Thinking in the age of AI: how not to forget how to think for yourself

The intellectual revolution and a new problem

Every major information technology has changed not only what people know but how they think. Writing freed memory from having to hold entire epics and made possible complex reasoning that can't be held in the head. The printing press turned knowledge from a rarity into a commodity. The internet and search engines made any fact available in seconds. Each time there were skeptics — and each time they were partly right and partly wrong.

The best-known example is the Phaedrus, where Plato has Socrates voice a worry: writing will give people "the appearance of wisdom"; they will rely on external marks rather than their own memory. Today this is often cited as a model of needless panic: writing obviously made humanity not dumber but more capable. But it's worth reading the objection more carefully. Socrates wasn't saying writing would destroy intelligence; he was saying that the appearance of knowledge is easy to confuse with knowledge itself. That observation hasn't aged — it has become more precise.

The calculator story is similar. It didn't destroy mathematics, but it changed what makes sense to practice by hand; understanding orders of magnitude and noticing an absurd result is still the human's job. The analogy is useful but imperfect: a calculator performs a narrow, verifiable operation, while a language model takes on a much broader share of the work — wording, structure, argumentation, interpretation.

That is what's new. Earlier tools mostly offloaded memory and computation. Generative AI is the first to offer, at mass scale, offloading what we considered the core of thinking: building an argument, synthesis, choosing words, sometimes judgment itself. A search engine gave you links, and you still had to read, compare, and assemble an answer. A model gives you a finished answer, coherent and confident. The intermediate work — where learning used to happen — can vanish entirely.

Hence the new problem. The question used to be: "How do I find the information I need?" Now it is: "How do I keep the ability to think when thinking is optional?" It's a problem not of scarcity but of abundance. When a plausible answer costs almost nothing, other things become expensive: the ability to tell whether it's good, the ability to ask a question that doesn't have an answer yet, and the willingness to sit with uncertainty instead of grabbing the first option.

It also matters that the effect is unevenly distributed. A person with deep expertise uses the model as an accelerator: they see errors, discard what's irrelevant, ask precise follow-ups. A novice who hasn't yet built their own map of a field risks taking someone else's description of the terrain for the map. So AI can simultaneously strengthen the strong and slow the development of those still learning — if they use it to replace the learning process itself. More on how the value of different skills is shifting in how AI changes what we should learn.

What quality of thinking is

Before asking how AI affects thinking, we need to agree on what we mean by it. "Smart" in the everyday sense is often reduced to erudition or speed. But quality of thinking is neither the number of facts in your head nor IQ. It's a set of abilities that show up in how a person works with a question, with information, and with their own beliefs.

We can identify at least seven components. They don't exhaust the topic, but they provide a working map — and let us see which of them AI touches most.

Clarity comes first and is underrated. Many difficulties in thinking begin with a fuzzy question. If you can't say in a sentence or two what you're trying to understand, no model will help — it will simply answer some other question, and you'll take that for a solution. Paradoxically, working with AI can both train clarity (a good prompt demands precision) and replace it (the model "guesses" what you meant, so you no longer have to articulate it).

Depth is the ability to ask "why?" one more time. Shallow understanding knows what happens; deep understanding knows why and under what conditions it would be different. A model easily produces a first- or second-level explanation. But if you stop there, you haven't gained depth — only the feeling of being "up to speed."

Coherence is what turns scattered knowledge into thinking. Experts differ from novices less in how many facts they know than in how those facts are organized: into connected schemas that let them notice analogies and anomalies. Such schemas are built only through your own processing — retrieval, comparison, application. Someone else's schema, read once, rarely becomes yours.

Critical stance is the habit of not taking a claim on faith but asking: how do we know this? What would follow if it were false? What are the alternatives? There's much more in the complete guide to critical thinking. With AI it matters especially, because models can be wrong confidently and fluently — and fluency is one of the strongest cues we use to judge credibility intuitively.

