I Let AI Teach Me for Six Weeks. My Recall Collapsed. Here's the 4-Rule Fix

For six weeks I studied quantum computing almost entirely through a chat window. Every session felt excellent. I asked good questions, got clear answers, followed the derivations, nodded along. I'd close the laptop feeling like I'd covered real ground.

Then a friend who actually does this for a living asked me to explain why a Hadamard gate matters. I opened my mouth and nothing structured came out. I could gesture at superposition. I could not derive anything, could not reconstruct the argument, could not tell him what I'd read three days earlier.

I hadn't learned quantum computing. I'd watched someone else understand it, forty times in a row.

The problem isn't AI. It's that AI removes the part where learning happens.

Here's the uncomfortable mechanism, and it has nothing to do with the model being wrong.

Learning is a byproduct of effortful retrieval and reconstruction. Not of exposure. Robert Bjork's research program at UCLA has spent forty years on this single distinction, and his term for it is "desirable difficulties." Conditions that make performance worse during practice, like spacing sessions out, mixing topics, testing yourself before you feel ready, reliably make retention better later. Conditions that make practice feel smooth, like rereading and following a clear worked explanation, reliably make retention worse.

A good language model is a desirable-difficulty removal machine. That is literally its product. It takes something confusing and makes it clear, instantly, in your preferred register, at whatever depth you ask for. Every friction point you would have had to grind through yourself gets sanded off before you touch it.

So you get fluency without encoding. The subjective sense of understanding is real. It's just measuring the model's competence, not yours.

Bjork and his colleagues have a name for that gap too. They call it an illusion of competence, and it shows up any time performance during study gets confused with learning. Rereading produces it. Highlighting produces it. Watching a lecture at 2x produces it. AI produces the strongest version I've ever encountered, because the explanation is custom-fit to your exact confusion, which makes it feel like your own thought.

What the studies actually found

I went looking for evidence after my quantum computing humiliation, partly hoping to find that I was an outlier. I wasn't.

The cleanest result is Bastani and colleagues, working with about a thousand high school students in Turkey. Students got access to GPT-4 while working through math practice problems. During practice, the AI group crushed it, performing substantially better than students without access. Then came the exam, closed book, no assistance. The group that had used unrestricted GPT-4 scored meaningfully worse than the control group who'd had nothing. Roughly seventeen percent worse, by their reporting.

The second half of that study is the part people skip, and it's the important part. The researchers also built a guardrailed version, a tutor that gave hints and prompts rather than answers. That group did not show the damage. Same model underneath. Different interaction pattern. The harm came from how it was used, not from its existence.

There's older work pointing the same direction. Sparrow, Liu and Wegner published a study in Science back in 2011 showing that when people expect information to remain available externally, they remember the information itself worse and remember where to find it better. They called it the Google effect. Your brain is a reasonable resource allocator. If retrieval is cheap and guaranteed, storage is a waste of metabolism.

More recently, a team at MIT's Media Lab ran EEG on people writing essays with and without an assistant, and found lower measured neural connectivity in the assisted group along with notably poor recall of what they'd just written, including inability to quote their own sentences. That one is a preprint with a small sample and I'd hold it loosely. But it rhymes with everything else.

None of this says AI is bad for learning. It says unstructured AI use is bad for learning, and structured AI use might be the best tutoring access most people will ever have. The whole game is in which one you're doing, and the two feel identical from the inside.

Why the feeling lies to you

Cognitive load theory, John Sweller's framework from the late eighties, splits the mental effort of a task into three buckets. There's the load intrinsic to the material's complexity. There's extraneous load, the wasted effort from bad instruction, confusing notation, hunting through a badly organized textbook. And there's germane load, the effort you spend actually building a mental model, the part that becomes learning.

AI is spectacular at eliminating extraneous load. It'll cut through terrible notation, bad textbook organization, and jargon that exists only to gatekeep. That's a genuine gift and I don't want to undersell it.

The problem is that it doesn't stop there. It eliminates germane load too, because germane load is made of exactly the same stuff as confusion, and from the outside they're indistinguishable. Struggling to reconstruct why a proof works looks a lot like being stuck on bad notation. The model helpfully resolves both.

And then there's the automation surface. When you offload a step, you don't just skip that step, you stop building the judgment that would let you evaluate it. A Microsoft and Carnegie Mellon survey of knowledge workers in 2025 found that higher confidence in a generative tool predicted less critical evaluation of its output. The trust and the scrutiny move in opposite directions, which is precisely backwards from what you'd want.

I've been writing software for ten years. I recognize this pattern from a different domain entirely. The engineers who never learned to read a stack trace because the IDE always jumped them straight to the line are the same engineers who are helpless the first time they're on a machine without the IDE. The tool was fine. The dependency was the issue.

Rule 1: AI goes before and after the thinking, never during

This is the single change that fixed most of it for me.

