The algorithm is the easy part
Artificial intelligence is often sold as the answer to the access gap in healthcare. From a radiologist's chair, the model reading the scan is the simplest part of the problem, and the responsibility lives everywhere else.
Every few weeks someone tells me that artificial intelligence is about to solve access to care. Radiologists are scarce across Africa, the argument goes, so let the software read the scans and the shortage disappears. I understand the appeal. I also spend my days reading mammograms, and from where I sit the algorithm is the easy part.
A model is a component, not a system
Software that flags a suspicious finding on a mammogram is real, and it is getting better. As a second pair of eyes it can help a tired radiologist catch what fatigue would miss, and it can triage a queue so the urgent cases rise to the top. None of that is in dispute. But a model is a component, in the same way a scanner is a component. Dropping it into a health system does not make the system intelligent, any more than parking a new machine in a broken clinic makes the clinic work.
The responsible use of AI in screening lives downstream of the algorithm, in the parts nobody demonstrates on stage.
Validate it for her, not just in the paper
A model learns from the data it was trained on. Train it mostly on women from one part of the world, and it can quietly underperform on women it was never shown. In a country as varied as ours, that is not a footnote. A tool that reports beautifully in a published trial can still miss the woman standing in front of the truck, and you will not know unless you check on your own population rather than trusting the brochure.
So the first responsible question is plain. Does this work for the people I actually serve, and how would I know if it stopped?
Keep a person accountable
The danger with a confident machine is that people stop arguing with it. Automation bias is well described: give a clinician a screen that says normal, and the temptation is to agree and move on. Used well, AI sits inside the reading process as an aid, with a named radiologist who owns the result and is free to overrule the software. Used badly, it becomes a way to read more scans with less attention, and the accountability leaks away to a vendor who never meets the patient.
Responsible AI keeps a human on the hook. The machine advises. A person decides, and can be asked why.
The follow-through still has to happen
A model that flags an abnormal finding into a system that then loses the patient has provided anxiety, not care. This is the same lesson the trucks taught me. The hard part of screening was never the image, it was the roster, the result that arrives in days, and the follow-up call that actually happens. AI can speed the reading. It does nothing for the woman unless the machinery of follow-through is already sound. Speeding up the first step of a process that drops people at the third just gets you to the failure faster.
AI multiplies the system it sits in
Here is the part that matters for the access thesis. AI is a multiplier of whatever system surrounds it. Placed inside sound access infrastructure, careful triage and remote reading let one urban specialist safely serve communities she will never visit, which is exactly the leverage a country short of radiologists needs. Placed on top of a broken system, the same tool industrialises the gaps: it misses the same people faster, at greater scale, with a veneer of objectivity that makes the misses harder to question.
That is why I get uneasy when AI is offered as a substitute for building the system rather than a component within it. The technology is not the scarce thing. A system that earns a woman’s trust, reaches her where she lives, and closes the loop after the scan is the scarce thing, and no model ships that in the box.
Where the responsibility sits
So when I am asked whether AI will fix access, I answer at a different altitude than the question. The responsible question sits past the model altogether. It is whether the whole arrangement reaches her, reads her scan well, and makes the call that changes her outcome. The algorithm is the easy part. Everything around it is the work, and the work is where the responsibility lives.