Independence is the ability to reach a conclusion based on your own reasoning. It doesn't mean "not using sources." It means the sources are material, the conclusion is yours, and you can defend it.

Metacognition is knowledge about your own knowledge. Do I really understand this, or do I merely recognize it? Where are my gaps? What strategy do I need right now? This is arguably the central ability of a learner, and it deserves its own guide to metacognition.

Stamina is the willingness to stay with a problem that doesn't resolve right away. Many important ideas arrive not in the first five minutes but after a stretch of confusion. If every bit of confusion is immediately extinguished by an answer, that stretch disappears — and with it the chance of a discovery of your own.

So AI's influence shouldn't be judged on a single dimension: a model can improve the clarity of your writing while weakening your stamina for hard problems.

How AI changes thinking

The changes generative AI brings can be described through four mechanisms. None is unique to AI — each was studied long before chatbots, using search engines, navigation apps, autopilots, and learning hints. But AI makes them larger in scale and harder to notice.

Cognitive offloading

Cognitive offloading is the use of external actions and tools to reduce the load on your own memory and processing. Writing down a phone number, setting a reminder, sketching a diagram on paper, tilting your head to read rotated text — all of these are offloading. Psychologists Evan Risko and Sam Gilbert, in a 2016 review, described it as a fundamental strategy of human cognition, not a weakness. We have always been "distributed" thinkers who lean on our environment.

Offloading brings real benefits: it frees working memory for more complex operations. But it has a cost, well documented in research. When people know information will be stored externally, they often remember the information itself less well — and remember better where to find it. This was described as the "Google effect" after work by Betsy Sparrow and colleagues (2011). A caveat is in order: some of the original findings replicated weakly, so the exact size of the effect is debated. But the general principle — memory adapts to the availability of external storage — is consistent with the broader literature.

Navigation is a similar example. Studies comparing heavy GPS users with people who find their own way report an association between habitual reliance on navigation apps and weaker spatial memory. The data are mostly correlational, and the direction of causality isn't always clear. But they illustrate the idea well: a skill that isn't used doesn't get trained.

With AI, the scale of offloading is different. We can offload not only facts but the phrasing of a thought, the structure of an argument, the first draft, the data analysis. The question isn't whether to offload — offloading is normal. The question is what you offload. Offloading routine to free attention for the substance is sensible. Offloading the substance — the very reason you took on the task — means giving up the task itself.

A useful distinction: storage offloading (I don't keep in my head what I can look up) versus processing offloading (I don't do mental work I could have done). The first is harmless in most cases. The second is harmless only if you don't need the processing skill or it's already firmly established.

The illusion of understanding

In 2002, Leonid Rozenblit and Frank Keil described the "illusion of explanatory depth." People were asked to rate how well they understood how everyday things work — a zipper, a flush toilet, a lock. Their ratings were high. Then participants were asked to explain the mechanism in detail, step by step — and afterwards they lowered their self-ratings noticeably. What had felt like understanding turned out to be familiarity, not the ability to explain.

Later, Matthew Fisher, Mariel Goddu, and Frank Keil (2015) showed that searching the internet can inflate people's estimates of their own knowledge: after searching, people tend to rate their knowledge higher even on topics unrelated to what they had looked up. Access to external knowledge quietly blends into the feeling that the knowledge is your own.

Generative AI amplifies this mechanism. Model explanations are usually well structured, use analogies, and sound clear. Reading them, you experience fluency — a sense of ease in processing. And fluency, as cognitive psychology has long known, is something we tend to mistake for understanding and memory. The smoother the text, the stronger the confidence that "I've got it." But understanding is tested not by how easy something was to read, but by whether you can:

• explain the idea in your own words without a prompt;
• apply it to a new problem;
• name the boundaries where it stops working;
• respond to an objection.

That test is the essence of active recall. Learning research consistently shows that trying to retrieve knowledge from memory strengthens it more than rereading. Reading an AI explanation is, at best, rereading. If it isn't followed by an attempt to reproduce and apply, the illusion of understanding is almost guaranteed.

