A comfortable illusion. There is a lie we tell ourselves collectively, every day, each time we open a chatbot and receive a response that sounds like wisdom. The lie is not spoken aloud. It doesn't appear in terms and conditions. It cannot be proved in any court. It is subtler than that: it is the feeling that we have understood something. That we have thought. That we have arrived, through our own intellectual effort combined with the power of the machine, at a valid conclusion. 'As stated above.' It is an introduction that refuses to be merely an introduction.
Listen to the audio dialogue summary
It is not true.
It is not true, and it is not harmless. 
And precisely because it is so comfortable, so fluid, so nearly inevitable — it deserves to be examined with the very instruments we are, paradoxically, allowing ourselves to lose in its presence: critical thinking, patience, the capacity to bear uncertainty without pressing a button.
Sit with that for a moment. Don't continue. Don't scroll.
There is a moment in any genuine act of reading when the resistance of the text makes you stop. When the sentence doesn't yield immediately and you have to work. That moment of friction — that small cognitive discomfort — is precisely where understanding happens. Not before. Not after. There, in that uncomfortable interstitial space.
Now imagine that moment disappearing. That there is always a voice to complete the sentence, summarise the argument, serve you the conclusion on a digital silver tray. Imagine that disappearing systematically, across millions of minds, day after day, over the course of a few years.
That is not a working hypothesis. It is a description of the present.
The machine that thinks in your place — and what that does to the brain
There is a concept in the specialist literature which, translated as directly as possible, sounds almost banal: cognitive offloading. The person delegates thinking to the machine. Receives the conclusion and accepts it.
The simplicity of that description conceals, however, a deeply troubling depth. The human brain is not a static processor that functions at full capacity regardless of demand. It is a living organ, plastic, continuously shaped by the kinds of tasks it performs. Neurones that are not activated thin out. Networks that are not used reorganise themselves. Cognitive muscles — the capacity for analysis, for synthesis, for counterfactual thinking — atrophy through disuse exactly as a muscle atrophies after six weeks of immobilisation.
Researchers at Aalto University found something apparently banal: the majority of users address the chatbot with no more than a single question per problem, with no further verification, blindly accepting the system's response. A single exchange. A single layer of interrogation. No return visit, no counter-argument, no questioning.
That is not a statistic about laziness. It is a statistic about how we have, without intending to, reconfigured our relationship with knowledge.
My father was a mechanic. Not the sort with diplomas and European certifications — a backstreet mechanic, who had learnt from an uncle, who had got it wrong hundreds of times before understanding why an engine knocks abnormally at low temperatures. When I came to him with a problem with the car, he didn't give me an immediate answer. He made me listen. Rev the engine cold. Rev it hot. Listen again. Different? Identical? Louder or quieter?
He wasn't educating me about engines. He was educating me about how to pose a question. About how a good question is worth more than a quick answer.
AI has removed the good question from the equation. It has given us the quick answer. And we accepted the exchange without negotiating, without reading the small print, without understanding what we were losing.
Shanmugasundaram and Tamilarasu reported that over-reliance on artificial intelligence for answers, assignments, and information can reduce an individual's capacity for critical thinking and independent thought. That sentence, buried in an academic study read by a few hundred people, describes something happening in millions of minds simultaneously, right now, as you read this.
It is not a prophecy. It is a diagnosis.
And the diagnosis carries an internal irony worth looking squarely in the face: the better the tool, the more severe the diagnosis. A poor hammer does no harm to the hand that holds it — it breaks by itself. A perfect hammer, in the inexperienced hand, can smash anything.
The reverse Dunning–Kruger effect — or how the machine makes you more confident precisely when you should be more cautious
Classical psychology describes a mechanism with an uncomfortable name: the Dunning–Kruger effect. Those who know little overestimate what they know. Those who know a great deal are more cautious, more aware of the limits of their knowledge. It is an almost consoling dynamic: the fool is too sure of himself, the wise person knows that he doesn't know.
Generative artificial intelligence has upended this dynamic. Not gradually. Not partially. Completely.
A study published in the journal Computers in Human Behavior, suggestively titled 'AI Makes You Smarter but None the Wiser', discovered something profoundly disorienting: when users interact with tools such as ChatGPT or similar systems, all of them — regardless of their actual level of competence — overestimate their performance. So far, perhaps unsurprising. What is surprising comes next: users with greater familiarity with these tools, with better digital literacy in the domain of artificial intelligence, display an even higher degree of overestimation.
