Clinical AI · Health Informatics · Digital Health Systems · Human Dimensions of Technology
Written from the perspective of a practising physician.
What Health Information Systems Taught Me About the Future of Healthcare AI — A nurse asked: “What is the output? Why are we doing this much struggle?” She had no citations. She had years of frontline experience. She was diagnosing a design pattern that is now being reproduced in healthcare AI. Three lessons from implementation experience — and why AI integration must learn them before it is too late. Article 1 of 4.
Read full articleListening Before Asking — I downloaded a wellness app and within minutes felt lost — not because of a technical fault, but because it asked me to classify myself before it understood anything about me. A reflection on choice paralysis, the design of healthcare AI, and why the best clinical encounters always begin with the same simple question: Tell me what brings you here today.
Read full articleWhen AI Said, "You Make Me Smile" — I study AI. I know it cannot feel anything. And yet, when those four words appeared on my screen, I paused. Just for a moment. If that can happen to someone who understands the technology, what happens when a frightened patient hears the same thing? A reflection on the illusion of emotional understanding in healthcare AI.
Read full articleWho Will Supervise the Supervising AI? — One evening, my son asked me three questions about AI in healthcare. I answered the first two confidently. The third one — asked quietly, after a pause — I could not answer at all. Who will supervise the supervising AI? A child had found, in one sentence, one of the oldest and hardest problems in governance.
Read full articleThe Day Health Advice Stopped Traveling Through People — In rural Sri Lanka, patients once arrived at my clinic carrying conversations — advice from daughters, neighbours, old friends. Then a young man held up his phone. He had asked an AI. It answered well. But something was missing. A reflection on what AI does to the human chain of care.
Read full articleThe Day My AI Assistant Disagreed With Itself — I built an ICD-O coding support assistant for cancer registry work and grounded it in fixed reference sources. Then the same system gave two different answers to the same question. Neither was wrong. Both were justifiable. That distinction changed how I think about AI in healthcare.
Read full articleA colleague described a difficult conversation with a patient’s family. Afterwards, she said: “I kept thinking about how I could have framed it better.” She used the word framed. Not said differently, or explained more clearly. Framed. The language of prompts has entered the language of care without anyone announcing it.
There is a specific discomfort that arrives when you ask someone a question and immediately, before they answer, begin composing your follow-up. You are no longer listening. You are processing. The person in front of you has become an input. Most of us do not notice when this happens. The ones who notice it in others rarely notice it in themselves.
A patient told me he had stopped keeping a diary because he had started using AI to summarise his week. “It is more accurate,” he said. “My diary was always what I felt happened, not what actually happened.” I did not tell him that this was the point of a diary. He seemed genuinely content. I have been thinking about that conversation since.
The most common use of the phrase I don’t have bandwidth for this is in response to someone who needs something that cannot be scheduled. Bandwidth is a network term. We have borrowed it to mean human attention, human capacity, human presence. The borrowing was quiet and has become so complete that people use it without hearing it.
In health informatics, a confidence score tells you how certain a system is about its output. Below a threshold, you review the result manually. Above the threshold, you may not. The threshold is set by humans, and then forgotten. Most people who act on the output never see the score. This is not a note about AI in medicine. It is a note about what happens when we forget where our confidence comes from.
A fifteen-year-old can identify prompt engineering in a conversation. His father could not. This is not a story about intelligence. It is a story about what you are trained to notice, and when the training begins.
Grief is inefficient. It does not produce output. It does not shorten when you manage it correctly. It does not respond well to systems. A mind trained for ten hours a day to reduce inefficiency will, eventually, begin to apply that training to grief. The results of this application are not yet well understood, because they tend not to appear in clinical settings until much later — when the person arrives and says: I should have cried by now. I don’t know where it went.
I was reviewing a clinical AI tool’s output when I noticed the system had flagged its own recommendation as “reconstructed from incomplete data.” It was more honest about its limitations than most people I have worked with. That is not a criticism of people. It is a question about what we have built into machines that we have not yet built into ourselves.
More notes will be added as they are written.
There is no schedule. Observations arrive when they arrive.