A special case of the illusion is summaries. A short recap of a book or article gives you the feeling of "knowing what it's about," when all you know is what the model considered important. More on this in the article on critical thinking and AI summaries.

Skill atrophy

The loudest fear is that AI will lead to skill atrophy: we'll forget how to write, calculate, reason. This calls for particular caution, because there's no reliable long-term data on generative AI yet. The technology has been widely available for only a few years, and cognitive changes at the level of skills and habits require long observation.

What do we have? First, research on automation in aviation, medicine, and industry. It has long described a phenomenon known as manual skill degradation: when automation takes over most of the work, operators practice manual control less often and may cope worse in a rare emergency. Alongside it is "automation bias," described for example in a review by Raja Parasuraman and Dietrich Manzey (2010): people tend to over-trust automated suggestions and miss their errors.

Second, early research on generative AI specifically. In a field experiment with high school students published in PNAS (2025), unrestricted access to GPT-4 improved scores during practice, but once access was removed, the students performed worse than the control group; a version with pedagogical guardrails largely eliminated this harm. It's analyzed in detail in the article on the performance–learning paradox. A survey of knowledge workers by researchers at Microsoft and Carnegie Mellon University (2025) found that higher confidence in AI was associated with less self-reported engagement in critical thinking, while higher self-confidence was associated with more. These are self-reports, not skill measurements. A small MIT Media Lab EEG study (2025, preprint) reported lower neural engagement among participants who wrote essays with ChatGPT — but the sample was small, the work hadn't been peer-reviewed at the time of publication, and interpreting brain activity as "weakened thinking" requires great caution.

The honest conclusion is this: there's no direct evidence that AI irreversibly reduces cognitive abilities. But there is a consistent picture from different sources: skills that aren't practiced don't develop, and in novices they may never form at all. For an experienced professional, frequent delegation may mean a slow loss of fluency. For a learner, the risk is more serious: a skill you never built can't "atrophy" — it simply never appears.

This is exactly where the calculator analogy needs nuance. The calculator replaced manual computation for people who already understood arithmetic — and curricula still require mastering it before the tool is used widely. If a tool arrives before understanding has formed, it doesn't free thinking for higher tasks — it removes its foundation.

Delegating judgment

The fourth mechanism is the subtlest and perhaps the most important. It's one thing to ask a model for a draft, to find a bug in code, or to remind you of a formula. It's another to ask it to decide what's right: which argument is stronger, which decision to make, what to believe, how to assess a person or an event.

Judgment is not computation. It involves values, context, responsibility, and knowledge of your own situation — none of which the model has. When you ask "tell me what's best," you get a plausible, averaged answer that sounds authoritative. Accepting it without doing your own work, you delegate not a task but a position — and with it, responsibility for that position.

There's a secondary effect too. Models are trained to be helpful and often to agree with the user. If you come with a ready-made position, the model will often confirm it and add supporting arguments. If you come without one, the model will offer its own — more precisely, whatever is most probable in its training data. Either way the risk is the same: you stop encountering the real resistance to your thinking that is what makes it sharper.

The same mechanism is at work as in belief formation generally: we tend to accept what sounds confident, gets repeated, and fits our existing views. More on this in the articles on the architecture of beliefs and how propaganda works. AI isn't a propagandist, but it can become a convenient source of ready-made judgments that are easy to adopt without checking.

The main danger: passive consumption of thinking

If we boil all the mechanisms down to one, the main danger isn't the technology itself but the mode in which it's used: passive consumption of thinking. That's when you receive the products of thought — conclusions, texts, explanations, decisions — without going through the process itself.

Passive consumption of information is long familiar: news feeds, TV series, endless scrolling. AI adds a new kind — consumption of reasoning. You don't just read other people's thoughts — you order them for your specific question, and they look like an answer addressed to you personally. This creates a strong feeling of participation, even when there may be none at all.

