The Error-Detection Layer
AI changes what you need to know. It doesn't change why you need to know it.
Nobody knows what to teach anymore.
I mean that seriously - not as criticism. If you’re designing a curriculum right now, the world is telling you that every job is about to change fundamentally. So you gather smart people in a room and try to produce something that will still make sense in four years, when your students graduate and walk into whatever comes next.
That’s an almost impossible task.
And yet the argument for teaching fundamentals has never been stronger - just for a different reason than before.
It used to be: you need to know how to do the work. Now it’s: you need to know enough to supervise the work. Those sound similar. They’re not. But they both require the same foundation.
A student who never learned to write code can’t tell an agent why its architecture is wrong. They can ask it to try again, sure. But they can’t steer it. A medical student who skipped the hard parts can’t catch a hallucinated drug interaction.
Domain knowledge is now the error-detection layer. Without it, you’re not really using AI. You’re just accepting what it gives you.
There’s a harder problem that most people don’t want to say out loud: if AI can pass most exams, how do you verify that any learning actually happened? You can ban the tools, but that’s just fighting reality. The entire assessment model needs rethinking.
The direction probably isn’t banning the tools. It’s making the reasoning visible - not “here is your answer” but “here is what I accepted, what I rejected, and why.” That requires domain knowledge to function. And it reveals quickly who has it.
That’s incredibly hard for a system that has functioned the same way for decades and has strong incentives to keep doing so.
What survives all of this is judgment. Not the word people use loosely - the actual thing.
Knowing which question to ask before you know the answer. Recognizing that something is wrong before you can prove it. Developing taste through years of exposure to good and bad work. Being wrong in low-stakes environments often enough that it starts to mean something.
That kind of judgment can be learned, but not really taught - not in the classical sense. It’s slow. It can’t be compressed or prompted or shortcut. It requires time, friction, repetition, and failure.
Which is, if universities are honest about it, probably what they were largely for in the first place.
Most of what they do is changing. That part doesn’t.
🙏
Be kind,
Manuel