The sequence that destroys learning is: encounter problem, ask AI, understand answer, move on. The sequence that builds it is: encounter problem, attempt it badly and alone, then bring AI in.

This is not a productivity tip. It's a specific finding. Pre-testing, sometimes called the generation effect, shows that attempting to answer before you know the answer improves later retention of the correct answer, even when your attempt is completely wrong. Richland, Kornell and Kao demonstrated this repeatedly. Failing at a question first primes you to encode the answer when it arrives. Kornell's "unsuccessful retrieval enhances learning" work is the tidiest framing of it.

So the AI answer that lands on a mind that has already wrestled with the question gets encoded. The same answer, on a mind encountering the question for the first time, gets consumed.

Practically, this means I set a floor before I'm allowed to open a chat. For a concept, it's five minutes of writing down what I think is going on and where exactly I'm confused. For a problem, it's one full attempt, even a broken one. For a paper, it's reading the abstract and methods myself and forming a guess about the result.

The five minutes are unpleasant. That's the signal that they're working.

Rule 2: Make it ask, not answer

The Bastani study's guardrailed tutor is the entire lesson here, and you can rebuild it yourself with a paragraph of instruction.

I keep a standing prompt I paste at the top of any study session:

You are a demanding tutor. Never give me a direct answer or a full explanation. Respond only with questions, hints, or a single next step. If I'm wrong, tell me I'm wrong and ask a question that exposes why. If I ask you to just tell me, refuse once and ask again. Only explain after I've produced an attempt and you've made me revise it at least once.

It's an annoying way to use a model. Sessions take maybe forty percent longer. But the thing that comes out the other end is mine.

The second mode I use is inversion. Instead of asking the model to explain, I explain, and the model interrogates. This is the Feynman technique with a partner who doesn't get bored, doesn't let vagueness slide, and is available at 11pm. I write my explanation of a concept in plain language, paste it, and ask for the three weakest points and the questions an expert would ask that I couldn't answer.

That's the highest-value thing I have found to do with a language model in a learning context, full stop. Not explanation. Interrogation.

Rule 3: Anything AI hands you gets re-derived once, unaided

Reading an explanation and agreeing with it is not encoding. You have to reproduce it.

My rule is that any derivation, proof, argument structure, or code pattern I get from the model gets closed and rebuilt from scratch within the same session. Blank page. No looking. If I can't reconstruct it, I didn't get it, I just recognized it, and recognition collapses within days.

This is the testing effect, and it's about as well-established as anything in learning science. Roediger and Karpicke's 2006 studies are the canonical reference. Students who studied a passage and then took practice recall tests dramatically outperformed students who restudied the same passage, on delayed tests a week out. The restudy group felt more confident. They did worse. That confidence inversion is the same trap AI sets, just at higher intensity.

For code specifically, I've got a version of this that costs me maybe ten minutes a day. Anything an assistant writes for me that I'd be embarrassed not to understand gets retyped by hand, not pasted, and I have to be able to say what each line does before it goes in. Most of it I skip, because most code isn't worth learning. The stuff in my actual domain gets the treatment.

Rule 4: If it isn't on a retrieval schedule, it didn't happen

Even perfect encoding decays. Ebbinghaus mapped this in 1885 and nobody has overturned the basic shape since: the steepest losses happen in the first day or two, and each successful retrieval flattens the curve.

AI accelerates the decay problem for a specific reason. Volume. You can cover four times as much ground per hour in a chat window as you can with a textbook, which means you're generating four times as much material with the same fixed retention budget. The output of a good AI study session is a pile of understanding with an expiry date of about seventy-two hours.

So the session doesn't end when the understanding lands. It ends when the understanding is on a schedule. Five to ten questions per session, written as questions I have to answer from memory, not as notes I'll reread. Then spaced review: day one, day three, day seven, day sixteen, roughly. The precise intervals matter far less than their existence.

This is the least glamorous rule and it's the one that decides whether any of the rest compounds.

What AI is genuinely excellent at, and where I'd never give it up

I don't want to overcorrect into "AI bad." Six weeks of failure made me sharper about where it's unmatched:

Finding my actual confusion. I can describe a fuzzy half-understanding in messy language and get back a precise diagnosis of which specific misconception I'm holding. No textbook does that. No lecture does that. This alone is worth the price.

Generating practice. Infinite problems, at the difficulty I specify, on the exact subtopic I'm weak on. Historically the bottleneck for self-study was that you ran out of exercises or the ones you had didn't match your weak spots. That bottleneck is gone.

Translating between representations. Give me this in matrix form, now in circuit form, now as an analogy from software engineering. Dual coding, in Paivio's sense, where verbal and visual routes reinforce each other. Being able to demand a representation that fits how you already think is a real accelerant.

Being the examiner. Cheap, patient, unlimited testing. See rules 2 and 3.