Consider two scenarios with the same task. A student needs to figure out why inflation accelerated in a particular country and write an analytical memo.

Scenario A. She opens a chat and types: "Explain the causes of inflation in country X and write a two-page memo." A minute later the memo is ready: structured, full of terminology, with conclusions. She reads it, seems to understand everything, tweaks a few phrases, and submits it. The result is good. A week later, in a discussion, she's asked why she considers the monetary factor the main cause rather than supply shocks. She tries to recall what was written but can't reconstruct the logic: the argument was never hers.

Scenario B. She first spends fifteen minutes writing by hand what she already knows and what her hypotheses are. She formulates three possible causes and, for each, what data would support it. Then she asks the model: "Here are my hypotheses. What alternatives am I missing? What data could distinguish between them?" After getting an answer, she checks two key claims against primary sources; one turns out to be inaccurate. She writes the memo herself and uses the model to critique the draft: "Find the weakest point in my argument." The result may be no better than in Scenario A and may take three times as long. But a week later she can defend every claim — and next time she'll do it faster.

The difference between the scenarios isn't whether AI was used. It was used actively in both. The difference is who did the key cognitive work: framing the question, generating hypotheses, choosing evidence, building the argument. In Scenario A that work was handed entirely to the model. In Scenario B it stayed with the person, and the model expanded it — pointing out missed alternatives and finding weak spots.

Passive consumption is insidious because it's pleasant. It saves time, removes the anxiety of the blank page, delivers a quick result and a sense of competence. Active use, by contrast, creates friction: you have to think before you ask, verify after you receive, rewrite in your own words. Learning researchers call this kind of friction "desirable difficulties" (Robert Bjork's term): conditions that slow you down now but improve long-term retention and transfer. The problem is that desirable difficulties subjectively feel like inefficiency — and people tend to avoid them.

It's worth stressing: passive consumption isn't always bad. If you need a quick reference on something you don't need to know deeply, getting a ready-made answer is fine. The problem is when this mode becomes the only one, especially in areas where you want or need to be competent.

Performance, competence, autonomy

To talk about AI's influence more precisely, it helps to distinguish three concepts that everyday language often blurs together.

Intellectual performance is how much intellectual work you do, and how well, with the tools available. The number of texts written, problems solved, documents analyzed, and their quality. Performance is measured by results and concerns the present moment.

Competence is what you can do on your own. Your ability to solve a problem, understand a text, build an argument when the tool isn't there or when it's wrong. Competence is measured by what remains once the help is taken away.

Intellectual autonomy is the ability to decide for yourself what to think about and how: to pose your own questions, form a position, evaluate sources — including AI itself — and take responsibility for your conclusions. Autonomy isn't "doing everything yourself"; it's keeping control over what you delegate and how.

AI affects these three levels differently.

It raises performance almost automatically. Numerous experiments with software developers, consultants, and customer support agents have shown gains in speed and output quality on typical tasks when using generative models — with important caveats about tasks beyond the model's capabilities, where help could actually make results worse. This is the main reason AI is spreading so fast: the benefit is visible immediately.

AI can raise competence or lower it, depending on the mode. If the model works as a tutor — asking questions, giving hints instead of answers, asking you to explain — it can accelerate learning. If it works as an executor, competence doesn't grow, and in novices it may not form at all. This is the "performance–learning paradox," analyzed in detail in a separate article.

Autonomy is the most vulnerable level. It isn't measured by tests and isn't visible in work output. It erodes without notice: you start more and more often with a prompt to the model rather than with your own thought; you less and less often form a position before learning "what the AI says"; you trust your own judgment less when it differs from the model's answer. Meanwhile performance may rise and competence may hold steady. But the ability to decide for yourself what and how to think gradually moves outside you.

Why is autonomy the most valuable? Because it determines how you use everything else. A person with high autonomy uses AI to amplify their own thinking: they know what they want, see where the model is wrong, and can reject its answer. A person with low autonomy becomes an extension of the tool: they can be very productive, but the direction of their thinking is set from outside. In a world where plausible answers are generated endlessly, the ability to decide for yourself what to believe and what matters becomes the primary intellectual capital.