Sit with that. Don't continue.
The better you are at using AI, the more convinced you are that it was you who did the thinking. Not the machine. You. The classic fool, the one from Dunning–Kruger theory, at least has an internal brake — somewhere, vaguely, he feels that he doesn't know. The fool assisted by AI no longer has even that brake. The machine has confiscated it, helpfully, efficiently, without asking his consent.
It is as though you had driven an automatic car for twenty years and were then convinced you knew how to drive a lorry with a manual gearbox on a mountain road. It's not that you can't. It's that you don't even know that you can't. And that is precisely the scale of the problem.
Performance increases. Understanding disappears. And nobody notices the discrepancy.
There is a rhetorical trap in any debate about artificial intelligence and cognition. Those who defend the tools point to results. 'Texts are better. Answers in tests are more accurate. Productivity has increased.'
And they are right. Partially.
Studies from 2025 show that groups using artificial intelligence tools report significantly higher self-evaluation scores after completing tasks, compared with groups that do not use them. Externally measured performance — the quality of the finished product — does indeed rise. But those same people, tested afterwards without tools, achieve genuinely lower scores than before they used the tool.
To put it plainly: the tool produced a better result. But the person behind the tool emerged from the interaction less capable than when they entered it.
An experiment on students enrolled in a course in creative thinking and problem-solving found that those assisted by AI achieved higher scores in fluency and flexibility. The same experiment found that the tools generated what researchers call cognitive passivity — a dependency that undermines the capacity for subsequent independent thought.
The brain operates on a 'use it or lose it' principle. That is not a metaphor. It is neurobiology. Synapses that are not activated are eliminated. Connections that are not called upon atrophy. A thought that is never followed through to its conclusion, without assistance, becomes a muscle that can no longer lift anything on its own.
My mechanic father would have understood that without having read a single academic study. 'A good tool in an unskilled hand,' he used to say, 'is more dangerous than no tool at all.' Not because it damages the bodywork. But because the man leaves the workshop convinced that he knows how to repair engines.
'Lobster' assistants and the illusion of autonomy — on OpenClaw and the infrastructure of intelligent dependency
There is a category of tools that raises these questions to a higher power. We are no longer talking about a chatbot you open on your computer when you need an answer. We are talking about permanent assistants, continuously active, connected to every channel through which you communicate — WhatsApp, Telegram, Slack, Discord, Signal — present in every corner of your digital life.
OpenClaw is one of them.
OpenClaw is an open-source framework for running autonomous and agentic artificial intelligence systems that interact with programming interfaces, tools, and messaging platforms in real time. digitalocean Through the DigitalOcean platform, it can be installed on your own infrastructure, running continuously, connected to the messaging platforms you already use — WhatsApp, Telegram, Slack, Discord, Signal, and others. digitalocean
The name is revealing. Claw — a talon, a pincer. The image of a lobster. That crustacean with its hard exoskeleton and powerful claws, which moves backwards when threatened and seizes everything in its path. It is probably an involuntary metaphor — but involuntary metaphors are always the most honest.
Let us unpack it.
A lobster has an exoskeleton, not an endoskeleton. Its structure of resistance is on the outside, not the inside. It does not grow by strengthening its internal structure — it grows by moulting, by shedding the old shell and exposing itself temporarily, vulnerable, to its environment, before the new shell hardens. If a lobster could entirely avoid the moulting process — if it could remain permanently in its old, comfortable, safe shell — it would be doomed. It would die a prisoner of its own protection.
Permanent AI tools work in precisely the opposite way to how cognitive growth functions. Cognitive growth requires exposure to resistance, to uncertainty, to the discomfort of not knowing and having to find out for oneself. An assistant that responds permanently, on every channel, at any hour, integrating conversational context from Telegram in the morning and Slack in the afternoon — that is not a tool. It is a permanent shell.
Deploying OpenClaw on a DigitalOcean Droplet configures it as a self-hosted AI assistant that runs at all times, acting as a control plane for managing sessions, channels, tools, and events. digitalocean
Always on. Remember that phrase. It does not describe a tool. It describes a presence. A presence that does not sleep, does not tire, does not have a bad day, does not say 'I don't know — think it through yourself.' A presence that has, amongst other things, support for tool execution: web browsing, API calls, file system operations, voice, command-line actions. digitalocean
In other words: a presence that can do almost everything you do, more quickly, without hesitation, without fatigue.
