There’s a strange gap that opens up when you learn with AI. The session feels productive. The explanations are clear, the summaries are tidy, the answers arrive in seconds. And then, a week later, you reach for what you supposedly “learned” and find almost nothing there.
That gap has a name in the research literature: it’s the difference between performance (how well things go while you’re working) and learning (what durably sticks afterwards). They are not the same thing, and they can even move in opposite directions. The concept that best explains why — and that turns out to be unusually useful for anyone learning alongside AI — is cognitive load.
This isn’t a productivity buzzword. Cognitive Load Theory came out of decades of instructional-design research (most associated with the psychologist John Sweller, from the late 1980s onward). Like most psychology, it has its debates and its limits, which we’ll be honest about. But its core picture is well-supported and genuinely practical, and it reframes the single most important question about studying with AI: not “did this feel easy?” but “did the effort land in the right place?”
Working memory is the bottleneck
Start with the constraint everything else hangs on. Your working memory — the mental workspace where you actively hold and manipulate information right now — is small and short-lived. The classic figure is “about seven items,” later revised down to something closer to four chunks for genuinely novel material, and whatever the exact number, the headline is the same: it’s tiny, and anything you don’t actively rehearse fades in seconds.
Your long-term memory, by contrast, is effectively limitless. Learning, in this model, is the process of building organised structures (schemas) in long-term memory so that what once took up four scarce working-memory slots eventually takes up one — or runs automatically with none to spare. An experienced reader doesn’t sound out letters; a fluent driver doesn’t think about the clutch. The clutter got compiled into something the mind can run cheaply.
The whole game of learning, then, is managing the traffic through that narrow working-memory bottleneck while the schemas get built. Cognitive Load Theory splits that traffic into three kinds.
The three kinds of load
Intrinsic load is the difficulty inherent in the material itself — how many elements you have to hold in mind at once because they interact. Learning individual vocabulary words is low intrinsic load; learning grammar that ties those words together is higher. You can’t wish this away, but you can sequence it: master the parts before the thing that combines them.
Extraneous load is the effort wasted on everything that isn’t the actual learning — a confusing layout, a badly worded problem, jargon you have to decode, flipping between two sources to assemble one idea, hunting for the information you need. This is pure overhead. Every slot it occupies is a slot stolen from understanding. Reducing extraneous load is almost always good.
Germane load is the effortful part that is the learning — the mental work of connecting new information to what you already know, noticing patterns, organising it, building the schema. This is the load you actually want. It feels like effort because it is effort, and that effort is the thing doing the construction.
Here’s the crux: working memory is a fixed budget shared across all three. If extraneous load eats most of it, there’s nothing left for the germane work, and little gets learned no matter how long you sit there. Clear away the extraneous clutter and you free up room for the effort that counts.
Now put AI into that picture, and you can see exactly where it helps and exactly where it quietly betrays you.
Where AI genuinely helps: it slashes extraneous load
This is the real and legitimate win, and it’s worth saying plainly because the rest of this piece is a caution. Used well, AI is one of the best extraneous-load reducers ever built.
It can take a wall of dense, badly organised text and give you a clean structure to hang the ideas on. It can define a term inline so you don’t lose your thread hunting through a glossary. It can turn an abstract explanation into a concrete worked example — and worked examples are one of the most robust findings in this whole literature for beginners, precisely because they let a novice study the structure of a solution instead of flailing through trial and error. It can hold the boring scaffolding so your scarce attention goes to the idea.
If you’ve ever been stuck not because the concept was hard but because the explanation was a mess, you’ve felt extraneous load. AI is very good at clearing that particular fog. (We’ve written before about the AI tools with the highest cognitive return and about how to actually learn with AI rather than just consume it.)
Where AI quietly sabotages learning: it offloads the germane load too
The trap is that the same tool that removes the wasted effort will, just as happily, remove the productive effort — and from the outside, the two feel identical. Both make the task easier. Only one of them was supposed to stay hard.
When you ask AI to summarise the chapter, it did the organising — the exact schema-building work that would have been yours. When you ask it to solve the problem, it ran the reasoning your brain needed the reps on. When you ask it to write the first draft, it did the retrieving, sequencing and connecting that is, cognitively, the learning. You got the output. Your long-term memory got almost nothing, because nothing passed through the construction step.
This is why AI study can feel so good and deliver so little. Fluency is seductive: watching a clear explanation produces a strong, confident sense that you understand it — a feeling researchers call the illusion of explanatory depth. But recognising a tidy answer is a world away from being able to generate it yourself. The ease is real. The learning is the thing that got skipped.
There’s a well-established flip side that makes this concrete. A family of findings known as desirable difficulties (from the work of Robert and Elizabeth Bjork) shows that certain things which make studying harder and slower in the moment make the learning stronger and more durable — struggling to retrieve an answer before checking it (the testing effect), spacing practice out, mixing problem types. The “generation effect” is the same story: you remember what you worked to produce far better than what you merely read. AI, pointed carelessly, is a machine for removing exactly these difficulties — and with them, the benefit.
How to tell the two apart (the one question to ask)
You don’t need to track three kinds of load in your head. You need one habit: before you hand a task to AI, ask “Is this effort the obstacle, or is this effort the point?”
If the effort is overhead — decoding a tangled explanation, formatting, finding the relevant passage, getting unstuck so you can continue — offload it freely. That’s extraneous load, and clearing it is what AI is for.
If the effort is the thing you’re trying to get better at — recalling it, reasoning through it, explaining it, composing it — keep it. That’s germane load, and handing it to AI hands away the learning.
A few ways that cashes out in practice:
- Retrieve first, then check. Try to produce the answer or explanation yourself before you ask the AI. Then use it to verify and correct. You get the germane effort and the clean feedback. (Pairing this with the testing effect is most of the value right there.)
- Make it ask, not tell. Have the AI quiz you, pose problems, or demand you explain your reasoning — rather than hand you conclusions. Flip it from an answer machine into a sparring partner.
- Use it to teach the structure, then do the reps yourself. A worked example or a clear scaffold is a great starting point for a beginner; just don’t let the worked examples be your practice. Study one, then solve the next one unaided.
- Explain it back without the screen. If you can teach the idea in your own words with the AI closed, it’s in long-term memory. If you can’t, you borrowed fluency you don’t own yet. (This is the Feynman move.)
- Watch for the “that was easy” signal. With genuine learning, ease is something you earn on the way out, not something you feel the whole way through. A session that was frictionless start to finish often means the friction — the learning — happened in the model, not in you.
The honest caveats
Cognitive Load Theory is a model, not a law of nature. The three load types are hard to measure cleanly and don’t always separate as neatly in practice as they do on a diagram; “germane load” in particular has been reframed and argued over in the research. The precise capacity of working memory depends on what you’re counting. And none of this means difficulty is good in itself — load that overwhelms a true beginner just causes failure and discouragement, which is why reducing extraneous load matters so much. The art is matching the challenge to the learner, not maximising struggle.
But you don’t need the theory to be exact for the practical point to hold, because it rests on something that isn’t in dispute: your working memory is small, your attention is finite, and the effort of building understanding has to happen in your head to count. AI can clear the clutter around that effort, or it can quietly perform the effort for you. The tool won’t tell you which it’s doing. You have to.
Learn with AI as a thinking partner that removes the noise and holds you to the work — not as an engine you outsource the thinking to. Keep the effort that builds the brain on your side of the screen.
This article is for general educational purposes and isn’t medical or psychological advice. If you want to go deeper, our courses turn this kind of science into daily practice.