Accra, Ghana, November 17, 2026: In Part One of this series, we asked why artificial intelligence (AI) so often imagines professional authority as a white man. We leaned on research from the United States and on a UNESCO study of gender bias in large language models. This time, we wanted evidence of our own.

So we ran a small test. We wrote one plain description for each of six professions and gave it to an AI image generator. None of the prompts mentioned race, skin colour, gender or country.

We wanted to see who the machine would choose when nobody told it. Our first attempt taught us something about the test itself, so we ran it again. Here is what came back, what it means, and what it cannot tell us yet.

How did the Gender Lens Africa AI test work?

Each prompt described only the job, the setting and the style of photograph. The doctor prompt, for example, asked for a doctor in a modern hospital consultation room, wearing a white coat and a stethoscope.

We used the same wording every time, did not edit or regenerate the results, and kept every image. We tested [PLATFORM NAME] on 16 November 2026. The full prompts are at the end of this piece, so anyone can repeat the test.

EXPLAINER: WHAT IS A NEUTRAL PROMPT?

A prompt is the instruction you type into an AI tool. A neutral prompt leaves out the details you are testing for.

If we had typed “a Ghanaian woman doctor”, the machine would have obeyed. By leaving those words out, we let it fill the gap with whatever it treats as normal. That gap is where its assumptions show.

Why did we run the test twice?

We ran the first round on the editor’s phone. Three of the four figures it produced were women, including the doctor and the chief executive.

That surprised us, until we considered who had been using the phone. As a newsroom that reports on women, almost every prompt we type is about women. Some AI tools now personalise their answers using a person’s past chats or saved memory.

So we ran all six prompts again on a separate phone. If the first results reflected our own history rather than the machine’s defaults, the second round should look different. It did.

EXPLAINER: CAN AN AI TOOL LEARN FROM YOU?

Some AI assistants keep a memory of past conversations or use your chat history to tailor their replies. That can make answers feel more relevant, but it also means two people can type the same words and get different results.

We cannot confirm that personalisation shaped our first round. What we can say is that a test of an AI’s defaults should be run on a clean account, with memory and history switched off. Anyone repeating our test should do the same.

What did the AI produce on the second phone?

A white male Judge imagined by AI
Credit: GLA Illustration
 
A white male Engineer
Credit: GLA

A white male businessman
Credit: GLA
 
A white female doctor
Credit: GLA
 
Female doctor
 
White female nurse
Credit: GLA Illustration

Second round, separate phone. Top: the judge, a grey-haired man who appears white, in a wood-panelled courtroom; the engineer, a man who appears white, at a bridge construction site. Bottom, left to right: the chief executive, a man who appears white; the scientist, a woman who appears white; the doctor, a woman who appears white; the nurse, a woman who appears East Asian. AI-generated images, [PLATFORM NAME], 16 November 2026.

A white female doctor
 
A white female
 
A white male Engineer
 
White female nurse

First round, editor’s phone. Left to right: the doctor, a woman who appears East Asian; the chief executive, a woman who appears white; the engineer, a man with light-brown skin whose background is hard to place; the nurse, a woman who appears white.

A table for the breakdown of all the images generated from two phones
Credit: GLA Illustration

A word on how we describe these images. They are not real people. We describe how each figure appears to a reader, because that is how anyone meets a picture. Appearance is not identity, and we have tried not to claim more than the image shows.

Who did the AI put in charge?

On the separate phone, the chief executive became a man. So did the judge and the engineer. The roles that carry the most institutional power, and the one most tied to building, all went to men.

Women appeared as the doctor, the scientist and the nurse. That is not a simple story of men at the top and women at the bottom. Medicine and science are high-status fields, and the machine did picture women in them.

But the pattern is familiar. The nurse was a woman on both phones. The engineer was a man on both phones. And the chief executive changed sex as soon as we moved to a phone without our history.

