AI writing tools show gender bias, study finds

LUSAKA — A study of four leading artificial intelligence (AI) chatbots is the latest evidence that the tools carry a gender bias, giving men futures and women feelings across 112 generated stories.

The finding matters because these tools are moving fast into schools, offices and government services across Africa, and a bias baked into the writing they produce does not stay on the screen — it shapes how people are described, assessed and imagined. This story is part of Kwacha News’s continuing technology coverage.

The research, by Professor Sioux McKenna of Rhodes University and Nompilo Tshuma of Stellenbosch University, submitted 112 story prompts to ChatGPT, Claude, Grok and Copilot, and reported the results in Daily Maverick. The prompts used names signalling different racial, ethnic and gender identities.

What the study found

The pattern was consistent across all four tools. Male characters were written with ambitions, plans and concrete next steps; female characters were written around emotions, relationships and caregiving, with little sense of a future of their own.

The researchers also found a racial layer. Black female characters frequently appeared without fathers, with mothers carrying all the emotional labour — a shape the authors traced to historical deficit narratives, including apartheid-era policy documents and older sociological research that framed Black families through what they lacked.

Men were given plans, ambitions, and futures, while women were given emotions, relationships and caregiving roles.

— Sioux McKenna and Nompilo Tshuma, study authors, <a href="https://www.dailymaverick.co.za/article/2026-06-19-the-hidden-bias-in-ai-narratives-that-gives-men-futures-and-women-feelings/">via Daily Maverick</a>

Why it matters

The authors’ central point is that the output is not neutral. The models learn from vast amounts of existing text, and they reproduce the patterns in that text — including the assumptions about gender and race that the text already carried. The bias is inherited, then repeated at scale.

That repetition is the risk. As AI tools draft school reports, screen job applications and write public communications, a slanted view of who has ambition and who has feelings can harden into something that looks like objective assessment. Kwacha News has reported how AI is spreading in African farming even as foreign capital dominates the tools, and how the most powerful models are reaching the public.

For Zambia, the stakes are practical. The country is building digital public services and a data-centre ambition; Kwacha News explained what a data centre is and why Zambia wants one. If the tools running on that infrastructure carry imported bias, the bias arrives with them.

Background — bias in the machine

Concerns about AI bias are not new, but this study is unusual in testing four mainstream tools side by side with the same prompts. By holding the inputs constant and varying only the names, the researchers could isolate how the models treated different identities.

The names mattered. Prompts built around names such as Mthokozisi Sithole, Jo-Anne Mitchell, Trevino Naidoo and Nomfundo Moyo let the researchers see whether the tools wrote different futures for characters coded as different races and genders — and they did.

The authors stop short of saying the developers intended the bias. Their argument is that the models mirror their training data, which means fixing the problem is less about a single switch and more about the data, testing and oversight behind each tool.

What to watch

The first thing to watch is how the developers respond. The four tools tested are among the most widely used, and whether their makers acknowledge and address the finding will signal how seriously the industry takes inherited bias.

The second is regulation. African governments are still writing the rules for AI, and studies like this give them evidence to demand testing for bias before tools are deployed in public services.

The third is local capacity. The more African researchers and developers shape and audit these systems, the better the chance the tools reflect the people who use them rather than the data they were trained on.

Frequently Asked Questions

These are the questions readers have been asking about the AI bias study. Short answers follow, drawn from the research and its reporting.

What is the AI bias study?

In short, it is research that ran 112 story prompts through four AI chatbots and found consistent gender bias. The answer, simply put, is that men were written with futures and women with feelings. The key is that the pattern held across all four tools.

How does the bias show up?

The answer is in the stories the tools write. Data from the study shows male characters got ambitions and plans, female characters got emotions and caregiving, and Black female characters were often written without fathers.

Why is AI biased this way?

Simply put, the models learn from existing text and reproduce its assumptions. Analysis by the authors shows the bias is inherited from training data rather than a neutral output, which is why they say fixing it means fixing the data and oversight.

What are the risks for Africa?

According to the researchers, the risk is scale. Evidence shows that as AI drafts reports, screens applications and writes public messages, imported bias can harden into what looks like objective judgement across schools, offices and government.

Which AI tools were tested?

Research shows the study tested four of the most widely used tools: ChatGPT, Claude, Grok and Copilot. The same prompts were run through each so the results could be compared directly.

Sources

Daily Maverick: the hidden bias in AI narratives that gives men futures and women feelings, by Sioux McKenna and Nompilo Tshuma. Kwacha News coverage: AI farm tech and foreign capital, the public release of a leading AI model and what a data centre is.