When AI can do most of the thinking, what is learning still for?

  <p>It is a fair question, and students ask it more honestly than their teachers would like: if a model can write the essay, solve the integral and explain the mechanism in seconds, what exactly am I learning for? Pretending the tools aren’t good is not an answer. They are good, and getting better. The honest reply starts by admitting that, then noticing what the question quietly assumes.</p>

  <h2>Getting an answer and judging an answer are different skills</h2>
  <p>A model will hand you an answer to almost anything. What it will not do, reliably, is tell you when that answer is wrong. In my own field a model will happily explain a single-molecule measurement and get the physics subtly, confidently wrong — the kind of wrong that looks right unless you have built the judgement to catch it. That judgement does not arrive by reading correct answers. It is built by working problems yourself, getting them wrong, and feeling where the reasoning gives way. Outsource the working, and you never build the judge.</p>

  <h2>Learning was never really about storage</h2>
  <p>The fear assumes learning is the storage of information, and that a better external store makes the internal one pointless. But a chemist does not carry the periodic table because they cannot look it up. They carry it because fluency frees attention for the harder thing: seeing the pattern, asking the next question. The facts you internalise become the ground you think on. Take away the ground and there is nothing to stand on while you decide whether the machine’s answer holds.</p>

  <h2>The skill rising in value is taste</h2>
  <p>When answers are cheap, the scarce thing becomes knowing which question is worth asking and which answer is worth trusting. Call it judgement, call it taste. It is what separates a researcher who uses a model to move faster from one quietly led off a cliff by a fluent paragraph. And taste is not downloadable. It grows the slow way, through your own hours of being confused and then less confused, which is exactly the process a model offers to skip for you.</p>

  <p>So learn differently, not less. Use the tools (they are extraordinary), but use them the way a researcher uses an instrument: to go further, not to avoid the walk. The point of learning was never to become a slower computer. It was to become someone who can tell when the fast one is wrong.</p>