Governance, Ethics & Legal Practical insights
Substantiate AI campaign claims before testing them
Build evidence into AI campaign production by checking numerical and comparative claims before variants enter experiments, localization or paid delivery.
An experiment can show which message gets more clicks. It cannot prove that the message is true. When AI produces dozens of persuasive variants, require evidence for the claims before those variants reach an audience. Otherwise the optimization process may reward the most exaggerated version.
Break the sentence into claims
Take an illustrative draft: ‘Our AI workflow doubles productivity and removes reporting errors.’ This contains at least two factual promises, each requiring support. Ask what was measured, for whom, against which baseline and under what conditions. If the team has no evidence for either promise, changing the wording to sound less technical does not solve the problem.
Attach the evidence and its limits
Maintain a claim record with source, method, date, owner and approved wording. Distinguish a vendor announcement, an internal controlled test and one customer’s experience. NIST’s generative AI risk profile identifies confabulation and information-integrity concerns. [1] A claim record provides a practical way to stop plausible invented evidence from entering the production chain.
Constrain the variant task
Give the writing tool the approved claim and its necessary qualifications. Ask for alternative openings or calls to action while preserving the factual scope. Compare the variants with the source record, paying attention to omitted conditions and broader audiences. Include localized versions: ‘may reduce editing time’ should not become an unconditional promise in Dutch.
Set a publication gate
Assign a reviewer who can reject an unsupported claim before testing begins. Route regulated or jurisdiction-specific advertising questions to the appropriate specialist. After publication, reopen the record when the product, comparison or evidence changes. The workflow still leaves room for creative experimentation, but the experiment happens within a defensible factual boundary instead of using audience response to decide what the brand is entitled to claim.
Sources and evidence
Sources checked on 4 October 2026. Proposed workflows and hypothetical examples are editorial analysis.
From insight to practice