And you, in the midst of this permanent presence — what do you do? Do you exercise your cognitive muscle, or do you let yourself be carried along by the current of ready-formed answers?
The question is not rhetorical. It is diagnostic.
Who truly benefits, and why the answer matters
There is an empirical answer to the question of who genuinely gains from this relationship with artificial intelligence tools, whether we are talking about a simple chatbot or a permanent agentic assistant of the OpenClaw type. The answer is uncomfortably clear: those who already know.
A comparative study found that AI-assisted participants achieved better results in logical reasoning, structuring, and problem definition, but showed significant weaknesses in generating new ideas, multidisciplinary integration, and critical rejection of unfounded conclusions. The practical translation: the tool is good at executing structures the user already knows. It cannot create them from nothing.
The expert uses AI as an amplifier — checks, corrects, completes. They have an internal framework against which they measure the machine's response. They know when the machine is wrong. They know when a response is plausible but false, when it is correct but incomplete, when it is complete but irrelevant to the actual problem.
The person without foundation uses AI as an oracle. They have no internal framework against which to measure. They accept. They consume. They believe.
And here a dimension emerges that academic research touches upon timidly, but which daily practice demonstrates brutally: assistants of the OpenClaw type are not neutral with respect to this dynamic. They deepen it. A tool that is permanently active, that knows your conversational history, that can browse the web on your behalf and call APIs on your behalf — such a tool leaves no space for cognitive resistance. It is too good. Too present. Too available.
And, paradoxically, that is precisely what makes it dangerous for the unprepared user.
A poor hammer forces the hand to understand the material. A perfect hammer lets the hand forget that the material exists.
The digital exoskeleton and what happens when you remove it
There is a thought experiment worth conducting honestly. Imagine you use an OpenClaw-type assistant for six months. You connect it to Telegram, to Slack, to e-mail. You configure it to manage communication routines, to summarise long conversations, to respond to frequently asked questions, to search the web when you need it. It works flawlessly. You are more productive. You have more free time. You are less stressed.
Now imagine that at the end of those six months, the server breaks down. Or DigitalOcean has an outage. Or you simply decide to stop paying the subscription.
What remains?
That is not a question about data. It is not a question about saved files or archived conversations. It is a question about capability. What can you do on your own, after six months in which you did not practise doing things on your own? What thoughts can you follow through to their conclusion without an assistant to complete, suggest, and reformulate?
A study involving 666 participants from diverse age groups and educational backgrounds found a significant negative correlation between frequent use of AI tools and critical thinking skills, mediated precisely by an increase in cognitive offloading. Younger participants — those who are growing up with these tools as a normality rather than a novelty — demonstrated greater dependency and lower scores in critical thinking.
To translate: this is not about a generation using better tools. It is about a generation growing up without developing the capacity to think without tools.
The lobster has claws. But if those claws are constantly supplemented by a mechanical pair, stronger and faster, the creature's own claws no longer get exercised. They no longer strengthen. They remain small, functional in theory, useless in practice.
And when the crustacean wants to moult — to grow, to outgrow the old shell — it discovers it has nothing with which to grasp the new world.
The infrastructure of dependency — or why the construction of the tool matters
It is worth looking more carefully at the architecture of a tool such as OpenClaw — not to condemn it, but to understand what it is, precisely, in its structure that produces the effects we are discussing.
The system functions as a control plane for managing sessions, channels, tools, and events, integrating the messaging platforms you already use. digitalocean It is designed for continuity, for permanent presence, for the elimination of friction from any interaction with information or communication.
Friction eliminated. Remember the phrase. It sounds like progress — and in many contexts it is progress. Unnecessary friction, the kind that produces no cognitive value, is worth eliminating. If I can automate sending a confirmation e-mail for a meeting, I have lost nothing cognitively by relinquishing that effort.
But there is another category of friction. Cognitive friction. The kind that arises when you must formulate a question on your own, evaluate an answer on your own, decide on your own whether the information is credible. This friction is not a defect in the thinking process. It is the thinking process. It is precisely where judgement is formed.
OpenClaw supports multi-agent routing with isolated sessions and configurations, tool execution — web browsing, API calls, file system operations, voice, command-line actions — and extensibility through plug-ins and custom capabilities. digitalocean
Each of these functionalities eliminates one moment of resistance. Web browsing on your behalf eliminates the moment in which you evaluate sources. API calls on your behalf eliminate the moment in which you understand what data you are looking for and why. Command-line actions on your behalf eliminate the moment in which you learn how the system you are using actually works.
