A generative model optimises for plausibility. Truth sits outside its remit. It has no mechanism for knowing that the award badge it has drawn on your packaging is fictional, because it is drawing what award badges look like.
In visual work the expensive hallucinations are the legally loaded ones: invented certifications, invented ingredient claims, invented awards. That is why the claim gate exists, and why it is separate from the label gate. A label can be checked by reading. A claim has to be checked against a document.
QUESTIONS PEOPLE ASK
How do you stop AI hallucinations in marketing content?
Constrain and verify. Give the model the source text and never ask it to recall, and run a claim gate where every factual assertion on screen has to trace to a document before it renders.