Most weeks in my clinic at Kenyatta National Hospital, I meet a tumour that has been growing for the better part of a year.
Not because it was invisible, but because the road between a county dispensary and a specialist runs through months of waiting, referral letters and bus fare that a family does not have. Into rooms like mine, we are now told, artificial intelligence is coming to close the gap.
Parliament is deciding what to make of that promise. The Artificial Intelligence Bill, 2026, now before the Senate's ICT Committee, rightly classifies healthcare as high-risk.
As the senators weigh it, they will hear three words said in a single breath: diversity, inclusion, equity. In health AI, these are not one breath. They are three different stories, with different villains, and each one ends differently.
Consider a warning from Thailand. Google built an AI that could read a photograph of the eye and detect diabetic retinopathy with high accuracy, around 95 per cent for vision-threatening disease, matching specialists in the laboratory.
Deployed across 11 Thai clinics, it stumbled. Rooms could not be darkened enough for clean images, so the system rejected a fifth of the photographs nurses took. Slow internet turned each upload into a 90-second wait.
A tool built to speed up care produced queues and patients who went home unscreened. The model was not wrong. It was accurate, and it still failed the people it was meant to serve.
The first story is about who is missing. American clinical AI, one major study found, learns overwhelmingly from patients in just three wealthy states while most of the country contributes nothing. If those datasets barely represent rural America, imagine how completely they miss Kisumu.
In skin cancer detection, researchers have shown leading AI models falter on darker skin because the image libraries they learn from are dominated by lighter skin. This is diversity failing, and because absence can be counted, it is the failure we talk about most.
The second story unfolds earlier, in rooms we were never in. Before any data is collected, someone decides what a model is for and what counts as truth.
IBM's cancer tool was trained on hypothetical cases reflecting the preferences of a small group of doctors at one elite American hospital, then sold across Asia, where its recommendations collided with local guidelines and locally available drugs.
Nothing was wrong with the data. The machine had learned the opposite of inclusion: it carried one hospital's judgment into every room it entered and treated the rest of the world as an implementation detail.
The third story is the one my clinic already knows. A model can hold our faces in its memory and still fail the patient at the end of the corridor, because equity was never inside the model to begin with.
It lives in the lighting and the bandwidth, the staffing roster and the drug stock-out, the referral that takes months. An answer the clinic cannot act on has helped nobody, however correct it was.
Now the uncomfortable part. These three problems do not heal at the same rate. Diversity is a wound that closes: the same researchers who exposed AI's failures on darker skin retrained the models on diverse images and watched the gap disappear.
Inclusion scars: once a tool's purpose is set in a boardroom abroad, no volume of Kenyan data poured into that mould changes its shape. And equity cannot be engineered inside a model at all. It is delivered, or denied, by the health system waiting outside.
Yet the world funds the wound and ignores the scar. Datasets get the grants and headlines. Inclusion gets a workshop after the architecture is frozen.
Equity gets a promise of evaluation someday. The order should be reversed: secure decision rights first, because they cannot be retrofitted; demand patient outcomes under Kenyan conditions early, because accuracy statistics flatter. The datasets are the one problem time can still forgive.
This distinction is already at work on our continent. The ALIVE Framework for governing health AI, standing for Adaptive, Locally-grounded, Inclusive, Vigilant, Equitable, which I helped develop and declare an interest in, is carried by a community of practice across nine African countries.
It keeps Inclusive and Equitable as deliberately separate commitments, because a system can consult widely and still widen gaps. Kenya's drafters do not need to invent this.
The Bill is the chance to write it into law, and the need is no longer hypothetical: the High Court has already issued an interim order finding that the current absence of safeguards for high-risk AI leaves fundamental rights unprotected. The question is not whether to fill that gap but how.
Certification of high-risk health AI should ask harder questions than where the training data came from: who held the pen when the tool's purpose was written, and what happened to patients when it met a Kenyan clinic. Vendors will resist. They will offer us diversity, because datasets are the cheapest concession on the table.
Nor is this a debate reserved for senators. The Ministry of ICT has just opened the draft Kenya Artificial Intelligence and Other Emerging Technologies Policy for public comment, with submissions due by August 4.
If you have ever wanted a say in how these machines will meet our clinics, the window is open now, and it is short.
Diversity gets you into the room. Inclusion decides who arranged it.
Equity is whether the patient at the end of the corridor lives differently
because the room exists. Kenya should stop accepting the first as payment for
all three.