The final result, after enough months of intensive use: a more productive, more efficient, and less competent user. The three terms are not mutually exclusive. That is precisely the paradox.
And the paradox has an ethical dimension that the industry prefers not to discuss. Tools of the OpenClaw type are designed for power and flexibility — the technical documentation is impeccable, security is taken seriously, the architecture is solid. Deployment through DigitalOcean includes portal authentication tokens, firewall-level rate limiting, non-root execution, and container isolation. digitalocean All of these are features of a serious tool.
But not a single line of documentation says: caution — inexperienced users risk atrophying their cognitive capacities through permanent use.
Not because the authors didn't know. But because that is not their remit. Their remit is to build the tool. Your remit is to decide how you use it.
And precisely that decision — conscious, calibrated, informed — is the first and most important cognitive function that the tool itself cannot perform in your place.
The amplifier and the transformer — a distinction that matters
There is a fundamental difference between amplifying and transforming. An audio amplifier increases the volume of an existing signal. A transformer changes the nature of the signal. Badly played music, amplified, remains badly played music — only louder. But a transformed signal can become something entirely different.
Artificial intelligence — including permanent agentic assistants such as OpenClaw — is an amplifier, not a transformer. It amplifies what already exists in the user's mind. If structured thinking, critical spirit, and capacity for synthesis are present there, the tool makes all of that better, faster, more efficient. If emptiness, confusion, and an inability to formulate a good question are present there — the tool amplifies the emptiness, the confusion, and the incapacity too.
That is the central thesis. And it is demonstrable empirically, not merely philosophically.
A skilled carpenter with an electric saw does better work than the same carpenter with a hand saw. The difference lies in the quality of the carpenter, not in the tool. A person with no idea of carpentry, with the same electric saw, produces warped planks — and is convinced he has made fine cabinetry.
Now take the same person. Give him a permanent agentic assistant that can browse the web for inspiration, call APIs for technical specifications, generate plans and materials lists, schedule deliveries, and manage orders. The person with no idea about carpentry now produces warped planks with extraordinary efficiency. And he is even more convinced that he has made fine cabinetry.
The better tool has produced more illusion, not more competence.
What it means to use a poorly understood tool well
I am not arguing for the abandonment of tools. I am arguing for something more difficult: for using them with clear-headedness.
There is a middle way between entirely denying the value of permanent AI assistants and surrendering to them blindly. That way has no technical name. It does not appear in DigitalOcean documentation or in OpenClaw configuration tutorials. But it can be described.
The first principle: use the tool for tasks that do not require judgement, not for those that do. Automating a confirmation e-mail atrophies no cognitive capacity. Allowing an assistant to formulate your central arguments in your place — does.
The second principle: practise regularly without the tool. Not occasionally. Not when the tool isn't working. Regularly, deliberately, as training. Exactly as an athlete who uses assistive equipment must also train without it, so as not to lose their basic capacities.
The third principle: verify, question, and counter-argue every response received from the tool — not out of distrust of the technology, but out of respect for your own cognitive apparatus. A muscle that is not called upon does not maintain itself. A critical faculty that is never set in motion does not remain on standby. It disappears.
The fourth principle: know the tool's limitations better than the tool knows yours. A permanent agentic assistant, with access to messaging, the web, APIs, and the file system, can do an enormous amount. But it cannot know when you are asking the wrong question. It cannot know when the premise you start from is false. It cannot know that the problem you have asked it to solve is not your real problem.
These are not functions of the tool. They are functions of judgement. And judgement does not develop through delegation.
Conclusion — or returning to the first sentence
You read the title. Perhaps you accepted it immediately. Perhaps it struck you as exaggerated. Perhaps you found it provocative and kept reading in order to find the error.
Good. That means your critical faculty is functioning. Use it.
AI does not make the incompetent person more intelligent. Not because the tool is poor. But because the tool, however good, cannot substitute for the process through which intelligence is formed — the process of resistance, of error, of revisiting, of doubt, of the slow and painful synthesis of information into judgement.
Permanent assistants, the digital lobsters that seize everything thrown at them and never sleep — OpenClaw or any other in this family of tools — are remarkable products of engineering. They are well built, well secured, well documented. And they are, in the hands of the unprepared user, exactly what an electric saw is in the hands of someone who doesn't know that wood has grain and that grain matters.
