An algorithm that reads chest X-rays learns by looking at tens of thousands of them. It learns what a normal lung looks like, what tuberculosis looks like, what a tumour looks like, by being shown examples until the pattern sticks.
The question nobody asks at the point of purchase is whose lungs it was shown.
If the answer is Boston and Berlin and Seoul, and the lung now in front of it belongs to a farmer in Bolgatanga, the machine is working from a reference population that does not include him. It will still return an answer. It will return it with the same confident interface and the same clean percentage it would give a patient in Massachusetts. It just has a higher chance of being wrong, and nothing on the screen will tell the clinician which case this is.
That gap is what I spent months reporting for Biased by Design: The Human Cost, an investigative documentary on the deployment of artificial intelligence in Ghana’s health system, supported by a UK Government media capacity building programme.
This is not a future problem
The temptation is to file AI in healthcare under things Ghana will have to think about eventually. That is already out of date.
AI-assisted tools are in Ghanaian facilities now. Computer-aided detection has been used in tuberculosis screening. Smartphone-based ultrasound support is in the hands of midwives in places where there is no sonographer within an hour’s drive. Ghanaian companies are building diagnostic models domestically, among them MinoHealth, founded by Darlington Akogo, who also chairs the UN ITU and WHO focus group work on artificial intelligence for radiology.
Some of this is unambiguously good, and the reporting did not pretend otherwise. A midwife with a tool that flags an abnormal presentation is better equipped than a midwife with nothing. In a country with the specialist shortages Ghana has, that is not a small thing.
But adoption has run ahead of governance. The people absorbing the risk are the people with the least ability to see it.
What the reporting showed
I went into this expecting a story about technology. What I came out with was a story about absent rules.
The reporting took me to the people on every side of the tool. Darlington Akogo of MinoHealth, building diagnostic models in Accra. Esi Appiah Hagan, a radiographer and global health strategist. Dr Yussif Adams, Senior Lecturer and Head of Biomedical Sciences at the University for Development Studies. Dr Achumboro Aghana Paul, founder of MorphoAI at UDS. Mary Otopah, a Senior Midwifery Officer using AI-assisted ultrasound on real patients in a real facility. Laboratory scientists and innovators working on the problem from inside Ghanaian institutions.
What connected them was not disagreement about whether the technology works. It was that not one of them could point me to a Ghanaian rule governing how it enters a hospital, who checks it, or what happens when it is wrong.
The clinician at the bedside is left carrying a judgement that should have been made long before the tool reached her, by a regulator, a procurement officer, or a validation study that was never run.
The three questions nobody is answering
Who approved it? When a hospital in Ghana installs an AI tool that influences a diagnosis, which body signed off on it? Medical devices fall under a regulatory regime. Software that makes a clinical recommendation sits in a grey zone that most countries, Ghana included, are still writing rules for. If no regulator has validated a tool’s performance on a Ghanaian population, the hospital buying it is relying entirely on the vendor’s own claims.
Who is accountable when it is wrong? A radiologist who misses a lesion is answerable to a professional body. An algorithm is not. The liability question runs through the vendor, the hospital, the clinician who accepted or overrode the output, and the procurement officer who signed the contract. In Ghana, that chain has not been tested, and nobody wants to be the first case that tests it.
Whose data trains the next version? Many of these systems improve by learning from the cases they process. If Ghanaian patient data is being used to refine a model owned elsewhere, Ghanaians are supplying the raw material for a product they will later pay to license. That is not a hypothetical arrangement. It is the standard commercial model, and it is operating in a country whose Data Protection Commission only began enforcement spot checks this month.
The African-led answer
The more hopeful finding was that the correction is already being attempted from inside the continent. There are Ghanaian researchers and founders building models on local data, in local institutions, with local populations in the reference set. That work is real and it deserves support.
But it is a handful of people against a global pipeline with enormous capital behind it. Without procurement rules that ask what a tool was trained on, without a regulator that can validate performance locally, and without data governance that tells Ghanaians what happens to their scans, the default keeps winning.
The default is a system that was never trained on us, making decisions about us, with nobody required to say so out loud.
Ghana’s answer cannot be to refuse the technology. The shortages are too severe and the potential is too real. The answer is to insist on the question that no procurement form currently asks.
Whose data taught this machine, and does it include mine?
Dr. Emmanuel Raphael Samani is a medical laboratory scientist and a broadcast journalist with JoyNews. His documentary Biased by Design: The Human Cost examines the deployment of artificial intelligence in Ghana’s health system.