What we don’t talk about when we talk about AI in medicine:
Standard of care isn’t standard care. New modalities of care (AI-assisted or otherwise) are always measured against a platonic “standard of care” that doesn’t exist in the real clinic. Medicine is as much art as science. Providers are overworked, tired, rushed, and routinely veer from textbook SOC.
Novel care modalities get killed in the cradle because they can’t measure up to the imaginary provider with infinite time and infinite knowledge. But that’s not what most patients get today. Less-than-perfect care that scales to many more patients is an obvious moral good, but our current bar makes it almost impossible.
Every provider understands that real SOC deviates. Not everything is black and white. Physiology doesn’t fit into clean textbook pathways. The gray areas are where great clinicians shine. We would never allow AI systems to operate in those same gray zones, even if that’s where most real care actually happens.
We also never talk about the patients we don’t see. Clinicians’ judgments about “how healthcare works” come from the patients in front of them. They don’t see the patients who never make it in the door, or the ones who are dead or much sicker because of something that happened upstream: bad access, no access, or bad care.
We struggle to reason about the enormous population that exists completely outside the current healthcare system — invisible due to lack of trust, lack of access, geography, scheduling, childcare, work, etc. AI-assisted, consumer-first care may be one of the only ways to reach many of these patients where they actually are.
Everyone is terrified of AI “hallucinations” and mistakes. But providers hallucinate too. Medical errors are estimated as a leading cause of death in the US. The difference is that when a human clinician makes a mistake, we have malpractice law, liability frameworks, and trillions of dollars of insurance wrapped around that reality.
As AI systems get better and account for a larger share of real decision-making, we’ll have to answer basic questions: do we license and insure models the way we do clinicians, or do providers simply inherit the risk of using them?
Healthcare also isn’t “lindy.” Private health insurance in its modern form took off in the 1950s. Medicare and Medicaid launched in the 1960s. Most of the system we treat as sacred is younger than many of the executives running it. There’s no reason to believe it should exist in its current form for the next century.
Consumer choice is the canary. Our crisis of cost and access is driven by a narrow set of rent-seeking actors who have no real incentive to improve anything. AI-assisted, consumer-first healthcare will eventually show that cheaper, better care is actually possible — and in doing so, expose a lot of the existing system as a scam.
It’s time to (re)build trust. Right now, many providers act like the second-worst thing that can roll into their department is a VC or a tech exec. That distrust is earned. Tech has overpromised and underdelivered in healthcare for decades. But staying in that posture forever won’t fix anything.
The only way out of this crisis is to take seriously the idea that this time might actually be different — and then hold people to it. Venture-backed implosions make for good schadenfreude. But we should feel no pride in delivering bad care, at high cost, to too few people, because we refused to take new tools seriously.