Those who know what they are doing with them will use them to do more, better, faster.
Those who don't know what they are doing will produce, with great efficiency, stupidity on an industrial scale.
And will be more convinced than ever that they have produced a masterpiece.
My mechanic father did not teach me to repair engines. He taught me to listen to engines. To distinguish one sound from another. To know that I didn't know, but to know what I didn't know — which is something entirely different from pure ignorance.
Now I have in my pocket a telephone with an agentic assistant that can diagnose an engine faster than he ever could. And I wonder how many sons will grow up having never listened to an engine started by hand, without help, without knowing in advance what they would hear.
And I wonder one more thing. A single thing, which I leave without an answer, because an answer received ready-made is worth nothing.
What will they know how to listen for, when there is no longer any machine to tell them?
Select bibliography
-
Fernandes et al. (2025) — 'AI Makes You Smarter But None the Wiser', Computers in Human Behavior
-
Yavich (2025) — Will AI undermine independent thinking?, Education Sciences / MDPI
-
OpenClaw — technical documentation, DigitalOcean Marketplace
-
Plato, Phaedrus — the critique of writing as a technology of memory
10 Questions and Answers — 'Does AI make you incompetent?'
Is it true that AI makes you incompetent, or is that a deliberate exaggeration?
It is a partially justified exaggeration. The empirical truth is more nuanced: AI does not produce incompetence from nothing — it amplifies pre-existing incompetence and, at the same time, can slow the development of competence in those who are just starting out. The exaggeration serves a precise rhetorical purpose: it shocks just enough to break through indifference, to force a pause, a moment of reflection. If you say 'AI can, under certain conditions, gradually reduce certain cognitive capacities in inexperienced users' — nobody stops scrolling. If you say 'AI makes you incompetent' — someone looks up.
Exaggeration is a tool. Like any tool, it is ethical if you know you are using it and acknowledge it when the moment calls for it.
What is the actual psychological mechanism by which AI affects competence?
The mechanism is called cognitive offloading — and it is not new. We have known it from computers, GPS, search engines. What is new is the scale and the depth. The human brain operates on a principle of effort economy: if a task can be externalised, it externalises it. The problem arises when the tasks being externalised are not the trivial ones but the formative ones — the formulation of a good question, the evaluation of a source, the construction of an argument from scratch.
Psychologically, it is similar to what happens in emotionally dependent relationships: the person who relies entirely on the other for internal validation no longer develops their own self-regulation resources. Not because they cannot — but because they are no longer required to.
Why can exaggeration on this subject be editorially productive?
Because the subject suffers from the opposite effect: systematic minimisation. The technology industry has every economic reason to present AI exclusively as an amplifier of human competence. The academic studies showing negative effects are real, but they are buried in jargon and published in journals that nobody outside the field reads.
Rhetorical exaggeration fulfils a function of balancing public discourse. It is not disinformation — it is counterweight. The condition is transparency: the reader must sense that the author knows where the line is.
What is the most solid argument in favour of the thesis, without exaggeration?
The most solid argument is the one from the study of 666 participants, which found a significant negative correlation between frequent AI use and critical thinking scores. Not a vague correlation. A statistically significant one, mediated by increased cognitive offloading, with more pronounced effects in younger participants.
That does not prove that 'AI makes you incompetent.' It proves that frequent and uncritical use of AI is associated with a decline in certain specific cognitive capacities. The distinction matters — but for an editorial headline, the first version works better.
Where does the logic break down if you exaggerate too much?
When you ignore the conditionals. 'AI makes you incompetent' is false as a universal statement. The surgeon who uses AI for medical imaging analysis does not become incompetent — they become more precise. The programmer who uses a code assistant to detect errors does not lose their architectural judgement — if they already had it.
The exaggeration breaks down when it no longer distinguishes between amplification and substitution, between the user with a foundation and the one without. If you ignore that distinction, you lose credibility with the informed reader — precisely the one you want to win over.
Is there a deeper psychological dimension, beyond cognitive offloading?
Yes. It is the dimension of identity. Competence is not merely a cognitive function — it is part of our identity. The feeling of knowing how to do something, of being able to solve a problem, of having worked through a difficulty and emerged with a solution — all of these contribute to what psychologists call self-efficacy, perceived personal effectiveness.
