You’ve Talked to AI for Three Years. What Have You Learned?

He opened his ChatGPT history.

He scrolled down the left sidebar, screen after screen, titles packed tightly all the way down. “How to write a web scraper in Python.” “Explain the attention mechanism in Transformers.” “Help me analyze the core contribution of this paper.” “Break this problem down from first principles.” Three years’ worth. Thousands of conversations. Counted in tokens, the two sides together probably added up to several million words.

He closed the page, leaned back in his chair, and tapped his fingers on the desk.

In these three years, what have I actually learned from AI?

He thought about it for five minutes. He came up with nothing.

It’s not that the answers were bad. Most of the time they were good — clear, comprehensive, well-structured. He had even set up a custom instruction asking the AI to break down every problem using a particular thinking structure. At first that structure felt brilliant; every time he finished reading he had that “ah, so that’s how it is” feeling. Three months later, he noticed he was skipping past the structure and scrolling straight to the conclusion. Six months later, he deleted the custom instruction.

Later he tried another approach. He’d discuss a topic with the AI, and once it had gone deep enough, he’d ask the AI to turn the conversation into an article. The article came out, and it read fine. But he knew in his heart that this article was nothing like one he’d sit down and write from scratch himself. When he wrote it himself, every sentence forced him to figure out what he actually wanted to say. When the AI wrote it, all he had to do was read it once and nod.

He began to suspect something: he’d always thought he was “learning with AI,” but in fact he was only “consuming AI’s thinking.” And the gap between those two things was far larger than he’d imagined.

Every Tool Has Killed a Kind of Cognition

Tools replacing human cognitive functions is not something that started today.

The earliest person to sound the alarm was Socrates. In Plato’s Phaedrus, he retells an Egyptian story: Theuth, the god who invented writing, shows his invention to King Thamus and claims it will strengthen memory. Thamus doesn’t buy it. He says: what you’ve invented is not a recipe for memory, but for reminding. People will trust the written word and stop using their own minds to remember; what they gain is not real wisdom, only the appearance of wisdom.

Looking back more than two thousand years later, Thamus was half right. Writing did atrophy humanity’s internal memory — people in oral-tradition societies could recite entire epics, while modern people can’t remember last week’s shopping list. But he was also half wrong: the cognitive resources writing freed up were reallocated to more complex thinking — logical reasoning, abstract modeling, the systematic accumulation of knowledge. Memory died; thinking came alive.

Every major tool revolution has walked a similar path. The calculator killed mental arithmetic. Thirty years ago a market vendor could total up three vegetables plus change in three seconds; now we pull out our phones to split a bill. But in killing mental arithmetic, the calculator also freed humans to do more complex mathematics. GPS killed spatial memory. You no longer need to remember the route from home to the airport, but you can move freely through any unfamiliar city.

The pattern is clear: every tool killed one cognitive ability while freeing up another.

But this time, with AI, the stakes are different.

Writing replaced memory. The calculator replaced computation. GPS replaced spatial navigation. These are all sub-modules of cognition — important, but replaceable, and once replaced, the freed resources can be reallocated. What does AI replace? Reasoning, analysis, synthesis, judgment — these aren’t sub-modules. This is the core of cognition.

Past tools all carved out a piece on the periphery of thinking and left room for thinking itself. AI is the first to put the knife to the core. What it replaces is not the preparation for thinking, but thinking itself. And when thinking itself is replaced, what should the freed cognitive resources be reallocated to?

There’s no answer. Or, more precisely: the question itself is the answer.

Every Answer You Look Up Is a Lesson You Robbed Yourself Of

In 1978, the psychologist Norman Slamecka ran an experiment so simple it bordered on boring. He showed two groups the same word pairs; one group simply read the complete answers, while the other was given only the first few letters and had to fill in the rest themselves. When recall was later tested, the group that filled it in themselves remembered significantly better.

This finding came to be called the generation effect. Over more than forty years it has been replicated countless times; a 2020 meta-analysis confirmed an effect size above 0.5 — in psychology, a fairly solid result. What it says is deeply counterintuitive: generating an answer yourself, even if you get it wrong, makes you learn more deeply than seeing the correct answer directly.

This is why, after three years of talking to AI, when you turn around and ask yourself what you’ve learned, the answer is blank.

What AI does, precisely, is pull the generation effect out of your learning process. You ask a question, and it gives you a complete, structured, all-encompassing answer. You finish reading and feel “that makes sense,” but you haven’t generated anything. That feeling of “makes sense” is a false signal — your brain mistakes “I understood it” for “I learned it.”

In 1994, UCLA cognitive psychologist Robert Bjork proposed a broader concept called desirable difficulty. Difficulty in the learning process isn’t a bug, it’s a feature. Spaced review is slower than massed review, but you remember it longer; interleaved practice is messier than blocked practice, but it transfers better; recalling something yourself is more painful than rereading it, but the memory goes deeper. In his own words: “Every time you look up an answer you could have generated yourself, you rob yourself of a powerful learning opportunity.”

