The panic says the machines are winning. The numbers say almost nobody is streaming what they make. The bigger question in South Africa is whose work those models were trained on.
More than half of the music uploaded to Deezer every day is now fully AI-generated. That is the platform’s own figure, roughly 90,000 machine-made tracks a day, more than two million a month. If you have spent time in a Spotify Discover playlist or deep in a TikTok feed lately, you have already met a few. Competent and convincing, catchy even, and made by no one in particular. No one played on them. A prompt and a dataset did the work.
The headlines call this a flood, and it is one. But the figure everyone quotes hides the figure that actually matters. On Deezer, AI-generated music now accounts for more than half of daily uploads while making up only around 1 to 3% of what people actually stream. Apple Music says AI accounts for less than 0.5% of listening on its platform, despite more than a third of the tracks delivered to it being fully AI-generated. Separate reporting suggests roughly 65% of AI tracks on the service have never been played at all. Of the streams these tracks do get, Deezer estimates 85% are fraudulent, bots inflating the count to skim the royalty pool. The apocalypse, looked at closely, is mostly an empty room with the lights left on.
Which should be reassuring, and in the northern hemisphere it more or less is. The claim that AI is out-writing humans falls apart the moment you check what anyone is willing to hear. Given the choice, people keep choosing the human. That is the comforting read, and it tends to get written from London and Los Angeles, where the industry under threat is big enough to absorb the hit.
We are not that industry. Down here the danger was never really about whether a machine can knock out a decent hook.
The export we never agreed to
In June, the local cybersecurity firm Cybersec Clinique went looking for South African music in the training data behind the big generative music models, drawing on the same publicly documented datasets The Atlantic had made searchable. It found more than 200 local artists across amapiano, kwaito, maskandi and gqom. Kabza De Small’s work appeared 360 times, DJ Maphorisa’s 270, Nasty C’s 193. Thomas Chauke and Spirit of Praise were in there too, and not one of them was asked or paid. The researcher behind the audit, Cybersec Clinique founder Doreen Mokoena, put it plainly: “The biggest export South Africa never authorised may be its culture.”
This is the part the global conversation keeps skipping. The datasets behind generative AI are built on real people’s work, and a lot of those people are here. Gallo Records has started fencing its catalogue off in response. GM Rob Cowling says the label is “ring fencing and creating a walled garden around its content” to protect it from AI scraping and training, and that Gallo will challenge any company found using its music for AI purposes. When a label as old as Gallo starts building walls, it has seen something coming.
Streaming was never going to pay the rent
The theft stings more because the thing being diluted barely paid local musicians in the first place. A 2022 national survey of South Africa’s music sector found 63% of respondents rating their streaming earnings as poor or very poor, a top-up to live income at best and, once the costs of getting music onto platforms were counted, a net drain at worst.
The headline numbers seem to tell a different story, and they need reading carefully. Spotify’s own figures put South African artists’ earnings above R504 million in 2025, up 28% on the year, with Sub-Saharan Africa MD Jocelyne Muhutu-Remy calling local artists a “globally dominant creative force.” True enough. But nearly 74% of that money came from listeners outside the country. The aggregate is growing fast, which tells you very little about what reaches the musicians lower down, the ones for whom streaming was never going to cover a month’s rent. That scene has always run on live shows and the door.
The law protects the song but not the singer
The law is nowhere near ready for any of this. South Africa’s copyright regime is still built on the Copyright Act of 1978. The amendment process begun in 2017 remains unresolved. President Cyril Ramaphosa returned the Copyright Amendment Bill to Parliament over constitutional concerns, and after Parliament reconsidered it, referred it to the Constitutional Court in 2024. The existing law has no dedicated regime for generative AI. Copyright can cover the song and the recording, and performers hold separate statutory rights, but the distinctive voice and timbre that make a singer recognisable have no clear protection against AI imitation.
It’s a tool, how you use it
It would be dishonest to draw the line as cleanly as some would like. When we put the question to our readers, most were in no mood to debate it. The reaction was blunt refusal, and under it a real question about what a machine could possibly know of the joy of discovery or the frustration of trying and failing at something. A smaller number made the counter-argument that gets made about every new technology, that it is only a tool and the only thing that matters is how you use it.
Both things are already true here. There are South African producers and topline writers using generative tools in international writing camps and remote sessions right now. For most it is a survival move made while the ethics stay unresolved, because the work pays and the rent is due. When the legitimate economy pays you in fractions of a cent, a tool that pays anything at all stops reading as a threat and starts reading as access. There’s nothing hypocritical in it. A scene that has never been paid what it’s worth will take the work that pays, and pretending otherwise would be its own dishonesty.
The thing a machine can’t fake
So if a machine can plausibly generate the song, and some of our own people are quietly using it to get by, what is actually left to defend?
The answer sits in a study Deezer commissioned late last year, run by Ipsos across 9,000 listeners in eight countries. In a blind test, 97% could not reliably tell the AI tracks from the human-made ones. If it is already that hard to hear, proof of the human has to come from somewhere beyond the audio, from knowing whose name is on it and that someone was in the room when it happened.
Cape Town is already standing on that ground. Think of the moment at this year’s Cape Town International Jazz Festival, no viral clip, just a drummer holding a phrase a beat longer than the arrangement asked for while the room leaned in rather than looked away. We called it the human event of sound at the time. It has aged into a competitive advantage. Bassist and composer Benjamin Jephta has argued that the most compelling use of AI he has encountered is the kind that imagines the technology alongside human performance and collaboration, rather than treating music as something to fully automate. Siya Mthembu of The Brother Moves On is blunter about the anxiety surrounding all of this: “I don’t think AI is gonna take anything that human beings won’t be able to make.”
You can see the same instinct in what people spend money on. Vinyl has grown for nineteen straight years and passed a billion dollars in the US alone, listeners paying a premium for the most physical, least optimised format there is. The platforms are racing to build that same reassurance into software. Spotify has started badging AI personas and marking out verified humans, and Deezer, Tidal, Bandcamp and Qobuz are now labelling AI tracks or refusing to pay royalties on them. The industry is quietly conceding that the human is the product. IFPI’s new rules keeping AI out of official charts apply first to the charts it runs directly, and the Official South Africa Chart is one of them.
Nobody’s romanticising the past here. This is the argument for what the scene already runs on, live players rebuilding a set on instinct in a full room, the thing bands like Operation Khataza and The Hii ROLLERS do live and collectives like AHCOM organise around, treating music as labour rather than content. Cape Town has been making that case for years without calling it a defence against AI. It turns out that is what it was.
Streaming, as we’ve said before, is a tool. AI is going to force a lot of people to work out where home actually is. Ours was never in the feed. It was in the room, watching someone think with their hands, and on the evidence so far, people here still turn up for it.
