Your Daily Dose of Ad World Buzz
Latest
The Ai Shift

Who Fact-Checks A Generated Ad?

Ad Tribe! Bureau3 min read
Small characters with magenta lenses inspecting a scatter of ad cards, one marked with a coral warning shape, illustrating claim checks

A generated ad can be written, illustrated and versioned into fifty sizes before anyone has checked whether a single claim in it is true. The tools are fluent, fast and confident, and they are entirely comfortable inventing a specific number. Which raises a question most agency workflows have not answered: who fact-checks a generated ad?

The honest answer in many teams is whoever happens to look at it last.

Where the errors in a generated ad come from

Models produce what is plausible, not what is verified. Ask for punchy copy for a pressure cooker and you may get a saving in cooking time the client has never tested. Ask for a festive offer line and the model may supply a discount, a date or a delivery promise. Ask for a product shot and the image may show the wrong variant, a bigger portion than the pack contains or a feature the model has borrowed from a competitor.

None of these look like mistakes. They look like copy. That is the problem.

A model does not know which of its sentences are claims. Someone has to.

Why the old checks miss them

Agency review was designed around a copywriter working from a brief and a product document. Claims came from the client, the writer shaped them and the client's legal or regulatory team cleared the master version. It worked because the claims had a known source.

Generation breaks that chain in two places. The claims may no longer come from the brief, and the volume means reviewers see a sample. Legal approves the hero film and the main banner. Nobody reads variant 147, the regional resize, or the line a tool quietly reworded to fit a smaller format.

Who should own the check on a generated ad

Responsibility does not move with the tool. The ASCI Code requires claims of objectively ascertainable fact to be capable of substantiation, and it applies to the advertisement, not to the method. If a generated line says a product does something, the advertiser has to be able to prove it.

In practice that means four habits. A claims library: an approved list of facts, prices, comparisons and benefits that the tools draw from, and nothing outside it. A named checker for each job, who is not the person generating, because the person who prompted is the one least likely to notice what the model added. Full checks on any variant carrying a number, a price, a comparison, or a health or environmental claim, rather than a sample. And image review against the real product, pack and variant.

Keep a record of what was checked and by whom. When a complaint arrives, the question will be what the brand knew and when. As we wrote in when the bottleneck moves to legal review, the speed gained in generation tends to pile up at the checking stage. That is not a flaw in the process. It is where the work now is.

More from The AI Shift.

Follow Ad Tribe