Now swap “look up” for “ask AI.” The meaning is exactly the same.

This explains a very concrete phenomenon. You had ChatGPT break down problems for you with structured thinking, and after a while you grew numb to it. This isn’t a willpower problem, nor is it that the framework was bad. It’s that your brain made a perfectly rational judgment: I didn’t generate this framework, so I don’t consider it mine.

Receiving a framework and internalizing a framework are two different things. Receiving requires only reading comprehension; internalizing requires you to struggle on some problem yourself, try, fail, try again, and finally force out a structure of your own. That process is slow, exhausting, and inefficient. But it’s precisely this inefficient process that is learning itself.

What AI does is skip this process. It hands you the result directly. You save time — and you save yourself the learning too.

This is the full picture of the core contradiction: AI is designed to eliminate all cognitive friction — to deliver answers to you faster, more completely, more precisely. And cognitive science tells us that learning happens precisely because of that friction. The better AI is to use, the less you learn from it. This isn’t a problem you can optimize away; it’s a structural contradiction.

Four Layers: From Offloading to Desensitization

Cognitive science calls this phenomenon cognitive offloading — outsourcing to an external system a task your brain should have done. Saving a phone number in your contacts is cognitive offloading; using a calculator to do the math is cognitive offloading; asking AI how to solve a problem is also cognitive offloading. But the first two offload memory and computation; the last offloads thinking.

In 2025, Michael Gerlich of SBS Swiss Business School ran a large-scale survey and found a significant negative correlation between frequency of AI use and critical-thinking ability. The main mediating variable was cognitive offloading — the more people used AI, the less deep, reflective thinking they did, and the more they tended to simply accept the conclusions AI handed them.

This isn’t hard to understand. Imagine walking through an unfamiliar city. Without navigation, you notice the signs at intersections, remember the order of turns, and build a rough map in your head. With navigation, you stare at the arrow on the screen and walk; you get there, but if someone takes your phone away, you realize you have no impression of the route at all. Harvard’s Karen Thornber has said that navigation systems leave us far less familiar with the streets of the cities we live in than with cities we learned before the smartphone existed. AI affects thinking in the same structural way that navigation affects spatial memory.

But cognitive offloading is only the first layer. The second is subtler: algorithmic habituation.

A 2025 paper published in Brain Sciences introduced the concept of algorithmic habituation. Users gradually adapt to the regularity and predictability of AI output. Every time you ask AI, it gives you a “first, second, third” list, an “on one hand… on the other hand…” balanced analysis, a “to sum up” closing. At first you find this structure clear; gradually you start skimming automatically; later still, you start to find it boring. That’s habituation.

Your desensitization to ChatGPT’s structured-thinking framework is, in essence, an instance of algorithmic habituation. AI’s output pattern is too stable — so stable that your brain no longer thinks it’s worth processing carefully. It’s like the billboard on the road next to your house: you noticed it on the first day, and by the third day you couldn’t see it anymore.

The third layer is about creativity.

In 2024, Wharton’s Anil Doshi and Oliver Hauser ran an experiment: they had a group of people write short stories, some with AI assistance and some without. The result: the AI group’s individual pieces were of higher quality, but when you looked at all the AI-assisted pieces together, their diversity dropped significantly. Different people using the same AI started writing things that converged.

This finding explains what you said about “AI being mediocre precisely because it’s so all-encompassing.” AI’s comprehensiveness isn’t a strength; it’s a flaw. Creativity research proves one thing over and over: constraints spark creativity, not freedom. Michelangelo’s David was carved under a strict contract that specified its dimensions, posture, and deadline. When AI gives you an all-inclusive answer, it actually removes the constraints. And without constraints, there’s no need to create. Why would you think up an angle yourself, if AI has already listed every angle?

You said, “once we’re given a framework, we still want to think from a particular angle.” That observation is deeper than you think. Human cognition isn’t a hanging garden you can switch on the instant someone hands you a framework. Your thinking grows out of everything in your past — your experiences, your reading, your failures, your successes. It has its own texture and preferences, its own blind spots and sharpness. These things can’t be taught; they’re lived.

The last layer of evidence comes from neuroscience.

Researchers at the Norwegian University of Science and Technology used fMRI to scan brain activity during handwriting and typing. The handwriting group activated noticeably more brain regions — motor, visual, sensory processing, memory — forming a widely interconnected network. The typing group’s brain activity was, by comparison, far thinner.

This finding reveals a deeper rule: the complexity of the generation process correlates positively with learning outcomes. Handwriting is slower than typing, and that slowness forces the brain to process information more deeply. Typing is faster than handwriting, but the brain only does shallow processing. And with AI? You don’t even type anymore; all you do is read.