When AI systematically takes over difficult tasks, you do not lose merely the technical skill. You also lose the experience of having struggled with something and won. And that experience is the foundation on which authentic self-confidence is built — not the kind inflated by artificially amplified performance.
How do you respond to the reader who says: 'But I use AI and I'm more productive than ever'?
With empathy and a question returned: productive in what sense? If you produce more in the same time — possibly true. If you are more capable of thinking independently, of evaluating critically, of constructing arguments without assistance — that is a different measure. And one that people rarely apply to themselves, precisely because the tool is always to hand.
That is exactly the paradox studied: participants who used AI reported higher performance — and achieved lower scores when tested without the tool. Productivity and competence are not the same thing. Sometimes they are inversely proportional.
What is the most powerful emotional argument for this thesis?
The argument of loss. Not abstract incompetence — but the concrete loss of something you were once able to do on your own. There is a specific category of sadness — described in psychology as functional nostalgia — connected to capacities we once had and lost through disuse.
The reader who wrote essays by hand, who navigated cities without GPS, who memorised telephone numbers — that reader already knows, viscerally, what it means to lose a capacity through delegation. AI is the same mechanism, at an incomparably greater scale and speed.
Is there an ethical risk in promoting this thesis too aggressively?
Yes. The risk is the stigmatisation of users, especially those from vulnerable groups — people with learning disabilities, with temporarily limited cognitive resources due to stress or illness, with limited access to formal education. For these groups, AI can be a tool of inclusion, not cognitive degradation.
A thesis promoted too aggressively can become, without intent, elitist: 'only the already competent deserve to use AI.' That is a morally unacceptable conclusion. The correct thesis is nuanced: conscious and calibrated use is what protects — not the absence of the tool.
How do you calibrate exaggeration so that it remains productive rather than manipulative?
Through three clear editorial practices.
The first: you exaggerate in the title and the thesis, but demonstrate the nuance in the body. The attentive reader must leave with a more complex picture than the one they arrived with.
The second: you acknowledge the thesis's limitations within the text, even if you acknowledge them with irony or in a passing sentence. Intellectual honesty, even partial, builds more authority than absolutist certainty.
The third: you propose an alternative, not merely a condemnation. 'AI makes you incompetent if you use it like this — and can amplify you if you use it differently' is a complete thesis. Only the first half is a warning without exit, which produces anxiety without direction.
The fundamental principle, borrowed from emotional psychology: strong emotions — including rhetorical ones — are messengers, not masters. Use them to open a conversation, not to close it.
Addendum — after re-reading three days later
There is a simple experiment I would recommend to any author: let the text go cold. Not for an hour. Not overnight. Three days. And re-read it as if someone else had written it.
I did that. And I found six places where the argument breathes irregularly.
I am not correcting them. I am naming them — because the tensions in a good text are more valuable than its certainties. Certainty closes. Tension invites.
The diagnosis that slides into condemnation
I wanted to write like a doctor. Sometimes I wrote like a judge.
There is a profound psychological difference between the two postures. The doctor stays beside the patient after the diagnosis. The judge delivers the sentence and leaves the room. 'The AI-assisted fool no longer knows that he doesn't know' — the phrase is accurate, but its tone drives away precisely the reader who should stay. The person who recognises himself in it closes up before he has absorbed the argument.
The psychological mechanism is classic: when we feel we are being judged, our defensive system activates before the information can reach the prefrontal cortex. Judgement blocks precisely the critical thinking the text is trying to defend.
The mechanic father and the blind spot of privilege
My mechanic father is real. His pedagogy is real. And precisely for that reason it has a limitation I did not see when I was writing.
Not all readers had that model — someone to teach them to listen before judging, to observe before concluding. For someone who never had that, AI is not a substitute for thinking — it is the first tool that has ever given them access to any form of competence. The article does not see that person.
Psychologically, the absence of an early formative model leaves a gap that people fill with whatever is to hand. To condemn the tool without seeing the gap it fills is an incomplete diagnosis. It is like treating the symptom without asking why it appeared.
The individual found guilty, the system unseen
I wrote a great deal about the unprepared user. I wrote little about the system that produces him.
The school that does not cultivate critical thinking. The economy that rewards speed, not depth. The technology industry that deliberately designs tools with interfaces that eliminate friction before the user notices what they are losing. All of these exist behind the text, unseen.
