Generated imagery in production interfaces

Nielsen Norman Group tested AI-generated images against stock photography in August 2026 and found that users did not perceive them as worse when they did not know the source.

That is a narrower finding than it first appears, and the narrowness is the useful part.

Reading what the study actually says

The comparison was against stock photography, not against commissioned work. Stock is already a compromise: generic, unspecific to the product, and chosen from what exists rather than from what the page needs. A generated image matching that standard is a statement about the standard.

The condition that users did not know the source also matters, because disclosure changes the response. A brand that labels generated imagery is running a different experiment from the one the study ran.

Deciding where it is a reasonable substitute

The places generated imagery substitutes cleanly are the places stock was already doing the job: illustrative headers, category tiles, empty states and anything whose function is to give the eye somewhere to rest.

The places it does not substitute are the places where the picture is evidence. A screenshot of the product, a photograph of a real customer, a real workshop, a real installation. Those carry information rather than atmosphere, and a generated version of them is not a cheaper alternative, it is a different and false claim.

The places generated imagery substitutes cleanly are the places stock was already doing the job.

BrilliantUX editorial principle

What it changes about the budget

The saving is real and it lands in a specific place: the long tail of secondary imagery that a marketing site or a product needs and that nobody was ever going to commission.

The budget that should not move is the one for the small number of images that are doing actual work. A product page benefits more from three commissioned photographs of the real thing than from thirty generated ones, and the study does not say otherwise.