A useful self-test: imagine that tomorrow you lose access to AI for a month. What changes? If you'd work more slowly but manage, you have high competence and autonomy, and AI is boosting your performance. If you couldn't even start tasks you now do regularly, it's worth asking how much of your skill exists only in combination with the tool.

How to develop your abilities

If thinking is a set of abilities, they can be trained. The good news: you don't need special programs. Most effective practices have long been known and well studied. The only new thing is that you now have to choose them deliberately, because the environment no longer forces them on you.

Reading

Deep reading is one of the strongest workouts for thinking available to us. A long, coherent text forces you to hold a line of reasoning across many pages, relate new material to what came before, notice contradictions, and build your own model of the author's thought. That's work a summary can't do for you: a summary gives conclusions but doesn't train the ability to follow an argument.

Reading researchers — Maryanne Wolf, for instance — have long discussed how a digital environment of rapid skimming may affect the habit of deep reading. There's a lot of uncertainty here too, but the mechanism itself is clear: the capacity for focused reading is a skill that needs maintaining. More on how reading changes the brain, and practical methods in the guide to reading books.

A practical rule: if a text matters to you, read it yourself first. An AI summary or recap is useful afterwards — to compare with your understanding and see what you missed or understood differently. If you start with the summary, you read through its frame and notice mainly what it highlighted.

Writing

Writing isn't just a way to record a thought; it's a way to create one. Many writers and scientists say they don't know what they think until they try to write it. When you search for a word, restructure a paragraph, notice that an argument doesn't hold — you're thinking. A model that writes for you takes away precisely that part.

That doesn't mean AI can't be used in writing. You can use it as an editor, a critic, an opponent: "Where is my argument weakest?", "What objection haven't I considered?", "What's unclear to a reader who lacks the context?" But the first draft — especially in areas where your own position matters — is worth writing yourself. Let it be clumsy: a clumsy draft of your own thought is worth more than a polished draft of someone else's.

A useful habit is short writing with no help at all: a few paragraphs a day about what you've read, what you're thinking about, what you disagree with. It's not literature; it's clarity training.

Problems without hints

Solving problems without immediate help is key to stamina and competence. When you're stuck, your brain searches for a solution, tries approaches, runs into the limits of your understanding. Even if you don't solve the problem yourself, the attempt makes the subsequent explanation far more useful: learning research describes this as the generation effect and "productive failure" — attempting before instruction often improves understanding, though the size of the effect depends on conditions.

A rule for working with AI: give yourself a real attempt before asking. Not thirty seconds, but long enough to understand exactly where you're stuck. And when you ask, ask for a hint, not a solution: "Don't give me the answer — point me in a direction to think." Many models can be asked to work this way — and, judging by early research, this fundamentally changes the effect on learning. Detailed scenarios are in the guide to studying with AI.

Curiosity

Curiosity is the engine of thinking, and it's easy to underestimate. It's the ability to notice that you don't understand something and to want to understand it. In a world of instant answers, curiosity can either flourish (any question can be explored) or wither (any question is closed so fast that interest has no time to develop).

To keep curiosity alive, it helps sometimes not to rush to an answer. Write the question down, think about it for a few days, come back to it. Look for answers in several sources, not one. Ask not only "what?" but "why this way and not another?" and "what would happen if…?" Use AI as a conversation partner that helps deepen the question rather than a reference book that closes it: "What are the open questions here?", "Where do experts disagree?", "What is still unknown?"

Rules of intellectual hygiene with AI

Hygiene is the right word. This isn't about prohibition but about habits that reduce risk without taking away the benefits. Washing your hands doesn't mean giving up food, and rules for working with AI don't mean giving up AI.

