The real risk isn't that AI will be "wrong" too often, it's that it will be right often enough that humans stop practising the skill. Pilots lose manual flying proficiency with autopilot, drivers lose wayfinding sense with GPS, and radiologists already double-check less when the AI agrees with them.

What makes medicine different is that the tail risks matter: you only need to miss one subtle but lethal case because you've dulled your instincts. And unlike navigation or driving, you don't get daily "reps" to stay sharp. Deskilling here isn't hypothetical, it compounds silently until a crisis forces a clinician to act without the crutch.

> Pilots lose manual flying proficiency with autopilot

Eloquently explained by Warren Vanderburgh in 1997: <https://www.youtube.com/watch?v=lIusD6Z-3cU>

> Pilots lose manual flying proficiency with autopilot

Depends on the type of plane. On small aircraft, all the autopilot does is keep the aircraft level; it's not doing anything which requires significant proficiency. (It's roughly akin to "cruise control" in cars, and is used in similar circumstances.)

>drivers lose wayfinding sense with GPS

My experience is the opposite. I find that while Google maps on my phone is a more than suitable replacement for the now almost impossible to get road Atlas of my youth, with the expanded metropolitan maps that were always out of date, my skills haven't been degrading. I'm now able to get real time feedback of traffic conditions to make better choices between routes I know by heart.

Just this past week I took a friend to visit someone out in the far suburbs, I used Google maps to get there, but he was astonished that I was completely comfortable driving home without it, despite all the twists and turns of modern US suburbs.

My experience with LLMs generating code is similar, they are better guides than the old school method of reading the manual and other books, but I remain able to get a handle on the code written when necessary.