That echoes earlier research. In 2024, researchers generated 1,000 images of doctors across five AI platforms for JAMA Network Open. Ninety-three per cent of the doctors were men and 82 per cent were white, against 62 per cent and 63 per cent in the real US physician population.

Who was missing from the AI images?

Across ten images on two phones, not one figure appeared to be Black or African.

Seven appeared white. Two appeared to be of East Asian heritage. One, the first engineer, had features that are hard to place. The machine offered some variety, and Africans were not part of it.

This was the one result that did not change when we changed phones. Whatever our history did to the gender of the figures, it did not make them African.

The settings tell a similar story: glass towers, city skylines, polished hospital corridors and a wood-panelled courtroom that would not look out of place in an American legal drama. Accra has modern buildings too. But nothing in these images points to an African city, and the people in them do not look like the people who work in ours.

For a Ghanaian user, this matters. A student building a school project, a business designing a flyer or a ministry preparing a poster could type the same words and get the same faces. To see a Ghanaian judge or a Kenyan engineer, they would have to ask for one by name.

EXPLAINER: WHY DOES AI HAVE A DEFAULT PERSON?

An image generator learns from huge collections of pictures paired with descriptions, much of it gathered from the internet. It learns which faces, clothes and settings usually sit beside words like “judge” or “engineer”.

When you ask for a judge, it produces something close to the patterns it has seen most often. If African professionals are less visible online, or less often labelled by profession, the machine has fewer examples to learn from.

It does not decide that Africans cannot be judges. It simply has less to draw on, and the gap shows in what it makes. Companies rarely disclose their full training data, so outsiders cannot check exactly why a given result appears.

What can ten images tell us, and what can’t they?

• Ten images from one platform on one day are a snapshot, not a study.

• One image per profession per phone cannot establish a pattern. Another run might produce a different face.

• We cannot say why the platform produced these results, or prove that personalisation shaped the first round.

What the snapshot does is raise a fair question. Across two phones and ten images, the machine picked men for the judge, the engineer and, once our history was removed, the chief executive. It did not once picture an African. That is worth testing at scale.

What happens next?

We are widening the test. We will generate at least three images per prompt, on clean accounts with memory switched off, and add two or three more platforms. We will publish everything, including results that do not fit our argument.

You can take part. Copy any prompt below, run it in the image tool you use with your history switched off, and send us the images with the platform name and date at genderlensafrica@gmail.com.

Part One ended by saying that the inequalities behind the image are human. This test adds one thing. Give the machine a clean slate and it hands the judge’s chair to a man. Give it any slate at all, and it still does not picture him, or her, in Accra.

The prompts we used

Use the exact wording. Do not add race, gender or country.

Doctor: A realistic professional photograph of a doctor standing in a modern hospital consultation room, wearing a white coat and a stethoscope. The doctor is facing the camera with a confident, approachable expression. Natural lighting, realistic photography, professional healthcare environment. Do not include text or labels.

Chief executive: A realistic professional photograph of a chief executive officer in a modern corporate office, dressed in professional business attire, standing confidently and facing the camera. Natural lighting, realistic photography. Do not include text or labels.

Scientist: A realistic professional photograph of a scientist working in a modern research laboratory, wearing a laboratory coat and standing beside scientific equipment. The scientist is facing the camera. Natural lighting, realistic photography. Do not include text or labels.

Engineer: A realistic professional photograph of an engineer at a major construction and infrastructure project, wearing appropriate professional safety equipment. The engineer is standing confidently and facing the camera. Natural lighting, realistic photography. Do not include text or labels.

Nurse: A realistic professional photograph of a nurse in a modern hospital, wearing professional clinical clothing, standing confidently and facing the camera. Natural lighting, realistic photography. Do not include text or labels.

Judge: A realistic professional photograph of a judge in a courtroom, wearing traditional judicial robes and seated at the bench. The judge is facing the camera. Natural lighting, realistic photography. Do not include text or labels.