It is more comfortable to say 'use the tool consciously' than to say 'the tool is designed to make consciousness difficult.' The first version places responsibility on the individual — and leaves him alone with it. The second places responsibility on the structure — and opens a conversation about power, not merely about behaviour.
I chose the first version. It was a choice, not a neutral truth.
Friction as an unacknowledged luxury
I celebrated cognitive friction — the moment of resistance, the discomfort of not knowing, the slow effort of thinking. I spoke about it as though access to it were uniform.
It is not.
The single parent with two children and three jobs does not have the luxury of productive friction. For them, an AI assistant that responds quickly is not a shell that atrophies them — it is a release valve that keeps them functional. To say to them 'practise without the tool, deliberately, regularly' means not seeing the conditions in which they live.
Psychologically, friction is tolerable only when basic resources — time, energy, security — are assured. Below a certain threshold of chronic stress, the brain can no longer invest in voluntary difficulty. It survives. It does not grow.
Two theses that never met
The title says: AI does not make the fool more intelligent. The implicit conclusion of the text is different: AI makes everyone more dependent, regardless of their starting level.
These are two different theses, with different moral implications. The first separates people by competence and says the tool only helps those who are already formed. The second is universalist — we are all vulnerable to the same mechanism.
I moved from one to the other without explicitly reconciling them. The attentive reader felt the slippage, even if they could not name it. And a text that slips unacknowledged loses authority precisely in the moments when it should be consolidating it.
The appeal to autonomy delivered paternalistically
'Use the tool consciously. Practise without the tool. Question every answer.' The advice is valid. The tone can feel like an explanation addressed to a child.
There is a psychological distance between telling someone what to do and making them see the consequence for themselves. The first produces temporary compliance — or resistance. The second produces understanding. Understanding is the only change that lasts.
The text does this partially. Sometimes it preaches where it could simply have shown. And by preaching, it does exactly what it criticises: it delivers the conclusion in place of the process.
The catastrophist thesis and what history forgets
I wrote that AI makes the ordinary user a productive incompetent. The sentence is accurate as a diagnosis of a moment. It is false as a prediction about a species.
History never confirms the thesis that a new tool definitively atrophies a human capacity. It confirms that it displaces one. The pocket calculator did not produce a generation incapable of arithmetic — it produced a generation that counts differently and solves different problems. GPS did not destroy spatial orientation — it shifted it away from memorising maps towards reading an environment in motion. Writing itself, which Plato condemned in the Phaedrus with arguments almost identical to those in this text, did not atrophy memory — it freed it to work on tasks that oral memory could not reach.
The human brain does not lose capacities. It renegotiates them. And renegotiation leaves subjective traces which, at short range, look exactly like loss.
That is the tension my text did not see: I confused the subjective, daily perspective with structural truth. The user who can no longer formulate a complex sentence on their own feels as though they no longer can. Perhaps they can no longer do so in the old form. But they can do something else — something they couldn't have imagined before. The problem is that the new is not immediately visible. The loss is visible tomorrow morning. The gain articulates itself over years.
That does not absolve the tool. It does not invalidate a single study cited above. But it places the thesis in time — and protects it from its own exaggeration. 'Everyone becomes incompetent' is true today, in the transitional phase, when the old skill has eroded and the new one has not yet formed. It is not necessarily true as a final state.
The most vulnerable moment of any cognitive adaptation is not before the tool. It is during it — when you have given up what you knew and have not yet built what is coming. Precisely in that interstitial space, which the text called the place of understanding, lies also the place of greatest risk. And of possibility.
'What is this article not?' — and 'What remains unsaid?'
There is an honesty that any argumentative text owes its reader, and that few texts practise: to say not only what it argues, but also what it cannot argue. What remains outside its field of vision. Where the author's competence ends and where his opinion begins.
I allow myself that honesty now, at the end, when the rhetoric has cooled.
What this article does not explain
The article does not explain how AI tools should be designed so as not to produce the effects described. It diagnoses. It does not prescribe technical or pedagogical solutions. It is not a guide to conscious use, though it approaches that in section IX. The distance between 'here is what is happening' and 'here is how it is structurally resolved' remains unbridged — and it is a real distance, not a rhetorical omission.
Nor does it explain the difference between the short-term and long-term effects of AI use. The studies cited measure moments. The human brain reorganises itself over years. What happens to a user who has used permanent AI assistants for a decade — and then stopped — we do not yet know. There are no data. There are hypotheses. The article sometimes treats hypotheses as if they were data, because well-argued hypotheses are more rhetorically convincing than honest uncertainty.
