The pilot of a commercial aircraft, descending into near-zero visibility, hands control to an automated system designed for precisely that moment. The landing is flawless. The passengers are safe. The pilot announces it proudly. Technology functions exactly as promised.
A doctor on board writes, “It made me think about the health care journey I am on with AI. The safety and well-being of 140 people wasn’t left to human intervention; the stakes were too high, and clearly machine has been proven to outperform human. I wonder how long it will be before the trickiest diagnosis, the earliest detection, or the best recommendations will be handled by the medical profession the same way my landing was?”
It is an elegant metaphor. It is also a dangerous one, at least when imported wholesale into medicine.
The limits of the aviation metaphor
In aviation, autoland exists for a narrow, well-defined purpose. The parameters are known. The runway is fixed. The physics are predictable. The system has been tested endlessly under tightly controlled conditions. And crucially, it does not replace the pilot. It is invoked rarely, for the hardest landings, under explicit rules, with human oversight before, during, and after the event. Less than 1 percent of landings use autopilot technology, not because the technology is weak, but because the system it operates within is disciplined, bounded, and standardized.
Medicine is none of those things.
The optimism embedded in the physician’s reflection rests on an implicit assumption: that medicine’s hardest moments resemble a foggy runway, rare, clearly defined, and amenable to algorithmic mastery once the data are sufficient. But the most difficult problems in health care are not low-visibility landings. They are ambiguous, evolving, deeply human situations in which the destination itself may be contested.
A diagnosis is not a runway. A patient is not a system state. And clinical care is not governed by a single set of invariant rules.
The complexity of clinical care
When aviation automation fails, investigators can usually trace the failure to a known category: sensor error, training lapse, software edge case, or human-machine interface breakdown. When medicine fails, the causes are more diffuse: social context, access to care, trust, bias, values, timing, fear, denial, financial pressure, and sometimes knowledge gaps. These are not “noise” to be filtered out. They are the substance of care.
The doctor’s analogy also obscures a crucial difference: who built the system.
Autoland was designed by engineers working within a tightly regulated industry, informed by decades of accident investigation, standardized aircraft designs, and a shared global safety culture. Medicine’s AI systems, by contrast, are often trained on incomplete, biased, and commercially mediated data. They are shaped not only by clinicians and scientists, but by incentives that reward scale, speed, and market capture. We should not pretend that these systems emerge from a neutral sky.
The accountability problem
When an AI system “outperforms” a human diagnostician, what does that mean, exactly? Outperforms on which population? Under what assumptions? Using which definitions of success? A correct diagnosis delivered late, without context, or without the patient’s trust may still be a clinical failure. Medicine’s outcomes cannot be reduced to touchdown.
There is also the matter of accountability. When a plane lands itself, responsibility remains unambiguous. The airline, the manufacturer, the regulators, and the pilot all operate within a clear chain of accountability. In medicine, that chain becomes murkier the more autonomy we grant machines. If an AI system misses an early cancer because its training data underrepresented a demographic group, who answers to the patient? The clinician who followed the recommendation? The institution that purchased the system? The company that trained it? Or the algorithm itself, which cannot explain, apologize, or bear moral responsibility?
This is not a theoretical concern. It is already playing out in some ways: clinical decision support tools that nudge rather than explain; algorithms that recommend without revealing their reasoning; risk scores that feel authoritative but conceal value judgments about what, and who, matters.
Augmentation, not abdication
The physician’s story suggests a future in which doctors proudly announce to patients that the “trickiest diagnosis” will be handled by AI, just as the pilot announced autoland. I find that vision disturbing, not because I doubt the power of technology, but because I doubt the wisdom of transferring authority without transferring understanding.
Patients do not come to medicine seeking optimal pattern recognition alone. They come seeking interpretation, judgment, and partnership. They come with stories that do not fit neatly into training sets. They come with fears that no model can quantify. They come needing someone to notice what doesn’t quite add up, to sense when the data are technically correct but clinically wrong.
Nothing in medicine corresponds to “near zero visibility at the field.” There is no moment when the human should step aside entirely because the machine sees better. There are moments when machines can assist, augment, warn, and support. And we should embrace those moments enthusiastically. AI can already help us detect patterns earlier, reduce certain errors, and manage complexity at scale. That is real progress.
But autopilot is not augmentation. Autopilot is abdication, however well intentioned.
A better model for medicine
Perhaps the better aviation analogy is not autoland, but the pilot who knows when to trust instruments and when to question them. The pilot who understands how automation fails. The pilot who remains responsible even when the machine is flying. That model preserves human judgment rather than replacing it.
Medicine does not need fewer humans at the controls. It needs better-supported ones, clinicians who understand both the power and the limits of AI, who can explain its recommendations, challenge its assumptions, and integrate its outputs into a broader human context.
Interesting times, indeed. But if we are going to borrow metaphors from aviation, let us borrow the right ones. Safety in flight has never come from turning pilots into passengers. It has come from respecting complexity, designing for failure, and insisting that humans remain accountable for the systems they create.
After all, planes do not heal people. Doctors do. And no amount of automation changes who must answer when something goes wrong.
Arthur Lazarus is a former Doximity Fellow, a member of the editorial board of the American Association for Physician Leadership, and an adjunct professor of psychiatry at the Lewis Katz School of Medicine at Temple University in Philadelphia. He is the author of several books on narrative medicine and the fictional series Real Medicine, Unreal Stories. His latest book, a novel, is JAILBREAK: When Artificial Intelligence Breaks Medicine.
