AI must have a ‘human-in-the-loop’ is the wrong hill for clinicians to die on

“AI-must-have-human-in-the-loop” is the wrong hill for clinicians to die on. If we’re out of the loop? Now I get why some clinicians push back—but if we’re going to lead on AI, we need to re-think our messaging.Months ago I tried to train Claude on how I write. Fed it hundreds of my posts. Trained it on my “voice” and “style.” I'd have the ideas, Claude would write them “the way I would.”It did not work. A weird frankenstein of phrases and tone I guess I use, but . . . boring. Repetitive. Definitely not as funny as (I think) I am!But I use Claude for writing all the time in other ways. For a business case, job description— anything worth templatizing—I've built a Claude skill. I feed it the context, say “use my template skill” and boom—it makes something pretty darn good. Yes, it’s boring and repetitive—except that’s what I want in those situations.Same is true for many parts of healthcare. Often standardization beats variation, and humans are the cause of that variation. Turns out AI is much better at standardization than we are. It won’t complain about following rules or evidence-based guidelines!Right now a clinician should still “review the AI scribe draft”—but we'll reach a point where notes are so customized and the error rate so low that reviewing one isn’t worth the opportunity cost. Clinicians will stop checking.I predict we’ll eventually see the same with clinical AI. Currently CDS alerts surface recommendations, but a clinician must act—which is why guideline adherence is so inconsistent. Increasingly we’ll have AI agents that act for us. First labs and imaging, which are low risk. Eventually automated medication/treatment orders. Not everything, but we’ll find the right things—and that surface area will increase as the AI gets better.Which is why it's counter-productive when some clinicians die on the “human-in-the-loop” hill. It’s a bit hypocritical to say “humans must be in the loop . . . except for jobs that aren’t mine, like call centres!” AI will increasingly take clinicians out of the loop—it’s a matter of time horizon and order of tasks. We should be OK with that if it leads to better access and outcomes.I don’t think any reasonable clinician opposes this in principle. But what clinicians are against is being told what should be automated against their will. We already see this with the one-third of clinicians who reject AI scribes even when offered. Technologists may not appreciate this, but change management doesn’t happen through force—it has to be earned.The perception right now is that doctors are saying “AI must have human-in-the-loop” when the right messaging should be “AI allows humans to be increasingly out of the loop, and that’s a good thing—but we as clinicians will lead on when, where and how quickly that happens to ensure it’s safe and effective for patient care.”Dr. Joshua Liu is a Toronto physician and the co-founder/CEO of SeamlessMD. This item originally appeared on LinkedIn.

AI must have a ‘human-in-the-loop’ is the wrong hill for clinicians to die on

Register for free access or login.

Related Topics