We tested it ourselves: ask an AI image generator to show a business leader and, as a rule, you get a white, male-presenting person. Ask for a portrait wearing "size 54" and the model often makes the person slimmer anyway. AI images take minutes to produce - but who they show is not decided by chance. It is decided by a built-in "normal": young, slim, white, without visible disability.
That is not a hunch - it has been measured rather thoroughly. And it affects every brand publishing AI images, which will soon be nearly all of them.
Where does the built-in "normal" come from?
AI image models learn from vast collections of photos and captions scraped from the web. They show the world they were fed - including every distortion already baked into it: stock-photo aesthetics, advertising imagery, decades of narrow visual language. The result is not a portrait of society, but a portrait of what happened to be photographed, uploaded and tagged.
The crucial point: the models do not merely pass that distortion on, they turn up the volume. A Bloomberg analysis of more than 5,000 generated images showed that Stable Diffusion draws gender and ethnicity stereotypes more sharply than reality: well-paid professions such as CEO or lawyer were rendered almost exclusively as light-skinned, male-presenting faces - considerably more one-sided than actual labour statistics suggest.
How large is the distortion really?
Body image makes it even clearer. A study in npj Digital Medicine analysed 9,060 images from four major generators (Adobe Firefly, Bing Image Generator, Meta Imagine, Midjourney): 88 to 96 per cent of the people depicted were of normal weight - in the real reference population it is 63 per cent. People in larger bodies appeared in only 3 to 5 per cent of images, against 32 per cent in reality. A group making up a third of the population is close to invisible in AI imagery.
Disability follows a similarly narrow pattern: a CHI study with the telling title "They only care to show us the wheelchair" and an analysis of popular image models show that disability is rendered almost exclusively as a wheelchair - and the people shown tend to look old and sad. The diversity of real lives, from invisible disabilities to prosthetics in everyday use, simply does not appear unless you spell it out in the prompt.
Why does this affect your brand?
Advertising has always shaped beauty ideals. The difference: it used to take a shoot, a briefing, a deliberate casting decision. Today a prompt produces a good-looking image in seconds - with the distorted normal included at no extra charge. Publish the first acceptable version unchecked, and you make that normal a little more standard - the standard real people compare themselves against.
Then there is the commercial side: your customers are more diverse than the average output of an image model. If your website and feeds show only one kind of person, many of your actual customers simply will not recognise themselves. It is the same logic as your brand voice: what your brand shows tells people what it stands for - whether you decided it consciously or not.
Why is a well-meant prompt not enough?
The obvious objection: "Then I shall simply write the diversity into the prompt." That is the right start - but it only works partially. In our own tests, even with an explicit instruction such as "size 54", the model often rendered the person slimmer than the prompt demanded. The model pulls its output back towards the normal it has learnt, and it does so quietly: you do not get a refusal, you get an image that takes your prompt only half seriously. Exactly this AI bias is what makes diversity manual work.
It gets more reliable with proper prompt engineering: concrete, respectful descriptions instead of vague labels, reference images where the tool supports them, generating several variants and choosing deliberately - plus a critical look before anything goes live. AI delivers the draft; the judgement stays human. The same principle that has proven itself for text as human in the loop applies to images all the more.
Three levers for images that show people as they are
Make diversity a prompt rule, not an exception. Put representation into your style guide: which people does your brand show, which terms do you use in prompts, what never goes out? That way it does not depend on chance or on whoever happens to be prompting.
Check before publishing with one simple question. Does this image show people as they really are - or just the model's built-in normal? If you lay ten generated images side by side and they all show the same kind of person, that is your signal.
Compare the output with your actual customers. Look at your real customer data, appointments, projects: who do you actually serve? If your brand's imagery systematically diverges from that, you are giving away recognition - and with it, trust.
AI images are here to stay - all the more reason not to let them decide what people look like. If you want to know how to produce AI content that represents your brand and your customers honestly, just drop us a line. 💌