Accra, Ghana, November 17, 2026: Imagine asking an artificial intelligence (AI) image generator to create a doctor. Then ask it for a chief executive, a university professor, a scientist or an engineer.

Look carefully at the faces. Who appears? What is their skin colour? Are they men or women? What changes when you swap surgeon for nurse, company director for domestic worker, or professor for cleaner?

The results will vary. The wording of a prompt, the tool and the images it learned from all shape what appears. But when the same faces keep emerging in positions of authority, one question is worth asking: why does the machine seem to have a default human?

And why, so often, does that human look like a white man?

This is not only a question about pictures. It is about how technology learns to represent the world, whose images go into that process, and what happens when old inequalities are reproduced at a scale that makes them look ordinary.

What does the research say about AI images of doctors?

In 2024, researchers writing in JAMA Network Open generated 1,000 images of physicians using five popular AI image generators. They used four prompts, including “face of a doctor in the United States”, and compared the results with data on real US doctors.

Eighty-two per cent of the AI doctors were white and 93 per cent were men. Among real US physicians, the figures were 63 per cent and 62 per cent.

The gaps went further. Three of the five platforms generated no Latino doctors, two generated no Asian doctors, and one produced no women doctors at all. The researchers concluded that the images disproportionately showed white men.

That study has limits. It tested five platforms with particular prompts in 2024, and it compared the results with US doctors. It does not prove that every AI tool behaves the same way today, or tell us how AI pictures Ghanaian professionals.

But it shows something important. Asked to picture a doctor, these systems produced a far narrower picture than the profession itself. That is where the problem begins.

Why does AI copy the world that made it?

AI is often discussed as if it were an independent mind, observing humanity and returning objective answers. In reality, people build these systems, people choose and assemble the data they learn from, and people decide how they should respond.

An image generator has no personal belief about who should become a doctor. It learns patterns from its training material: which faces, clothes and settings tend to sit beside which words. It then uses those patterns to produce something that fits a prompt.

If that material repeatedly links certain professions with certain appearances, those links shape what the machine makes. No developer has to instruct it to leave out Black people or women. Training data and design choices differ between systems, and companies rarely disclose enough for outsiders to check.

The point is simple. A system can reproduce inequality without anyone telling it to.

Is the bias only in images?

No. In March 2024, a UNESCO study of three large language models, the systems behind popular AI chat tools, found that women were linked far more often with words such as “home”, “family” and “children”. Men were linked with “business”, “executive”, “salary” and “career”.

In stories written by one of the models, men received varied, high-status jobs such as engineer, teacher and doctor. Women were more often cast as domestic servants, cooks and prostitutes.

Think about what that means in practice. A woman asks an AI tool for a story about a successful entrepreneur. A student generates an image for a school project on scientists. A business designs material about leadership. A teacher asks for pictures of people at work.

If the technology keeps presenting men as leaders and women as carers, or white people as the face of professional authority, it is not just producing pictures and paragraphs. It is circulating ideas about who belongs where.

A single picture of a male doctor or a woman cooking is not discrimination. The concern arises when the same associations recur, when some groups are consistently absent from certain roles, and when those patterns are presented as natural.

Why does AI bias matter for Africa?

For African users, the question cannot stop at whether AI generates more Black faces. Which Black faces does it generate? Can it tell a Ghanaian woman from a Nigerian man, a Kenyan doctor from a South African engineer? Can it picture African professionals without reaching for a narrow set of visual clichés?

Representation is not about darkening the skin in an otherwise unchanged image. It is about whether people appear as individuals with different professions, ages, abilities, cultures and ambitions. It is about whether an African woman can appear as a surgeon, a chief executive, a judge or a scientist without being specifically requested every time.

Part of the answer lies in what these systems learn from. An analysis of two widely used public image datasets found that more than 32 per cent of the identifiable images came from the United States, with the United Kingdom also among the best represented. The researchers described a clear American and European bias, despite the creators’ efforts to encourage diversity. That study dates from 2017, but many systems were built on foundations like these.

This is not evidence that Africa lacks knowledge, talent or expertise. It raises a different question: whose work is visible online, whose images are labelled and collected, and whose knowledge is included in the systems being built.

The question is now a domestic one. On 24 April 2026, President John Dramani Mahama launched Ghana’s National Artificial Intelligence Strategy, a ten-year plan running to 2035. As government, schools, businesses, newsrooms and ordinary users adopt these tools, Ghana will be importing their assumptions along with their convenience.

We should not accept a technology’s limited imagination as an accurate picture of our societies.

Why should users not have to fix AI bias themselves?

It is tempting to dismiss all this as a minor flaw. A user can simply type “Black woman doctor” or “African engineer” and try again.

But that response puts the burden on the user to correct the system. Why should a woman have to specify that a doctor is female? Why should an African user have to keep adding instructions to see people who look like her own community? And what happens to people who do not know enough about AI to notice the bias at all?

No single AI image decides what anyone believes. The concern is that these systems are becoming part of the everyday material through which people learn, communicate, advertise, teach and decide.

When one image of authority appears often enough, it starts to feel familiar. Familiarity can be mistaken for normality. And normality quietly shapes expectations.

That is why developers cannot treat representation as a cosmetic fix for when users complain. They need to examine their training data, test their outputs across populations and professions, disclose more about how their systems are built, and take responsibility when their products reproduce harmful stereotypes.

Governments, researchers, educators and civil society must demand that accountability. Users, in turn, need the knowledge to question what AI produces rather than treating it as an objective answer.

Who gets to define the default human?

For generations, access to knowledge, technology and the power to shape global narratives has been unequal. Some societies have had far greater resources to produce, publish, preserve and distribute information. Others have been under-represented, or described largely through outsiders’ eyes.

AI is being built inside that unequal world. Being new does not make it immune to the old imbalances.

The issue is not whether white men should appear in AI images. They should, just as Black women, Asian men, disabled people and every other community should. The issue is why one group so often occupies the centre while everyone else must be deliberately summoned into view.

The white man did not become AI’s image of professional authority because a machine judged him the most qualified human being. These systems draw on patterns shaped by human decisions, unequal representation and the material made available to them.

If we do not question those patterns, we allow technology to give old prejudices a new appearance of authority. AI may generate the image. But the inequalities behind it are human.

COMING IN PART TWO: THE GENDER LENS AFRICA TEST

We wanted evidence from an African newsroom, not only from studies conducted elsewhere. So we gave an AI image generator six professions (doctor, chief executive, scientist, engineer, nurse and judge) with no mention of race, gender or country, and ran the test on two phones.

Across ten images, not one figure appeared Black or African. Part Two sets out what the machine produced, what changed between the phones, and what a small test can and cannot prove.