AI Images Match Stock Photos When Users Can't Tell the Difference

A Nielsen Norman Group study with 77 U.S. adults found that AI-generated hero images triggered no measurable trust penalty when viewers couldn't identify them as synthetic. Participants rated identical fictional consulting-firm pages nearly the same whether the hero shot came from an AI generator or a stock catalog.

"When users didn't know whether an image was AI-generated, the images we tested did not create any perception penalty compared to stock photos," the researchers wrote.

The headline number is six. That's how many page variants the research team built. Three carried AI imagery, three carried real photography. Everything else — copy, layout, color, type — stayed frozen.

How was the comparison built?

Each participant spent 10 seconds on each page. That timing matters. Ten seconds mirrors a typical first scan of a landing page — long enough to register the hero image, short enough that nobody lingers over captions. After viewing, participants rated the firm on trustworthiness, professionalism, and authenticity. They also answered an open prompt: "What other factors influenced your impression of this company?"

The twist is what they weren't told. Nobody mentioned AI. Participants believed they were evaluating standard website designs. The researchers wanted gut reactions, not filtered ones.

What does the finding actually mean?

When provenance stays hidden, AI imagery performs on par with stock photography. The study used mixed-effects analysis to handle variation across participants and individual images. Pairwise t-tests compared each AI image against each stock image. A Bonferroni correction kept the multiple comparisons honest.

The practical takeaway sits inside a question most design teams already ask: if we swap stock photos for AI, will users punish the brand? The data says probably not — so long as nobody points at the swap. Without a label, viewers can't build the suspicion that prompts the penalty.

One caveat threads through the methodology. The study measured perception, not behavior. A 10-second rating captures first impressions, not purchase intent or sign-ups. Designers reading this as a green light for AI imagery at scale should treat it as one input among several.

Why did the researchers pick a consulting firm?

The fictional firm gave the researchers a neutral canvas. Each participant evaluated all six variants in randomized order, so every AI image got judged against every real image inside the same viewer's head. That's a stronger design than between-subjects comparison, where one group sees only AI and another only real.

Numbers worth holding

• 77 participants, recruited from the U.S. adult population
• 6 webpage variants, identical except for hero image
• 3 AI images, 3 real stock images
• 10-second viewing window per page
• 3 trust dimensions rated: trustworthiness, professionalism, authenticity

Roughly one in every 12 evaluations came from a viewer who never saw the same image twice. That randomization killed order effects. It also meant each comparison carried about 462 data points — 77 viewers times 6 variants.

What stays unresolved?

Disclosure is the variable the study didn't test. The researchers deliberately excluded any condition where participants knew about the AI element. NN/g's test nails the no-label case. The labeled case remains a separate study waiting to happen.

The forward question for design leaders: how do users respond once you tell them? The methodology holds that door open. Until then, the data covers silent swaps, not flagged ones.