Rule 1. I think first, then check. Before asking the model a question, spend at least a few minutes on your own version: What do I already know? What are my hypotheses? What do I expect the answer to be? Write it down. Then the model's answer becomes not a substitute for your thought but a point of comparison. You'll see where you matched, where you were wrong, what you missed — and that in itself is powerful learning feedback.

Rule 2. I verify key claims. Models can be wrong — inventing sources, mixing up dates, simplifying to the point of distortion. You don't need to check everything, but whatever you rely on in a decision or argument is worth checking against an independent source. A good question to ask yourself: if this claim is false, what breaks? If a lot breaks, check it.

Rule 3. I explain in my own words. Once you've received an explanation, close it and try to reproduce it in your own words — out loud, in writing, or as a diagram. If you can't, you didn't understand; you only read. This is the simplest and most reliable antidote to the illusion of understanding.

Rule 4. I keep judgment for myself. A model can help gather arguments, show alternatives, find weak spots. But formulate the final conclusion — what's true, what matters, what to do — yourself, and be ready to defend it without saying "the AI said so." If you can't justify a position with your own arguments, you don't have a position — you have a borrowed answer.

Rule 5. I keep AI-free zones. Choose areas or situations where you deliberately don't use an assistant: a personal journal, reading important books, solving problems in a field you're learning, first drafts. This isn't asceticism; it's a training gym where skills get maintained that would otherwise get no load.

It's worth occasionally asking yourself: in which tasks over the past week did I start with the model before I'd had a chance to think? This isn't a reason for self-reproach but a way to notice automatic behavior.

The thinking environment beyond AI

It would be a mistake to look at AI in isolation. It's embedded in a broader information environment that poses serious challenges to thinking even without it. If we're talking about preserving independence, that environment has to be taken into account too.

Information overload. The volume of available information has long exceeded what a person can process. Overload leads not to better understanding but to skimming, quick heuristics, and decision fatigue. AI is double-edged here: it can help filter and structure, but it can also add even more text — summaries, recaps, answers — that you also have to process.

Fragmented attention. Notifications and task-switching break attention into pieces, and every switch carries a cognitive cost, even when it feels like we're coping.

Information bubbles and algorithmic personalization. News feeds and recommendations adapt to our preferences, and we see more of what we already agree with. The strength of the "filter bubble" effect is debated in the research literature and is probably smaller than popular writing suggests. But the tendency to choose confirming information is itself a well-studied fact about human thinking. An AI assistant that adapts to your requests and style can become yet another — very personal — mirror.

Manipulative messaging. Propaganda, advertising, and political rhetoric use the same mechanisms as the illusion of understanding: repetition, a confident tone, emotional engagement. With generative models, producing persuasive text has become dramatically cheaper. That raises the value of being able to recognize the techniques — they're described in detail in the article on how propaganda works.

What to do about it? The same principles as with AI, only broader. Create periods of focused work without distractions. Deliberately seek out sources that disagree with you. Give yourself permission not to have an opinion on a question you haven't studied. Notice when a text is trying to provoke an emotion instead of making an argument. And remember that AI is not the only element — just the newest — of an environment in which thinking needs protection.

Three modes: consumer, operator, partner

A person's relationship with AI can be described through three modes. These aren't personality types — the same person can switch between them depending on the task. But it's useful to notice which one dominates for you.

Consumer. Uses AI as a source of finished products: answers, texts, solutions. The question is phrased quickly, the answer is accepted with almost no checking, and the result is used as is. Performance can be high, especially on routine tasks. But competence barely grows, and autonomy gradually weakens: the consumer starts less and less often from their own thought. This mode is appropriate for tasks where you really don't need to dig in — quick lookups, drafts of technical emails, formatting. It becomes dangerous when it's the only mode.

Operator. Manages AI as a tool: knows how to write precise prompts, break a task into steps, check the result, fix errors. The operator achieves high performance and keeps control over quality. But they may never bring the model into their own development: the tool does the work, the operator coordinates it. Their domain competence may stand still, even as their competence in managing the tool grows. That's a valuable skill, but not a substitute for thinking.

