I need to tell you about a moment of shame.
It was a Tuesday, late, and the house was quiet. My wife Marthe was asleep. She's Senegalese, and I mention it because it matters for what comes next. I was working on a promotional image for an artist at the European label where I serve as press officer. We specialize in West African music, the music I've lived inside for decades. I needed a press photo concept: an African producer at work in a professional studio.
I typed my prompt into the AI image generator. I was specific. I thought I was being thoughtful. I wrote about a home studio, about the contrast between modern technology and its surroundings. I hit generate.
What came back was a young Black man in a concrete room with a cracked wall, hunched over a laptop on a wobbly table, a single bare bulb overhead. The resilient creative. The Third World improviser. The charity-poster cliché, delivered in four seconds.
I sat there looking at it and felt sick. Not only because of what the machine had made, but because of what I had asked for. I had wanted a professional studio. Somewhere between that thought and my keyboard, it had become a "home studio" with "contrast." My words, my supposedly enlightened, forty-years-in-African-music, married-to-a-Senegalese-woman words, had led the machine straight to the poverty cliché. Not consciously. But I asked for it.
I closed the laptop. I didn't tell Marthe. How do you tell your wife that after twenty-five years of loving her family, her culture and her country, your default imagination still runs on colonial software?
What the machine knows and doesn't know
My prompt was only half the story. The other half is what the machine had to work with.
Here's the thing about AI that tends to get lost in the hype: it doesn't understand anything. It's a mirror. And right now it reflects a world built by and for the West.
The numbers are brutal. Researchers at the Mohamed bin Zayed University of Artificial Intelligence surveyed the datasets used to build music-generating AI. Roughly 86 percent of all the hours of audio come from the Global North 2. European music accounts for more than 6,000 hours. African music, from a continent of more than 1.4 billion people and thousands of distinct traditions, accounts for about 28 hours 2. Not 28,000. Twenty-eight. Another breakdown of the same research puts Western genres at 94 percent of the data and Africa at 0.3 percent 1. The people doing the research look much the same: only 6.1 percent of the papers have a first author from the Global South, and not a single one has a first author from Africa 2.
Commercial platforms don't disclose what they train on. But there is little reason to believe their data looks any different.
I tested this myself on Suno, one of the biggest AI music platforms. I asked for mbalax, the Senegalese genre built on sabar drum patterns that I've been listening to since before most Suno engineers were born. What came back was a vaguely West African-sounding mush. Pleasant enough. Utterly generic. Then I asked for highlife. Similar mush, slightly different seasoning. Kuduro. Same. Amapiano. Same.
Then I asked for country music, and the machine came alive. Outlaw, bluegrass, honky-tonk, Bakersfield, polished country-pop: it knows them all, in detail. It can hear the gap between Waylon Jennings and Taylor Swift. It cannot hear the gap between Youssou N'Dour's mbalax and Fela Kuti's Afrobeat: different countries, different decades, different traditions, different languages. To the machine, they are the same thing: African.
I sat with that for a while.
And it isn't just my experience. Researchers have documented that when Suno attempts Arabic maqam music, it may round off the microtones to the nearest Western note, stripping the music of the very thing that makes it distinctive 2. A piece meant to evoke an Indian raga can come out sounding like a Western pop melody played on a sitar 2. Ask the model to step outside its comfort zone and it falls back on Western harmony and rhythm, because its comfort zone was built almost entirely on Western music.
It's not just music
At the label, we were brainstorming a name and visual identity for a dance festival in Africa. My colleagues and I worked through the AI-generated options. Every name suggestion came drenched in the kind of cultural shorthand Western marketing departments have been recycling since the seventies. Every poster concept: masks, sunsets, silhouettes against orange skies. The visual language of the tourist brochure, not the club flyer.
We tried to steer the AI away from it. We couldn't. Not because we didn't know what we wanted (we knew exactly what we wanted) but because the machine's entire vocabulary for "African event" was built by Western designers, photographers and art directors over decades of well-meaning exoticism. The cliché wasn't a bug. It was the only language the system spoke.
I've seen the same thing in text. Ask a large language model for a story set in Dakar and you get heat, dust, market noise, a woman carrying something on her head and a vague aura of spiritual wisdom. Ask for Paris and you get a three-dimensional character in a specific café with a specific emotional problem. Dakar is a backdrop. Paris is a place.
Ask for a poem in Wolof and you get something that sounds like Wolof the way a fortune cookie sounds like Confucius.
The pattern repeats everywhere: image generation, music production, text, design, marketing copy. The entire creative AI stack is Western infrastructure with a global user interface painted on top. By one estimate, 90 percent of natural language processing tools are available only in European languages 7. The contrast between multilingual Africa and today's English-centric AI models could hardly be starker 4.
The Kodak echo
I keep thinking about Kodak. For decades, the company calibrated its color film against a "Shirley card," a reference photograph of a white woman. That was the benchmark for skin tone. Dark skin came out flat and muddy, without detail. Not out of malice. Out of market logic: the customers who counted were white, so the technology was optimized for white.
Do you know when they finally fixed it? When chocolate makers and furniture manufacturers complained that their products, dark chocolate and dark wood grain, weren't coming out properly in advertising photographs. The adjustment that eventually did justice to Black skin was made for candy bars and tables. Not for people.
