Most talk about artificial intelligence (AI) in medicine starts with what it will take over. Will it read the scans, write the notes, replace the doctor?
Pro-human AI starts from the other end. It asks what AI lets people do that they couldn't do before, and whether they come out more skilled. The aim is to extend human reach, not to imitate people and replace them.
In medical education that has a practical meaning. The scarce thing in training is expert attention: practice, watched by someone who knows the job, followed by feedback. Pro-human AI should give more of that to more trainees, and leave them better at the work.
Will AI replace doctors? That depends on what we build
The economist Erik Brynjolfsson draws the line clearly. AI can be built to imitate what people do, or to augment them. Imitation is the more obvious goal, but "automating labor ultimately unlocks less value than augmenting it to create something new." Augmenting people, he writes, "opens an endless frontier of new abilities and opportunities" (Brynjolfsson, 2022, Daedalus).
The computer scientist Ben Shneiderman adds that this isn't a trade between automation and human control. "Well-designed technologies that offer high levels of human control and high levels of computer automation can increase human performance" (Shneiderman, 2020).
So whether AI replaces doctors or extends them is a design choice. We've set out the different kinds of AI tool used in medical education before. Each can be built either way.
AI can make people more skilled
Augmentation would be a thin idea if AI only made people faster while it was switched on. The evidence says it can do more than that.
Professional Go players are a striking case. Researchers analysed more than 5.8 million moves made by professionals between 1950 and 2021. After superhuman AI arrived, "humans began to make significantly better decisions", partly by playing moves no one had tried before (Shin et al., 2023, Proceedings of the National Academy of Sciences). The players learned from the AI and became better players.
The same thing has happened in ordinary work. When 5,179 customer support staff were given an AI assistant, the number of issues they resolved per hour rose by 14% on average and by 34% for novice and low-skilled workers. Staff two months into the job with the assistant performed as well as or better than those with more than six months' experience without it (Brynjolfsson, Li and Raymond, 2025, Quarterly Journal of Economics).
They also kept the gains. During system outages, when the assistant wasn't available, staff who had been using it continued to perform better. The authors conclude that tools like this "can help workers develop durable skills."
The biggest gains went to people early in their careers. That's who medical education serves.
In medical training, the scarce thing is expert attention
Doctors learn a consultation by doing it, then hearing what to change. The feedback is the expensive part. It needs an experienced clinician's time, and that time is already stretched. We set out in why postgraduate training has a different problem from medical school how much of a registrar's learning happens with no one watching.
This is where AI can extend reach. It can't stand in for a supervisor's judgement. It can give a trainee more realistic practice between supervised sessions, with specific feedback each time.
The early evidence points the same way as the evidence from outside medicine. In a small randomised trial in Germany, medical students practised history-taking with an AI patient. After four sessions, the group that also received structured AI feedback outperformed the group that practised without it on clinical decision-making (Brügge et al., 2024, BMC Medical Education). The practice helped. The feedback made it count.
That's augmentation. The AI does the part that doesn't need an expert present every time, so the expert's time goes further.
Extending people versus imitating them
The same technology can be pointed either way. The difference shows up in what it does for the learner, the supervisor and the decision.
| Built to | Imitate | Extend |
|---|---|---|
| Learner | Gets the answer | Does the work |
| Supervisor | Replaced | Reaches more trainees |
| Assessment | AI decides | A person decides |
| Afterwards | Less able | More able |
The imitation route has a known cost: a tool that does the learner's thinking can leave them less able. We've covered that in why medical students will need to be better, not worse, in the AI era. The extension route is the alternative, and the early evidence for it is encouraging.
A person makes the decisions that matter
Extending people means people stay in charge. The World Health Organization's first principle for AI in health is protecting human autonomy. In health care, it says, "humans should remain in control of health-care systems and medical decisions" (World Health Organization, 2021).
In training, the decisions that matter are about the learner. Is this trainee competent? Ready to progress? Ready to work unsupervised? Those calls belong to a person who can explain them and be questioned on them. AI can widen what that person has to go on. It shouldn't make the call.
What this means for what we build
Gestalt builds consultation practice with formative feedback. A learner works through a realistic consultation and gets specific feedback on how they reasoned and communicated.
The aim is reach: more practice with feedback than supervisors alone can give, for trainees wherever they are.
The test of pro-human AI is easy to state. After using it, can people do more than they could before? That's the standard we think every AI tool in medical education should meet, ours included.
AI will be part of every doctor's working life. The tools worth building are the ones that leave doctors able to do more.