Advertising AI-generated performance claims: proof and transparency

Under section 2(1) UWG, a commercial practice is misleading if it contains incorrect information or is otherwise capable of deceiving market participants about a relevant point and causing a business decision that would not otherwise have been made. For performance claims, the essential characteristics of the product and the characteristics of tests or investigations are central.

The provision asks what the claim says and how it affects the audience. A language model is not a legal justification and cannot replace review. A sentence that sounds technically plausible may still be wrong if it describes a non-existent feature, an unsupported measurement or a conclusion that goes beyond the available data.

The complete presentation must also be assessed. The headline, image, figures, qualifications, link destination and audience selection may create an impression different from an isolated sentence. The claim should therefore be reviewed in the form in which it is actually distributed.

A performance claim can be assessed only against a clear benchmark. Record which performance is meant, under which conditions it should be achieved, which period applies and which product version was tested. Statements such as “twice as fast” or “suitable for every use” require a concrete factual basis.

The test method, setup and result must support the wording used. A test using one configuration or a small sample does not automatically support a general promise for every user. An average result should not be presented as an outcome every customer will achieve.

The advertising claims self check helps organise the wording, audience, qualifications and evidence. For AI-generated text, also record which data or instructions produced the claim and who approved it from a subject-matter perspective.

Safety promises often create a particularly strong expectation. Wording such as “safe”, “risk-free” or “reliably protects” may mislead about product characteristics and risks if it does not reflect the intended use and known limits.

The evidence review must cover the actual scope of use. This includes the product version, testing or standards basis, tested conditions, relevant exceptions and the period for which the result applies. Technical approval for one use does not automatically support an unrestricted promise for every situation.

Automatically generated variants must not strengthen the claim through audience targeting or shortening. A short advertising sentence can remove a qualification from the source text. Review every distributed variant and preserve the related approval.

Claims about sustainability, environmental impact or climate action are factual claims when they communicate a verifiable characteristic of the product or business. Terms such as “climate friendly”, “resource efficient” or “sustainable” therefore need a clear reference point.

The business should record what the claim covers: the product, packaging, production, transport or the full life cycle. The data source, calculation, period and included factors must match the reach of the statement. One improvement does not justify a general claim about the entire environmental impact.

The article on green claims and environmental advertising develops the review of environmental statements further. The same rule applies to AI-generated wording: the system may describe existing data, but it cannot replace missing data.

In practice, AI copy often comes from a platform, an agency process or an internal system. The legal review should therefore show who described the product, supplied the data, approved the wording and published the advertisement.

A contract with a provider or agency does not by itself establish that a claim is factually correct. The business should define requirements for data quality, approval, change records and correction. For external models, it is also important to know which source material was processed and whether the result refers to the correct product version.

The roles do not have to be performed by one person. They should be documented well enough to reconstruct the path from the source to the published claim. A purely technical log without the final advertising wording will usually not be enough.

A reliable approval process starts with a claim brief. It records the product, audience, intended statement, conditions, evidence and permitted qualifications. The AI draft is then checked against the records and read again in the actual advertisement or landing page.

Save the model or prompt version, input data, draft, subject-matter review, approved wording and publication time. For dynamic advertising, also record audience rules, distribution variants and the change logic.

The article on influencer advertising disclosure shows why commercial context and presentation must be reviewed together. The same applies to AI campaigns, where content, channel and approval are connected.

If a business discovers an incorrect or overstated claim, it should first preserve the affected advertisements and their distribution. The original wording, time and audiences must not disappear through a simple overwrite.

The specific deviation should then be identified. A correction for one term, one product version or one missing qualification may differ from a case where the claim has no supporting test basis. Review automatically generated variants, translations, landing pages and reused text modules as well.

A new version needs its own approval and a clear change record. Whether customers, platforms or business partners also need information depends on the reach, content and circumstances. The article on customer reviews and evidence illustrates why the earlier visible version should remain available after a correction.