It happened in the car, on an ordinary drive, with all three of my kids in the back. I did what I now do a dozen times a day without thinking—I talked to an AI (it’s a single press in my car, and something I don’t usually do in their presence) to ask about the best ice cream in town. I got back an answer in that fluent, endlessly agreeable voice, that gives multiple options, which all sound nice. And from the back seat, silence. Then my 7-year-old, genuinely startled: “Baba, who is that?”
I gave the answer you’d expect—it’s a computer assistant, it helps me with work. But the question stayed with me long after the drive ended, because the honest answer is that I’m not entirely sure. And whatever that voice is, it felt strange having it in the car with my children before I had decided what I wanted them to make of it. That feeling—the weirdness of it—is what this column is about.
Where I part ways with Sam Altman
Sam Altman has been arguing lately that the voice belongs in exactly that seat. He has suggested that parents connect their family calendars and their kids’ interests to ChatGPT so it can generate a personalized daily podcast for the school run, and he has said he “cannot imagine” having raised a newborn without it. Readers of this column know I am no AI skeptic—I have spent years arguing that these systems will transform our field. But here I disagree with him.
I am torn, and I am certain my views will evolve—probably several times before my 2-year-old reaches the age my 7-year-old is now. My children will grow up in a world saturated with these systems no matter what I decide at bedtime. But right now, today, I am keeping them away.
Here is what we do instead. I have been reading to them from my old encyclopedia—the set from the mid-1980s that I grew up with, dated entries, occasional charming wrongness and all. But the reading isn’t really the point. The point is what happens before the reading: I make them use the index. You want to read Jack and the Beanstalk? Let’s look it up and page through the numbers. Oh—we stumbled upon another story you want to read? Let’s take that detour.
It is slow. It is friction. The friction and the surprises are the curriculum.
What an index actually teaches
On a recent episode of our Real Talk podcast, I covered a position paper out of Google DeepMind with a title that doesn’t hedge: “LLMs can’t jump.” Its argument, built on Einstein’s own account of discovery, is that AI has mastered pattern-finding and logical derivation but lacks the leap—the moment Einstein imagined a man falling from a roof, felt in his mind that the man would be weightless, and jumped from that sensation to a new theory of gravity. The models have read everything and experienced nothing. The machines cluster; people jump.
If that is even half right, then the scarce skill of the next generation is not retrieval. The machines have won retrieval, decisively and forever. So what does a child gain from an index that she doesn’t gain from a pleasant voice that answers before she’s finished asking?
Three things, in my view. One: knowledge has a shape—it is organized, it lives somewhere and you can learn your way around it. Two—finding is an act you perform, not a service you receive, and the effort of the search is precisely what makes the answer stick. Three—and this is the one I keep coming back to: an index teaches you what exists without requiring you to hold it all in your head. You don’t memorize the encyclopedia. You learn that the entry is there, so that when the world hands you something that doesn’t fit, you know there is somewhere to look.
That third skill has a name in our profession. We just don’t teach it on purpose yet.
The harder job
I think about this as a father, but the same question is coming for us as physicians, and sooner than most of us would like.
In my view, diagnosis is on its way to becoming a commodity. Not this year, not next—but the trajectory in our own field, from autonomous diabetic retinopathy screening onward, points in one direction. And when a superintelligent diagnostic machine reads the scan and drafts the differential, the physician does not disappear. The physician becomes the check. Someone must own the judgment, catch the miss, carry the responsibility no model can hold. That sounds like a smaller job. It is a harder one—because being the check means proving, case after case, that you add value on top of a system that is very good and relentlessly getting better.
What does competence look like in that world? You still have to pattern-match; that never goes away. You definitely shouldn’t reach for OpenEvidence on every patient! But maybe it is finally all right if you can’t rattle off every white-dot syndrome on the tip of your tongue—so long as you know the white-dot syndromes exist, so that when you see something weird at the slit lamp, you recognize weird for what it is and know exactly where to look. Tip-of-the-tongue recall is what we are commoditizing. Knowing-it-exists, noticing-it-doesn’t-fit—that is the index skill, and it is upstream of the jump. I want residents to have it. I want my kids to have it. And I don’t yet believe a frictionless answer-machine builds it.
None of this is an argument against AI in education—quite the opposite. Coming up on AI in Eye Care, I’ll be featuring educators in our field who are working with students and residents on exactly this frontier, and I think there is enormous promise here: AI tutors that probe a trainee’s reasoning instead of handing over the answer, AI surgical coaches that watch every case and surface the feedback no attending has time to give. The distinction I care about isn’t AI versus no AI. It’s whether the tool builds the index skill or bypasses it—whether it adds the right kind of friction or removes all of it.
Whether I can hold this line is a separate question, and an honest one. The current is against me, the tools will only get more wonderful, robotics will make physical AI a reality, and my own children will make sure my thinking evolves whether I plan to or not. Ask me again in five years.
But when my 7-year-old asked who I was talking to, what unsettled me wasn’t her question. It was how easily that voice could become the one that answers all of hers.
So for now, in our house, bedtime stories still work the old way: let’s read something we’ve read a thousand times before—but oops, there’s another story we stumbled upon. Let’s feel the page on our fingers and the 30-year-old ink fading away, think of how their own father once read from this very book. Then put the book away, and jump into bed—to dream thoughts an AI hasn’t thought of yet.