Partner. Uses AI as a conversation partner for developing their own thinking. The partner arrives with their own position and asks for it to be challenged. Asks not for answers but for questions. Uses the model to see alternatives, test arguments, find gaps in understanding — and then draws conclusions themselves. The interaction is slower and takes more effort, but both competence and autonomy grow as a result. Here AI plays a role once filled by a strong opponent, a mentor, or a demanding editor — a role that used to be available to very few.

Moving from consumer to partner is a question of mindset, not technique: who is doing the thinking here?

All three modes have their place: there's no point being a partner while reformatting a spreadsheet. But in your profession, your studies, and your beliefs, partnership separates the person who grows along with the tool from the person who depends on it.

The 30-day protocol

Awareness is useful, but habits change through practice. Below is a four-week protocol that helps rebuild your relationship with AI without giving it up radically. It isn't based on a specific study as a ready-made program — it's an assembly of well-known principles of learning and habit formation, worth adapting to yourself.

Week 1. Observation. Change nothing; just notice. Each time you use AI, jot down briefly: what was the task, did I try it myself first, which mode (consumer, operator, partner) was I in? At the end of the week, look over your notes. Where did you reach for the model automatically? Which tasks did you used to do yourself? The goal of the week is to make the invisible visible.

Week 2. The "me first" rule. For all important tasks, introduce a mandatory pause before prompting: a few minutes of your own thinking, written down in at least two or three sentences. What do I think? What answer do I expect? After the model answers, compare. Note how often your first version was close and where the model showed you something new. At the same time, choose one AI-free zone — reading a book or keeping a journal, for example.

Week 3. Partner mode. Try changing your prompts in several tasks. Instead of "write," say "here's my draft, find the weak spots." Instead of "explain," say "ask me questions to check whether I understood." Instead of "what's better," say "here's my choice and my reasons — give me the strongest objection." Add the explain-in-your-own-words rule: after every important explanation, close it and reproduce it.

Week 4. Autonomy check. Pick a task you usually do with AI and do it entirely without it. Note what was hard, where you got stuck, what took longer. This isn't a punishment but a diagnostic: you'll find out which skills are being maintained and which exist only in combination with the tool. Then decide which of those skills matter to you and make them permanent AI-free zones.

After 30 days, it's worth writing down your own rules — a short list that answers these questions: Where do I use AI freely? Where only after my own attempt? Where not at all? How do I verify important answers? Revisit the list every few months: both the tools and your tasks change.

The goal of the protocol isn't to minimize AI use but to make it deliberate.

Going deeper: what else is worth preserving

Beyond skills and rules, there are deeper qualities of thinking that become especially valuable in the age of AI, even though they're hard to measure.

Intellectual stamina

The ability to stay with a hard problem — a question without a ready answer, a text that doesn't open up right away, a problem that takes days of thought. Many important discoveries and insights come not in a moment of focused search but after a long period of work, and even of rest from it. If every difficulty is immediately extinguished by an answer, this ability doesn't get trained.

Stamina develops gradually: more time before asking, texts slightly harder than usual, returning to unsolved problems a few days later.

The right not to know

In a culture of instant answers, "I don't know" sounds almost like defeat. But the ability to admit ignorance is the foundation of honest thinking. Only someone who knows what they don't know can learn. Metacognition begins with that admission.

AI creates a temptation never to remain in not-knowing: any question gets a plausible answer in seconds. But a plausible answer obtained without understanding isn't knowledge — it's an imitation of it. Sometimes it's more honest and more useful to say "I don't know, I need to look into it" — and actually look into it — than to voice someone else's conclusion as your own. There's separate value in allowing yourself not to have an opinion on a question you haven't studied.

Knowing how to ask

When answers get cheaper, questions get more expensive. A good question determines what you'll see at all: it sets the frame, chooses what counts as important, separates the known from the unknown. A model answers the question you asked — and if it's shallow, the answer will be too.

