In 2016, one of the world’s most respected AI researchers said we should stop training radiologists, because AI would clearly outperform them within five years. A decade later, the NHS has a radiologist shortage and is training more, not fewer. Keep that story in mind for everything you are about to read.
This is the final module, and it is different from the others. Modules 1 to 5 dealt with what exists: how AI works, how to use it safely, how to run it in your consultation room and your back office. This module deals with what is coming — which means it deals with uncertainty.
So before we look at any technology or any NHS plan, we need to talk about how to think about predictions. Because the biggest risk in this territory is not ignorance. It is confident wrongness — in both directions.
The two ways clever people get this wrong
Watch the discussion about AI in medicine for a while and you will see two failure patterns, often in the same conference programme.
Overestimating the short term. The radiology prediction is the classic. The technology demonstration was real — AI genuinely can detect abnormalities on images at impressive levels. What failed was everything around the prediction: regulation, liability, integration, workflow, edge cases, and the discovery that reading a scan is a fraction of what a radiologist actually does. Demonstrations move fast. Deployment moves slowly, and healthcare deployment moves slowest of all.
Underestimating the long term. In 2019, the idea that a machine could take a fluent history, draft a referral-quality letter, and explain a complex diagnosis in plain English at midnight was science fiction. By 2023 it was a free website. The people who dismissed language models as autocomplete in 2022 were as wrong as the people who cancelled radiologists in 2016.
Both errors come from the same source: extrapolating in a straight line, either from excitement or from scepticism. The future arrives unevenly — faster than expected in capability, slower than expected in adoption.
A more useful frame: capability versus deployment
Here is the frame I find most useful, and it explains most of the confusion you will encounter.
The capability curve is what AI models can do in demonstrations and benchmarks. It has been climbing steeply for years and shows no sign of stopping. Lesson 2 covers where it currently stands.
The deployment curve is what is actually running, governed and paid for, in your surgery on a wet Tuesday. It climbs far more slowly, because it is constrained by everything the capability curve ignores: safety cases, regulation, procurement, integration with thirty-year-old NHS systems, training, trust, and money.
The gap between the two curves is where hype lives. A vendor sells you the capability curve. Your working reality is the deployment curve. When the Royal College of General Practitioners surveyed GPs in December 2025, 28% reported using AI tools at work — mostly for documentation and admin. That is the deployment curve: real, growing, and years behind the demonstrations.
When you encounter any claim about AI in healthcare, ask which curve it describes. “A model passed a medical exam” is capability. “Our ICB has approved this tool and forty practices use it” is deployment. Both matter, but only one affects your Monday morning.
Why general practice specifically resists prediction
There is a reason confident predictions about general practice keep failing, and it is worth naming because it applies to every claim in the rest of this module.
General practice is not a diagnostic pipeline. It is undifferentiated complexity: the patient with four conditions, three medications from two hospitals, a housing problem, and a question they only ask with their hand on the door. It runs on continuity, trust, and risk held over time — properties of relationships, not transactions.
Technologies that automate transactions arrive quickly. Technologies that touch relationships arrive slowly, partially, and with more resistance than their inventors expect. That does not mean general practice is immune to change — Lesson 4 takes the question seriously. It means the change will be messier and less linear than either the enthusiasts or the doom-mongers predict.
The stance this module takes
So here is the honest position, and the one this module will hold throughout.
Nobody — not the vendors, not NHS England, not the researchers, and certainly not me — knows what general practice looks like in ten years. Anyone who claims to is selling something, even if it is only a conference ticket.
What we can do is more modest and more useful: understand the technologies that actually exist and their trajectory, read the NHS’s published plans carefully, distinguish the committed from the speculative, and prepare in ways that pay off across many futures rather than betting on one.
You do not need to predict the future to be ready for it. You need to understand the direction of travel, keep your practice adaptable, and stay current enough that nothing important surprises you. That is what the next five lessons are for.
Key Takeaway
Predictions about AI in medicine fail in two directions: overestimating the short term (demonstrations are not deployments) and underestimating the long term. Judge every claim by which curve it sits on — capability or deployment — and remember that general practice, built on relationships and held risk, changes more slowly and more messily than any straight-line forecast suggests.