Killing extraneous load. Bad notation, unexplained conventions, jargon, historical baggage. Cut it all. None of it is the thing you're trying to learn.

What I won't let it do: produce the first attempt, produce the summary I'm supposed to write, or answer a question I haven't tried to answer. Those three are where the learning lived.

The workflow I actually run

I'm currently a few months into what I've been calling the Side Quest Series, where I learn things outside my domain in public. Mandarin and quantum computing at the moment. Here's the loop, unglamorous version, ninety minutes:

0 to 10 minutes. Retrieval on the last three sessions. No notes. Answer the questions I wrote, out loud or on paper. Whatever I blank on goes to the top of today's list.

10 to 25 minutes. New material, unassisted. Textbook, paper, or problem set. I write down what I think is happening and where I'm stuck, in my own words. This section is uncomfortable by design.

25 to 55 minutes. AI in tutor mode with the standing prompt. It questions, hints, and refuses to hand me answers. I produce attempts and revise them. When something finally clicks I close the window and rebuild it on a blank page.

55 to 75 minutes. Practice problems generated to hit whatever I just failed at. Attempt first, check after.

75 to 90 minutes. Write five to ten retrieval questions from the session and get them onto the schedule. Then stop.

Lagos power situation permitting, I run this four or five days a week. The Mandarin version swaps the problem set for production practice, but the shape holds: attempt, then assist, then reconstruct, then schedule.

How to tell if it's rotting

The honest self-audit is one question, asked cold, twenty-four hours after a session: can you explain what you covered to another person, from memory, without the chat history open?

If yes, the session worked. If you can only recognize it when you see it again, you consumed, you didn't learn. There's no third state, and the feeling during the session tells you nothing about which one you're in. Mine felt fantastic for six straight weeks.

Two other tells. First, if you can't remember what you asked, only that the answers were good, you were a passenger. Second, if you find yourself reaching for the model before your own brain has produced anything at all, the dependency is already forming. That reflex is the thing to watch.

The last fifteen minutes are the ones people skip

Go back and look at my ninety-minute block. The final fifteen minutes, writing retrieval questions and putting them on a schedule, is the least interesting part of the session and the only part that determines whether the other seventy-five survive the week.

It's also the part I failed at for years, and the failure mode is predictable. At minute seventy-five you feel finished. The concept landed. Your brain has already filed the session as a success and has no interest in doing clerical work about it. So you skip it, and four days later you're back where I was with the Hadamard gate, holding a warm feeling and no structure.

I solved it by removing the authoring step. I use Edirae for the retention layer because it pulls cards straight out of the source material and my session notes, so what's left at minute seventy-five is editing rather than writing from scratch, which is a small enough ask that I actually do it. Anki does the same job if you're willing to author the cards yourself, and plenty of serious people prefer it that way. There's a 3-day free Pro trial at edirae.com if you want to test the automated version against your own discipline. Either tool is fine. Skipping the layer is not, because without a schedule everything above is just a more sophisticated way to forget.

---

FAQ

Does using AI actually make you dumber?

Unstructured use measurably hurts learning outcomes. The Bastani et al. study found students with unrestricted GPT-4 access during practice scored roughly seventeen percent worse on a later unassisted exam than students with no AI at all. The same study found that a hint-based tutor version produced no such harm. It's the interaction pattern, not the tool.

What's the single most important rule if I only adopt one?

Attempt before you ask. Five minutes of your own thinking before you open the chat window changes what happens to the answer when it arrives. Pre-testing research shows that even a wrong attempt improves retention of the correct answer.

Is it fine to use AI to summarize things I need to learn?

For material you need to know, no. Writing the summary is the learning. For triage, deciding whether a paper is worth your time, absolutely, that's a great use.

How do I stop the model from just giving me answers?

Instruct it not to, explicitly, at the top of the session, and include a rule about what to do when you beg. Models comply with a Socratic constraint reasonably well if you state it up front. Re-paste it when the conversation drifts, because it will drift.

Does this apply to coding assistants too?

Yes, with a filter. Most code isn't worth learning and offloading it is fine. Code in the domain you're trying to get good at is different. My rule is that anything in my core domain gets retyped by hand and explained line by line before it goes in.

How much slower is this?

Sessions run roughly forty percent longer than pure AI consumption. Compared to learning without any AI, it's still a large net speedup, mostly from targeted practice and instant diagnosis of confusion. You're trading against consumption speed, not against traditional study speed.

Do I still need spaced repetition if the AI can just re-explain things?

Yes, and more than before. The ability to re-explain instantly is exactly what tells your brain not to store it. Cheap external retrieval reduces internal storage, which is the Google effect Sparrow and colleagues documented in 2011.

Is the MIT brain-scan study reliable?

Treat it as suggestive. Small sample, preprint, and EEG connectivity is a coarse measure. It's consistent with the behavioral evidence, which is stronger, but I wouldn't build an argument on it alone.

---