AI-Powered Diagnostics

I want to push back on something before we even start. The framing of 'AI versus doctor' — the lesson handles it well, but it still comes up constantly in public conversations about this, and I think it does real damage.

It does, yes. And the people perpetuating that framing are roughly split between journalists who find 'AI takes your job' more compelling than 'AI reads a scan,' and doctors who have confused their professional identity with their diagnostic utility. Both groups are being unhelpful, for their own reasons.

That's a bit uncharitable to the doctors. There are legitimate questions about liability, about trust, about what happens when the AI is wrong and no one's quite sure who owns that error.

Those are excellent questions. They are also entirely separate from whether AI can detect an early-stage tumour in a chest X-ray faster and more reliably than a radiologist who has been on shift for nine hours. One is a governance problem. The other is just an empirical fact.

Fair enough. And the data processing point is genuinely staggering when you sit with it — millions of images, patient histories, research literature, all being cross-referenced simultaneously. That's not a marginal improvement on what a human can do. It's a different category of thing entirely.

Which is why I find the 'AI as a tool' metaphor slightly underselling it. A stethoscope is a tool. What's being described here is more like giving a doctor access to the pattern recognition of every radiologist who has ever lived, simultaneously, without the inconvenience of asking them.

Although a stethoscope doesn't occasionally hallucinate a diagnosis because the training data had a demographic blind spot.

...that is a fair point, and I will allow it.

The objectivity claim is where I get a little cautious, actually. The lesson says AI provides 'objective, data-driven insights' and reduces subjective interpretation. But the model was trained on data that was selected by humans, labelled by humans, and weighted by humans. At what point does that stop being objective?

It doesn't, strictly speaking. What it offers is a different kind of bias — one that is at least auditable in principle, rather than being locked inside a consultant's intuition. That's not objectivity. That's reproducibility. They are not the same thing, and the lesson could be slightly more precise on this.

Though for a student encountering this for the first time, 'AI reduces reliance on subjective interpretation' is a reasonable starting point. You can't teach the epistemological limits of training data before you've explained what training data is.

One can, in theory. One simply chooses not to, for reasons of pedagogy. I accept that. My concern is students who stop at the first lesson and walk away believing the AI is a neutral oracle, which is the kind of belief that leads to very poor decisions at the implementation stage.

Then the lesson after this one does some work. The personalised medicine section actually gestures at this — the idea that AI recommendations are only as good as the data a particular patient population has contributed, which is quietly one of the more important points in the whole piece.

Quietly, yes. It could be considerably louder. But I will concede the lesson does the most important thing correctly, which is establishing that the doctor is not being replaced — they're being given better information. A student who leaves with that distinction intact has learned something genuinely useful.

High praise from you. Relatively speaking.

I said it was useful. I didn't say it was sufficient.