signal

Visitors trusted AI hero images as much as real stock photos

A blind study of 77 U.S. adults found AI-generated hero images scored as well as real stock photos on trust and professionalism, and slightly higher on authenticity.

Go Voltic · published

Put an AI-generated photo on your home page, the theory goes, and something in a visitor's gut notices even when nothing is labeled. A new study from the Nielsen Norman Group tested that directly, and the fear did not hold up. Told nothing about which images were AI-generated and which were real stock photography, 77 U.S. adults rated the two about the same on trust and professionalism, and rated the AI images slightly higher on authenticity.

A blind test of 77 adults found no trust penalty for AI images

Nielsen Norman Group built six versions of a fictional consulting firm's homepage that were identical except for one hero image. Three versions used generative AI, which a separate NN/g glossary entry defines as "AI systems that generate new content (text, images, code, audio) from patterns learned during training, rather than only retrieving or classifying what already exists." Three used real stock photography.

Researchers recruited 77 participants from the general U.S. adult population, showed each person one version of the page for ten seconds, and had them rate the company on trust, professionalism, and authenticity on a scale of 1 to 7. Nobody was told which images were AI-generated. The team analyzed results with mixed-effects models and pairwise t-tests carrying a Bonferroni correction, the standard way to guard against a false positive when three separate outcomes are tested at once, and coded participants' open-ended comments about what shaped their impressions.

The only statistically significant gap ran in AI's favor, and it was small

Ratings came out slightly higher for the AI images on all three measures. Trust and professionalism each moved by 0.2 points on the 7-point scale, a gap too small to separate from chance. Authenticity moved by 0.4 points and did clear statistical significance, the only one of the three that did, though the researchers themselves describe the size of that effect as very small.

Read plainly: a visitor who did not know which image was which did not rate the AI photo as less trustworthy, less professional, or less authentic than a real one. If anything, the numbers ran the other direction, by an amount too small to build a strategy on.

What actually moved people was who was in the picture

The open-ended comments point at a different variable than origin. Participants who saw images of people working together, reading as diverse and collaborative, wrote things like "I like that they are all working together and that makes me feel like this company values teamwork" and "I like this picture more than the others. It shows the company has inclusion and a good teamwork system." A participant who saw a less diverse image wrote, "I don't notice the diversity as much."

The content of the image, not its origin, is what people reacted to. A generic-looking AI photo and a generic-looking stock photo drew the same flat response. An image that read as a real, collaborative workplace drew a better one, whichever way it was made.

One study, one company, one context, and the researchers say so themselves

NN/g's own write-up hedges its result directly: the results "may not generalize to other types of imagery, industries, or contexts." Every image in the test sat on a consulting firm's homepage, a setting where a hero photo sets a mood rather than carries the whole pitch. A product photo on an ecommerce listing, a "meet the team" page naming real employees, or a testimonial photo attached to a real customer's name make a different kind of claim: each implies the picture shows something specific and true, in a way a generic hero shot does not. None of those were tested here.

What to check before you publish an AI image

  1. Look for the tells that actually read as fake on close inspection: warped hands, extra or missing fingers, text that dissolves into gibberish, shadows falling in two directions, or a background pattern that repeats.
  2. Reserve AI images for mood and illustration, not for claims. A hero image implying "our office" or "our team" when neither exists is a different problem from an unlabeled stock photo, and no study is needed to know it is dishonest.
  3. Check the generator's commercial-use and indemnification terms before putting an image on a page that sells something. These vary by provider and by pricing tier, and they change without much notice.
  4. Decide your disclosure position before someone asks, not after. This study tested images nobody suspected were AI-generated. It says nothing about the reaction once a visitor asks directly.

None of this requires guessing. Look at an image the way NN/g's participants did: for ten seconds, for what it shows, not for how it was made. A photo that reads as real people doing real work scored the same to visitors whether it came from a camera or a prompt. A photo that reads as generic scored the same too, however it was produced.

Sources

  1. AI-Generated Images Can Perform as Well as Stock Photography. Nielsen Norman Group, read 2026-08-22
  2. Artificial Intelligence Glossary. Nielsen Norman Group, read 2026-08-22

Run a free Instant Scan on your own site · More from the Newsroom