Bias in the frame: making sure AI shows Africans as they are

Ask a general-purpose AI image model for "an African businesswoman" or "a street in an African city" and you will sometimes get something accurate. You will also, too often, get a collage of clichés: dusty roads, a generic "tribal" pattern, every person with the same skin tone and the same clothes, a skyline that belongs to no real city.

That is not malice. It is data. Models learn from the images that exist online, and for decades many of those images showed Africa through a narrow lens: poverty, wildlife, crisis and costume, usually photographed from outside.

What bias looks like in practice

• Sameness: a continent of 54 countries reduced to one look
• Poverty by default: homes, streets and people shown as poorer than the prompt implied
• Costume: traditional dress added to modern scenes where it does not belong
• Skin tone errors: dark skin rendered too light, poorly lit or with texture smoothed away
• Wrong places: landmarks, signage, vehicles and architecture from the wrong country

For brands, these are not abstract problems. An ad that shows your customers inaccurately tells them you do not see them.

Practical steps

1. Be specific. "A woman in her thirties in a modern Westlands office, in a tailored blazer" leaves less room for stereotype than "an African woman at work".
2. Name the place. Kisumu, Lagos, Kigali and Dakar look different. Say which.
3. Attach real references. Photos of your actual shop, staff (with consent) and customers' settings anchor the model in reality.
4. Describe skin and features respectfully and precisely.
5. Review every result and reject ones that stereotype, even subtly.
6. Involve people from the community you are depicting.

What Tari Studio does

Tari Studio's AI director adds instructions to every request for African scenes, people, languages and accents, and for faithful, correctly exposed East African skin tones. The AK1 model leans towards documentary realism, real homes, shops and streets; AK2 towards modern, confident, Afro-cinematic style. Neither defaults to poverty or costume. Meet AK1 and AK2.

You can see exactly what was sent to the model under Prompt we sent, and when a result is off, a thumbs down with a reason ("stereotyped setting", "skin too light") becomes part of the notes the Studio keeps about your brand. Brands and brand memory.

The bigger fix

Prompting can only do so much. Lasting change needs better data: more images of African life taken by Africans, more African languages in training sets, more African researchers and companies building models. The sovereignty question.

Until then, the responsibility falls on everyone using these tools to notice, reject and correct bias. Africa's image has been shaped by others for long enough.

Make ads that show your customers as they are.