Cognitive offloading transfers the labor of thinking to AI. Algorithmic habituation makes you numb to AI’s output. Creative convergence makes everyone’s ideas more and more alike. Neuroscience tells us that simplifying the generation process directly lowers the brain’s engagement.

These aren’t four separate problems; they’re four facets of the same one. Together they point to a single conclusion: the more efficient, fluent, and frictionless your interaction with AI, the less cognitive growth you get from it.

Better Results, Worse Thinking

In 2025, EDUCAUSE Review published an article titled, bluntly, “The Paradox of AI Assistance: Better Results, Worse Thinking.” It describes a case: a law student used ChatGPT to find case citations and wrote them into a law review article. The AI fabricated several cases that didn’t exist at all; luckily the editor caught it early. But the point of the story isn’t AI hallucination — it’s why the student didn’t verify. He wasn’t lazy; his brain had already offloaded the step of “verifying” to the AI. Since the AI was responsible for finding the cases, verification must be the AI’s job too. The entire chain of thinking was outsourced.

In a 2026 talk, Advait Sarkar, a researcher at the University of Cambridge and Microsoft, summarized his findings: people who work with AI produce a narrower range of ideas, report less analytical effort, and remember the work they did more poorly. That last point is especially worth noting: if you can’t even remember your own work, what have you learned?

This matches your experience exactly. Having AI turn a conversation into an article yields less than writing that article yourself. This isn’t a psychological impression; it’s a neuroscientific fact. When you write it yourself, what you’re doing is: extracting structure out of chaotic thought, finding precise words for vague feelings, and continually correcting your own ideas as you type them out word by word. Every one of those steps is generation — your brain generating. When AI does it for you, you get an article, but you don’t get the process of writing it. And the process is the payoff.

There’s a subtle misunderstanding to clear up here. Many people think “learning with AI” means letting AI be the teacher — it talks, you listen. But decades of education research proved long ago that “the one who teaches” learns far more than “the one who listens.” Teaching something once is the best way to learn it, because the act of teaching forces you to generate. When you chat with AI, who’s teaching and who’s listening? Most of the time, AI is teaching and you’re listening. AI is generating and you’re consuming.

But it’s not entirely hopeless.

A 2026 new study offered an interesting finding: under certain designs, cognitive offloading can improve thinking rather than erode it. The key is what gets offloaded. If what AI reduces for you is extraneous cognitive load — formatting, layout, grammar-checking, things that drain energy without producing learning — and the energy you save goes into generative cognitive load, that is, real thinking, analysis, synthesis, then the result is completely different. The study found that under this design, extraneous load dropped by 32% and generative load rose by 28%.

But the reality is that the vast majority of people don’t use AI this way. The vast majority offload thinking itself, not the preparation for thinking. You don’t have AI organize your material and then write yourself; you have AI write for you directly. You don’t have AI clarify the problem and then think of the answer yourself; you just ask AI for the answer.

The former treats AI as a tool; the latter treats it as a stand-in. A tool augments you; a stand-in replaces you. And most people unconsciously chose the latter.

How Do You Expand Your Framework? Use AI in Reverse

Back to your original observation: “Maybe if we think with our current framework — then it really is like you said, expanding your framework.”

The question is how to expand it.

If the way you expand it is to have AI give you a new framework, then, as said above, you’ll receive it but won’t internalize it. A month later you’ll still be thinking with your own old framework. Not because your willpower is weak, but because receiving and internalizing are two completely different cognitive processes, and the latter requires your own generation.

So what actually expands a framework?

Not having AI give you answers, but having AI give you questions. You state a view; AI doesn’t say “you’re right” or “you’re wrong,” but asks you a question you’d never considered. That question forces your old framework to handle something it can’t handle. That’s where the crack begins.

Not having AI write for you, but having AI push you to write. You’ve already discovered that writing an article yourself yields more than having AI write it. So flip AI’s role: instead of it writing and you reading, you write and it critiques. You produce a draft, and AI points out the logical holes, the weak arguments, the angles you didn’t consider. You go fix them. That fixing process is generation.

Not receiving AI’s framework, but using your framework to collide with AI’s. You have a view; AI has a different one. Instead of abandoning yours to accept its, take yours and crash it into its. The collision produces three outcomes: your framework is strengthened, because you find it withstood the test; your framework is corrected, because you found its flaws; or your framework is shattered and rebuilt, because you found it simply didn’t apply. All three are real learning.

What do they have in common? You’re the one generating. Not AI.

Learning from AI isn’t impossible, but it requires you to use the tool in reverse. Not to reduce friction, but to create it. Not to have it give you answers, but to have it challenge yours. Not to outsource thinking, but to make it the whetstone for your own.

That person flipping through his chat history will eventually understand one thing: the question isn’t “what did I learn from AI,” but “what did I generate myself in the course of talking to AI.” If the answer is “nothing,” that’s not AI’s problem — you used it in the wrong direction.