Not Forgetting How to Think

You read an excellent explanation. Everything is clear, it all connects, you close the tab satisfied. Two days later you need to use it, and what's left in your head is the memory that the explanation was good.

Anyone who has studied properly recognises this: it's the illusion of competence from the How to Learn sphere, the one produced by rereading with a highlighter. What's new is that AI produces it on demand, at any hour, about anything — perfectly clear, perfectly adapted to you, and perfectly frictionless.

That's the honest version of the worry. Not that the tool makes anyone stupid, but that it's extremely good at removing effort, and some of the effort it removes was the part doing the work.

Three different risks

They get lumped together and they have different fixes.

The illusion of understanding. A good explanation feels like knowledge. It isn't — it's someone else's chunk, borrowed. You'll discover the difference at the moment you have to produce it without the tab open.

Skill atrophy. Skills are built by reps, and you only get reps on the part you actually do. Hand over the part that is the skill, and you keep the output while losing the growth. This is invisible for months, then obvious.

Convergence to the median. The model's default is the average of what's been written. Take that as your starting point often enough and your work drifts toward the middle — competent, unobjectionable, indistinguishable. The danger here isn't error, it's sameness.

What's actually been measured

Two studies worth knowing, with their caveats.

In 2025, Nataliya Kosmyna's group at the MIT Media Lab ran an EEG study on essay writing: one group used an LLM, one used search, one worked unaided. The LLM group showed weaker neural connectivity during the task and — the striking bit — often couldn't quote from the essay they had just submitted. Handle this one carefully: it's a small, short-term study of one task, and it was over-reported as "AI rots your brain." What it does support is narrower and more useful: text produced through a model is worse encoded as your knowledge than text you generated yourself. Which is exactly what the generation effect from How to Learn predicts.

The second is more practical. Also in 2025, Hao-Ping Lee and colleagues at Microsoft Research and Carnegie Mellon surveyed several hundred knowledge workers about their real AI use, and found the pattern that should shape your habits: the more confidence people had in the AI, the less critical thinking effort they reported — and the more confidence they had in their own expertise, the more. The tool doesn't determine the outcome. The relationship you have with it does.

The rule

Almost all of this is fixed by one ordering habit.

Your move first, then its move.

Before asking, produce your own version: the answer as you understand it, your draft, your hypothesis, your list. It can be bad. It can be three bullet points. What matters is that it exists before you see the model's.

Three things happen at once. You do the retrieval — the single most effective learning operation there is, and the one AI most easily displaces. You turn the model's answer from an anchor into a comparison, which is a completely different cognitive event: instead of accepting a frame, you're diffing two frames, and the gaps are where the learning is. And you keep your own angle, since you committed to it before the median arrived.

The reverse order — ask first, then think — feels faster and produces neither learning nor originality. It's also the order almost everyone defaults to.

Goal skills and overhead skills

Here's the line that dissolves most of the agonising, and it's yours to draw, not a rule anyone else can set.

Goal skills are the ones you're trying to have: for a developer, designing systems; for a writer, writing; for an analyst, reading data. Delegating a goal skill costs you the thing you were building — even when the output is fine.

Overhead skills are the ones you'd happily never develop: boilerplate, formatting, a language you use twice a year, the summary nobody will read closely. Delegating these is pure gain, and refusing to on principle is just ceremony.

The same task sits in different categories for different people. A developer writing marketing copy and a marketer writing marketing copy are not in the same situation, and neither is wrong. What breaks is delegating goal skills by accident, one convenient turn at a time, because in the moment it's indistinguishable from delegating overhead.

The test is a question you can answer honestly: am I trying to get better at this, or trying to be done with it?

The self-check

Atrophy is silent, so audit it deliberately. Once in a while, do a real task in your goal skill without any assistance and see how it goes. Not to prove a point — to get information.

If it's slower but fine, you automated something. If you can't start, or you can't hold the whole problem in your head any more, you didn't automate it — you replaced it, and the replacement only exists while the tool is available.

Use it to add difficulty, not remove it

The deepest connection to How to Learn is this: Robert Bjork's desirable difficulties — the effort of retrieval, spacing, having to work things out — are what produce durable learning. AI is, by nature, a difficulty-removal machine. Used thoughtlessly it strips out the desirable ones along with the pointless ones.

But it can just as easily be pointed the other way, and this is where it becomes genuinely superb for learning:

• Make it quiz you instead of explain: "ask me questions on this topic, one at a time, and don't tell me the answer until I've tried."
• Explain to it, then have it find the holes. You do the generating; it does the marking.
• Ask it to argue against you — the counter-case habit from article 3, aimed at your thinking rather than at its own.
• Ask for problems, not solutions. "Give me five exercises that would show whether I actually understand this."
• Ask what you're not considering. This is the direct antidote to convergence: use the median as a map of the territory you should be deviating from, not as the destination.

Stage 4 turns this into a full method for learning with AI. The principle is already here: it's most valuable when it's making you work, not when it's saving you from working.

Who decides

One last thing, and it isn't about learning.

A model has no stake in your outcome and no capacity to be accountable for it. When you take its recommendation, the decision remains yours completely — the reasoning was outsourced, the responsibility wasn't. "The AI said so" isn't an explanation you'll be able to offer anyone, including yourself, when it matters.

This isn't an argument for using it less. It's the reason the previous articles insisted you keep the ability to evaluate: judgement is the one part of the job that cannot be delegated, because there's no one on the other end to take it.

In practice

Always your move first. Draft, hypothesis or bullet points before you ask. Even a bad one.

Never state your conclusion before asking for analysis — you'll get agreement instead of thought.

Know which of your skills are goals and which are overhead. Delegate the second freely.

Audit a goal skill unaided from time to time. Slower is fine; unable is a finding.

Point it at difficulty: quiz me, mark me, argue against me, give me problems.

Use the median as a baseline to beat, not as an answer to submit.

Check yourself

Close the article and answer in your own words:

1. What are the three distinct risks, and why does each need a different response?
2. What does the MIT EEG study support, and what does it not support?
3. What was the pattern in the Microsoft/CMU survey, and what does it imply about habits rather than tools?
4. Why does "your move first" produce learning that "ask first" doesn't — name all three effects.
5. How do you tell a goal skill from an overhead skill, and what's the honest test?
6. What's the difference between automating a skill and replacing it, and how would you detect the second?
7. How can AI add desirable difficulty rather than remove it?

In short

• Three separate risks: the illusion of understanding, atrophy of skills you delegate, and convergence toward the median.
• MIT's 2025 EEG study: text produced through a model is encoded as your own knowledge less well than text you generated — a small study, over-reported, but consistent with the generation effect.
• Microsoft/CMU 2025: confidence in the AI predicted less critical thinking; confidence in one's own expertise predicted more. The habit decides the outcome, not the tool.
• Your move first, then its move: you get retrieval practice, you get a comparison instead of an anchor, and you keep your own angle.
• Delegate overhead skills freely; guard goal skills. The honest test is whether you want to get better at this or be done with it.
• Audit unaided occasionally. Slower means automated; unable means replaced.
• Desirable difficulties are what teach. Aim the model at adding them — quiz me, mark me, argue with me, give me problems — not at removing all of them.
• Reasoning can be outsourced; responsibility can't.
• Your move first, then its move.