence. In a small business, the same person may hold more than one role, but the review should still ask three distinct questions. What is being promised? Can we deliver and prove it? What will someone reasonably take away?

Use a simple risk tier. Low risk covers decorative backgrounds and routine copy polishing that do not change claims. Medium risk includes product images, prices and comparative statements. High risk includes synthetic testimonials, realistic people, health or financial outcomes and any claim aimed at a vulnerable audience. Require recorded evidence and explicit sign-off for medium and high risk. Give any reviewer a stop rule: if a visual or sentence creates a material claim that cannot be traced to evidence, it does not publish. That is not an invitation to debate how convincing the creative looks.

Reed Smith notes that the UK's Digital Markets, Competition and Consumers Act unfair-commercial-practices provisions came into force in April 2025, with the CMA able to impose penalties of up to 10% of global turnover in relevant cases. That is a statutory maximum, not a prediction about any individual advert, but it makes the point that consumer law is not optional because a tool produced the asset. The counterargument is that these checks kill the speed advantage of AI. They need not. Approve reusable claim blocks, product image standards and disclosures once, then let creative teams vary the presentation within those boundaries. Escalate only when the meaning changes. Speed comes from a prepared evidence library, not from skipping review.

Keep a defensible record and revisit live creative

Once the asset is approved, preserve the final image or video, the prompt or brief where practical, the claim register version, evidence files, reviewer, approval date and channels used. This is not about archiving every experimental prompt. It is about being able to reconstruct why a published claim was considered fair. Record the specific disclosure decision as well. If a complaint arrives three months later, “the agency handled it” is not an evidence trail.

Set expiry triggers rather than relying on a calendar alone. A price change, software release, revised service scope, new testimonial withdrawal or changed regulation should reopen affected approvals. For paid campaigns, check the live platform preview because crops and overlays can obscure qualifications. For organic social, confirm that the caption, first comment and image remain coherent if viewed separately or shared out of context. A disclosure hidden in a landing page does little for an impression formed in a feed. Sample live ads every month and prioritise the highest-spend or highest-risk variants.

Baker McKenzie highlights that EU rules are more specific for some realistic synthetic images, audio and video, and that ordinary technical edits are not necessarily treated like deepfakes. That distinction will matter for UK firms placing ads across borders. Keep a market column in the register and recheck placements when a campaign expands. A local supplier advertising only in the UK has a different review task from a brand using the same creative across the EU. The practical goal is not to make marketers into lawyers. It is to make the evidence and escalation point visible before a customer, regulator or competitor has to ask for it. Start with your five highest-spend AI-assisted creatives, review the implied claims, and fix any gap this week.

Frequently Asked Questions

Must every AI-generated UK advert be labelled?

No. There is no blanket UK labelling requirement. Assess whether omitting AI use would materially mislead consumers.

Does an AI label make a misleading product image acceptable?

No. Correct or remove the misleading depiction. A label does not substantiate an inaccurate claim.

Who should approve an AI-assisted advert?

Give the campaign owner, product or service owner and final consumer-impression reviewer explicit responsibilities, even if a small team combines roles.

Are generated backgrounds lower risk than generated testimonials?

Usually, because decorative backgrounds may not affect a product claim. A synthetic testimonial can imply a real experience and needs much stronger review.

What evidence should we keep?

Keep the final asset, claim register, supporting evidence, disclosure decision, approver and approval date.

When should a live advert be reviewed again?

Reopen approval when prices, features, evidence, testimonials, markets or applicable rules change. Also sample high-risk live variants regularly.

Do UK rules cover adverts aimed at EU customers?

UK rules remain relevant, but EU AI Act transparency obligations may also apply to certain realistic synthetic content. Review each placement and market.

Will a review gate slow down every campaign?

It need not. Preapprove reusable claims and creative boundaries, then escalate only variants that change the consumer impression.