Knowing how to ask involves several levels: clarifying questions ("what exactly is meant?"), questions about evidence ("how do we know this?"), questions about alternatives ("how else could this be explained?"), questions about consequences ("what follows if this is true?"), and questions about the question itself ("am I framing the problem correctly?"). These skills develop through practice — including deliberate work with AI, if you use it to deepen the question rather than to close it.

Independent judgment

Independence doesn't mean setting yourself against everyone. It means your position rests on your own reasons, not on the authority of a source — whether an expert, the majority, or AI. An independent thinker can agree with the model, but because they checked the arguments, not because the model said so.

AI answers seem "objective," even though they reflect training data, developers' choices, and the wording of your question. They're one point of view, not a verdict.

Preserving effort

Finally, the simplest and the hardest thing — keeping the willingness to make mental effort when it can be avoided. The brain is economical: it tends to choose the path of least resistance. That's not a flaw but an evolutionarily sensible strategy. But in an environment with almost no resistance, it can lead to effort disappearing altogether.

Preserving effort is a conscious choice: where I save my energy, and where I spend it, because it's precisely in that effort that ability grows. Thinking, like muscle, is maintained by load.

FAQ

Research and related foundations

This article draws on several lines of research, and it's important to understand their strengths and limitations.

Cognitive offloading. The review by Risko and Gilbert (2016, Trends in Cognitive Sciences) describes offloading as a basic cognitive strategy and discusses its cost for memory. The work by Sparrow and colleagues (2011, Science) on the "Google effect" became a classic, but some of its findings replicated weakly in follow-up studies — worth keeping in mind. Research on GPS and spatial memory is mostly correlational.

The illusion of understanding. The illusion of explanatory depth (Rozenblit and Keil, 2002) replicates well. The study by Fisher, Goddu, and Keil (2015) shows how internet search inflates estimates of one's own knowledge. The fluency effect — confusing ease of processing with understanding — is widely studied in cognitive psychology.

Learning and desirable difficulties. The testing effect (active recall) and Robert and Elizabeth Bjork's concept of desirable difficulties are among the most robust findings in the science of learning. More in the guides to active recall and metacognition.

Automation. Research on automation bias and manual skill degradation (for example, the review by Parasuraman and Manzey, 2010) has accumulated in aviation, medicine, and industry. Transferring it to generative AI is plausible but not automatic.

Generative AI. Here the data are the newest and the least settled. The PNAS field experiment (2025) on GPT-4 and high school learning is analyzed in the article on the performance–learning paradox. The Microsoft and Carnegie Mellon survey of knowledge workers (CHI 2025) is based on self-reports. The MIT Media Lab EEG study (2025) is a preprint with a small sample. All of these are useful as early signals, but not as a final verdict on AI's long-term effect on thinking.

Reading and environment. On reading: how reading changes the brain and how to read books. On belief formation and manipulation: the architecture of beliefs and how propaganda works. On critical thinking: the complete guide and working with AI summaries. On how learning priorities are shifting: how AI changes what we should learn.

Conclusion

We live at a moment when access to knowledge — and even to reasoning — has become almost free. That's an enormous opportunity: a person with good questions and independent thinking gets a conversation partner, editor, and opponent they could once only dream of. But the same technology offers a convenient path around thinking — a path on which performance rises while the ability to think for yourself quietly moves outside you.

The choice between these paths isn't made once. It's made in every small situation: start with my own thought or with a prompt? Verify or accept? Explain it to myself or just read it? Form a position or take a ready-made one? Individually these decisions seem insignificant. Together they determine what kind of thinker you'll be in a few years.

The key intellectual competence of our time isn't knowing how to use AI, important as that is. It's remaining the author of your own thinking when a machine can think alongside you. AI will give you performance anyway. Competence you'll have to build deliberately. Autonomy you'll have to protect.

If you want to practice critical reading of texts, including AI-generated ones, try ThinkLens: it helps you spot assumptions, weak arguments, and manipulative techniques — and reach your own conclusions.