How to learn better with AI

Almost everyone uses AI to learn the same way: open a chat, type "explain neural networks," read what comes back. The answer arrives clean, ordered, well-paced. It makes sense on the first pass.

That is exactly the problem.

Clear writing is not your understanding

When something reads easily, your brain treats that ease as evidence that it landed. It's an honest mistake — effortless processing feels identical to knowing. Which is how you can finish an explanation nodding along, close the tab, and be unable to reconstruct the idea half an hour later.

A badly written textbook told you the truth immediately: you'd read the same paragraph three times and still have nothing. The friction was the signal. AI removes the friction, and takes the signal with it.

The practical consequence is uncomfortable. If your study method is asking for explanations and reading them, what you're measuring is how well the model writes, not how much you're learning.

Ask it to test you, not to tell you

The highest-return change is to stop treating the chat as an encyclopedia and start treating it as an examiner. Pulling something out of memory fixes it far better than reading it again, and an AI is a study partner with unlimited patience for being wrong at.

Quiz me on [topic]. One question at a time. Do not give me the answer until I've attempted it. When I'm wrong, tell me the exact point where my reasoning broke, not just the correct answer.

Those last two sentences do all the work. Without them the model hands you the question and the answer in a single message, which is precisely what you don't want — you're back to reading instead of recalling.

Get the syllabus before the conversation

A chat is reactive. You only learn what occurs to you to ask, and what you don't know exists is what you will never think to ask about. That produces patchwork knowledge: five disconnected concepts, no map, and a vague sense that something's missing without knowing what.

Before your first question, get an outline. Ask for eight or ten modules ordered from foundational to advanced, one line each on what it covers, and treat it as a script. There are tools built for exactly this: Sapiens generates the syllabus from a topic and a specific audience, then writes the lesson for each module — the step almost nobody has the patience to do by hand, one prompt at a time.

The tool isn't the point. The order is: structure first, questions second.

Explain it back and ask to be torn apart

When you think you've got it, write it out in your own words and hand it over. But say explicitly what you want, because models default to encouraging and encouragement is worthless here:

I'm going to explain [X] in my own words. Don't be encouraging and don't soften your corrections. Find the exact point where my explanation breaks down, goes vague, or hides a logical jump, and say so plainly.

This is the most useful minute of the session and the least pleasant. The cracks always show up in the same place: the sentences where you wrote "basically" or "somehow."

Tell it who you are

A generic explanation is aimed at nobody. The same question deserves different answers depending on what you already know, and the model has no way to guess.

Compare "explain embeddings" with "explain embeddings to someone who's been writing software for ten years but hasn't touched linear algebra since university." The second version skips three paragraphs of preamble you don't need and starts where you actually are.

Distrust anything you can't check

Models are wrong with total confidence in their voice. There's no hesitation to warn you, and that's their most dangerous flaw as a teacher.

On settled material — school mathematics, grammar, the fundamentals of a programming language, well-established history — it's an excellent tutor, because that content appeared thousands of times in what it learned from. The closer you get to the recent, the contested, or the very niche, the further reliability drops.

Simple rule: if money, health, a legal decision, or code that reaches production depends on it, check against a source that isn't the chat.

Come back in three days

One session isn't learning. It's the minimum condition for learning later.

Close the laptop, let three days pass, then write down everything you remember without looking at anything. What comes out is what you actually learned; the rest was reading. Then go back to the material only to patch the gaps you just found.

One detail matters more than it sounds like it should: get the content out of the chat window. Reading a lesson in an actual reader — no blinking cursor, no temptation to ask something else mid-paragraph — resembles studying. Reading it in the same tab where twenty other conversations are open resembles browsing. If you study on a Kindle, sending the material to the device takes about thirty seconds and changes the quality of the reading completely.

The old advice, minus the old excuse

None of this is new, and that's the interesting part. Test yourself, explain out loud, space your review, check your sources, follow an order: these are the same things that worked when all you had was books and a teacher with office hours on Tuesdays.

AI doesn't change how learning works. What it removes is the last reasonable excuse — not having anyone on the other side with infinite patience to quiz you on the same thing twelve times.