AI image generation is the Shirley card of this decade. And the prompt writer (me, you, anyone typing words into a box and expecting magic) is the photographer who never learned to light a dark face because nobody taught him it mattered.
To get a dignified image of an African musician, I have to tell the machine: make it look Western.
So here's what I'm actually saying
I'm not writing this to complain about AI. I use AI every day. I'm writing this to say: stop using it blindly for work that touches cultures the technology doesn't understand.
That's the message. Simple as that.
If you're a Western creative agency making a campaign for an African artist, stop generating images and hoping for the best. You are producing clichés at machine speed. You're not saving time. You're doing damage faster.
If you're a label, a promoter or a journalist using AI to write about African music, read what comes out. Really read it. Does it sound like a human being who knows this music, or like a Wikipedia summary with adjectives? If it's the second, throw it away.
If you're an African creator using these tools to work with your own heritage, know that the deck is stacked. Not against you personally, but against your entire tradition. The AI will flatten your mbalax into generic Afrobeats, your griot storytelling into "African folklore," your visual identity into tourist-brochure aesthetics. Not out of malice. Because it simply doesn't have the data to do better.
And this is the part that keeps me up at night. The workaround I found, namedropping Western aesthetic standards in my prompts, Apple Music editorial photography or Dazed magazine, to force the model into a higher visual register, works. It produces better results. But think about what that means. To get a dignified image of an African musician, I have to tell the machine: make it look Western. That is not democratization. That is assimilation with a progress bar.
The responsible thing, the honest thing, is not to use these tools for this work unless you know exactly what you're doing and why. And even then, you should be uncomfortable. If you're not uncomfortable, you're not paying attention.
The people building what the West won't
I would be lying if I said nobody sees this.
Masakhane (isiZulu for "we build together") is a pan-African grassroots network of thousands of volunteers, coders and researchers building language technology for African languages 5. They have produced translation research for dozens of African languages 6. They're doing it with older models, limited computing power and the kind of stubbornness that only comes from knowing your language matters, even when Silicon Valley doesn't think so.
When Jade Abbott, who cofounded the South African company Lelapa AI with Pelonomi Moiloa, asked ChatGPT to count from one to ten in isiZulu, a language spoken by more than ten million people, the results were, in her words, "mixed and hilarious" 5. In 2024, Lelapa launched InkubaLM, billed as Africa's first multilingual small language model, covering Swahili, Yoruba, isiXhosa, Hausa and isiZulu 8. Inkuba means dung beetle: a creature that moves loads hundreds of times its own weight.
GhanaNLP in Accra. Digital Umuganda in Kigali. The EthioLLM project in Ethiopia 4. Small teams, small budgets, enormous ambition. They're building from scratch what billion-dollar companies never bothered to include.
But here's the honest scale of it. In the survey of music-generation research I quoted earlier, not one paper had a first author from Africa 2. These are passionate people fighting a structural deficit with dedication and duct tape. They deserve funding, recognition and, most urgently, for the rest of the world to stop pretending the problem doesn't exist.
The last song
Here's where I refuse to end on a note of despair. I've spent forty years in music, and if those years have taught me anything, it's that the original always outlasts the copy.
So much of the music the West loves traces its bloodline back to Africa. Jazz, blues, rock and roll, funk, soul, house, techno, hip-hop: follow the rhythm far enough back and you end up on that continent. Every time the global music industry has eaten itself hollow and run out of ideas, it has turned to Africa. This is not romance. This is the historical record.
And now the world is leveling itself into sameness. AI makes it possible for anyone to produce something that sounds competent. One platform alone was reported to be turning out ten songs every second 3. A flood of content that is getting very, very good at being average. The middle is expanding. The edges are disappearing.
In my own creative work I'm no purist. I write songs and play bass; I've recorded with Graham Gouldman and Paul Burgess of 10cc and with Ton Scherpenzeel of Kayak. And I use Suno. But I use it knowing what it can and cannot do. It can give me a competent arrangement. It cannot give me the thing that makes a song matter. That still has to come from a human being who has lived something.
And here's the turn I keep coming back to: the creators these tools left out, whose traditions were compressed into 28 hours of data, whose languages were dismissed as "low-resource," whose visual identities were reduced to sunset silhouettes, may well outlast everyone else.
Not because they chose authenticity. Because they never had the luxury of faking it.
The West is drowning in its own cleverness. Its creative class has more tools than ever and less to say. Authenticity has become a marketing term. Originality is measured in clicks. The machine feeds on itself and calls it innovation.
But the griot under the baobab, kora on his knee (and I've sat next to enough of them to know this is not a metaphor but a Tuesday afternoon), doesn't need the algorithm. His music predates the training data by centuries. His tradition is the data the machine never bothered to collect. And when the flood recedes, when generated content collapses under its own weight because there is only so much average the world can absorb, his voice will still be there.
In a world where it gets harder every day to be a genuine human being, in music, in writing, in art, in life, there is something unbreakable about people who never could be anything other than what they are. The West is scrambling to rediscover authenticity. These traditions never lost it. Not because Africa is pure or simple or any of the other patronizing things Western observers like to project, but because a tradition that was never deemed worth digitizing was never available for diluting either.
The last song will not be generated. It will be played. On an instrument the algorithm has never heard, in a tuning the training data doesn't recognize, in a language the model can't spell.
And it will be magnificent.