Content Benefit

Content Strategy | UX Writing | Content Design | Globalization | Communications

AI images or stock photos?

AI images or stock photos?

What the Nielsen Norman Group study actually measured

On 21 August 2026 Rachel Banawa published a small but treacherous experiment for the Nielsen Norman Group. Seventy-seven American adults visited six versions of the homepage of an invented consulting firm, all identical except for the opening picture: three built with ChatGPT Images 2.0, three licensed from Unsplash+ and iStock. Ten seconds of viewing, then a score from 1 to 7 on trustworthiness, professionalism and authenticity. None of the participants had been told that the survey concerned automated generation.

The verdict surprises anyone who expected a penalty. Pages carrying synthetic visuals collected 0.2 points more on trustworthiness and professionalism, a gap without statistical significance, and 0.4 points more on authenticity, a difference the researchers themselves called “very small”. Translated: when people browsing ignore where a picture comes from, the advantage of traditional photography evaporates.

Suspicion, rather than technique, is what flips the outcome

There is one detail that deserves to be highlighted, though. Whenever participants suspected an artificial origin, they lowered their rating of the site. Even in front of perfectly genuine shots, bought and paid for. Reputational damage therefore springs not from the generated pixel but from the doubt creeping into whoever looks.

That shifts the problem away from technical quality and towards trust. And trust, for anyone working in content, stays hard to rebuild once lost. On the question of attention thresholds I had already written in Attention please, if you will.

I admit I cannot tell a synthetic photo from a real one

Here comes the confession of a middle-aged lady who has crossed twenty-five years of localization and still remembers fax machines. If an image or a video generated by AI is well made, I don’t notice. Not at all. I inspect the hands, count the fingers, check the reflections in the eyeglasses, and I get it wrong anyway. I feel decidedly boomer every time someone exclaims “but it’s obvious!” in front of a scene that looked to me like reportage.

Consolation arrives from the data, and it counts for something.

The figures say almost nobody spots a generated picture

A 2025 study published in Frontiers in Artificial Intelligence by Högemann, Betke and Thomas presented 104 German-speaking people with fifty images, half authentic and half produced by five text-to-image models. Average accuracy in identifying the synthetic material stopped at 63.7%.

The interesting element concerns the variance across models. Creations by Kolors were unmasked in 86.73% of cases, those by Playground v2 in 82.69%. With FLUX.1-dev the share collapsed to 29.04%, below chance level: participants erred more often than random guessing would have produced. Worse still, confidence in one’s own answer grew while correctness declined.

Age shows a negative correlation with accuracy (odds ratio 0.88 per age bracket). So yes, my sense of generational inadequacy rests on an empirical basis. At least I have company.

Hyperrealism: artificial faces look more human than human ones

The most disorienting phenomenon carries the signature of Elizabeth Miller and colleagues, who documented it in Psychological Science in 2023 under the label AI hyperrealism. Across 124 adults, algorithmically generated White faces were classified as human in 65.9% of cases, against the 51.1% obtained by the countenances of actual people.

Machines, in short, turn out visages more convincing than the originals. A further 59% of participants displayed negative insight into their own mistakes: they slipped up exactly where they felt most certain, a textbook Dunning-Kruger effect applied to visual perception.

One last unpleasant note: hyperrealism appeared only on Caucasian portraits, because the training dataset consisted of roughly 69% White individuals. On other ethnicities performance dropped to chance level. Corpus imbalance thus spills over into perceived credibility, a matter that intertwines with what I recounted in Accessibility by default.

Fifteen billion pictures in a year and an award handed back

Grasping the scale of the phenomenon requires an order of magnitude. According to the count kept by Everypixel Journal, in about twelve months generative systems churned out fifteen billion visuals, a quantity comparable to what photography produced across its first hundred and fifty years.

And then there is the anecdote every content strategist should bear in mind. In 2023 the German artist Boris Eldagsen won the Creative category of the Sony World Photography Awards with “Pseudomnesia: The Electrician”, a black-and-white portrait of two women. He turned down the honour, explaining that the work had come out of a generative model and that he had entered it as a test, to verify whether juries were ready to identify the phenomenon. They were not.

What the European AI Act requires from 2 August 2026

The regulatory picture is changing in these very weeks. Article 50 of the AI Act obliges providers of generative systems to mark outputs in a machine-readable format, detectable as artificially produced material, across text, imagery, audio and video. Transparency provisions apply from 2 August 2026, with a window until 2 December 2026 for systems already placed on the market before that date.

A partial derogation exists for evidently artistic, creative, satirical or fictional works: there the disclosure must be made in a manner that does not spoil enjoyment. The judicial front is moving too: in November 2025 the English High Court dismissed the secondary copyright infringement claim brought by Getty Images against Stability AI, the first British ruling in a dispute of this kind.

The decorative photos nobody really looks at

There is an older strand of research worth putting forward again. Back in 2010 Jakob Nielsen had shown through eyetracking that purely ornamental shots get skipped by the gaze. On a Yale School of Management page, generic student pictures received no ocular fixations despite occupying abundant space.

Conversely, genuine portraits of FreshBooks employees captured 10% more time than the textual biographies while taking up an enormously smaller surface. Nielsen’s summary holds up: people observe images that carry information useful to their task and disregard the rest.

If a visual gets skipped, the question of who made it loses much of its weight. I like to recall the parallel with the placeholder copy described in What is “Lorem ipsum” and why does it matter: filling a space does not amount to communicating.

Why I chose minimalist illustrations for the blog

My editorial decision grows out of all this. On the blog I publish no realistic photographs, neither authentic nor synthetic. I use spare drawings, few lines, a warm recurring palette. The reason is simple: I have no wish to imitate reality, I want to evoke it.

A stylised sketch declares its own nature without needing labels. Nobody mistakes it for a document, nobody feels deceived, nobody wastes time hunting for the sixth finger (which has indeed shown up). The pact with readers stays clean, and chromatic consistency builds recognisability better than any licensed shot. On the daily struggles with these tools I had gathered a few reflections in 5 struggles when using AI.

Anyone working in fields where documentary credibility matters, journalism, healthcare, justice, would do well to keep authentic photography. For a professional blog reasoning about content and interfaces, graphic abstraction settles the dilemma at the root.

A checklist before publishing a generated visual

The Nielsen Norman Group suggests five checks, which I summarise with a few personal additions:

  • Purpose: does the picture support the message of the page, or does it plug a hole in the layout?
  • Representation: who appears, who is missing, how people are portrayed and in which visual hierarchy.
  • Plausibility: do gestures, glances and interactions feel believable, or do they betray the contrived posing of stock?
  • Errors: deformed characters, improbable hands, incoherent reflections, background details that melt away.
  • Context: final crop, size, placement within the page.

To these I would add a sixth item, contractual in nature: knowing the model’s training data and the commercial usage rights before adopting it in production. The perceptual result may well prove favourable, while the legal and ethical exposure remains entirely open to assessment.

Related sources

Chiara Bonifazi