What this article is not
It is not original research. It cites studies, but produces no data of its own. The author has not conducted any experiment, has not followed any group of users over time, has not measured any cognitive capacity before and after the use of a tool. The entire force of the argument rests on the synthesis of secondary sources and personal observation — both valuable, neither sufficient on its own.
Nor is it a neutral text. It was written with a predetermined thesis, sought out and demonstrated, not discovered. Sources that contradict the thesis exist — some show that AI improves critical thinking in well-structured pedagogical contexts, that how the tool is integrated makes an enormous difference, that the negative effects are conditional and reversible. These sources appear marginally in the text or not at all. That does not invalidate them. It means that the article is an argument, not a balance sheet.
The author's limitations
I am not a neurologist, a cognitive psychologist, or an education researcher. I am a practitioner who builds digital presences and who has observed, in daily work, a change in the way clients, collaborators, and he himself think through problems since AI tools entered the workflow. The observation is real. But the observation of a practitioner does not carry the weight of a controlled study.
I have a personal relationship with the subject — and that is both a resource and a risk. A resource because the authenticity of the argument comes from direct experience, not from academic distance. A risk because personal involvement produces blind spots. The section 'Addendum — after re-reading three days later' identifies seven of them. There are probably others I did not see after three days, nor after a month.
Finally: I am writing in 2026, at a moment when the field moves faster than any article can follow. Some of the claims made here will be contradicted by research that does not yet exist. Others will be confirmed. I do not know which are which. And that, more than anything else, is the real limitation of any text written now about a phenomenon that has not yet finished unfolding.
These tensions do not annul the central argument. They make it more honest.
A text without blind spots is not a profound text — it is a text that has not looked hard enough at itself. Re-reading after three days is exactly the kind of cognitive friction I recommend to the reader. It proved useful to apply it to the author as well.
Before you close this page, think of a question you have right now and follow it all the way through on your own. Not tomorrow. Now.
That is all the reader needs to do. The rest comes by itself — or it doesn't, and there is nothing to be done about that.
Addendum — After Rereading One Month Later
I received a critique. Harsh, competent, partially correct. I won't reproduce it in full — I've digested it and I'm responding to what deserves a response, not to what stings.
The critique says, in essence: the text is a well-written rhetorical pamphlet, partially grounded, but undone by exaggeration, selective nostalgia, and structural hypocrisy. That I fetishise cognitive friction from a position of privilege. That I cherry-pick studies. That I repeat "sit with this, think about it" like a motivational seminar host. That I probably used AI to write a text about how AI ruins us.
Let's be honest. How many of these are true?
The exaggeration: guilty.
The central thesis — AI produces productive incompetence — is true as a tendency, not as a universal law. I wrote it as a diagnosis, I delivered it as a verdict. The difference matters, and I ignored it deliberately, because a verdict sounds more convincing than a tendency. That's a rhetorical choice. It's also a lack of honesty.
The human brain adapts — that's not a counterargument I acknowledge late, in an addendum, to appear balanced. It's a structural truth that the text treats as an exception. It shouldn't have.
The romantic nostalgia: partially guilty.
The mechanic father is real. His lesson is real. But I turned a personal experience into a universal model — and that's exactly the mistake I reproach others for. Not everyone had that model. Not everyone needs the same kind of friction to think well. There are people for whom AI is the first tool that ever gave them access to a form of competence the education system never offered them.
I acknowledged this in point six of the three-day reread. And yet the overall tone of the text contradicts it. The body of the text and its self-critique don't inhabit the same register — and the reader feels that.
The cherry-picking: guilty with mitigating circumstances.
Yes, I selected the studies that confirm the thesis. I didn't ignore AI's benefits — I mention them at points — but I treated them as background, not as arguments of equal weight. An honest text would have built the tension between benefits and risks in the main body, not parked the benefits in parentheses.
The mitigating circumstance: a text that tries to be a complete account of an ongoing phenomenon can no longer call itself an argument. It becomes an inventory. This article is an argument. It knew that from the start. The problem is it didn't acknowledge it explicitly — it claimed to be a diagnosis, which is something more neutral than it actually is.
The OpenClaw hypocrisy: true, but irrelevant to the thesis.
I write on an SEO-optimised blog. I use digital tools. I may have used AI to structure parts of this text. The critique is correct in letter. It's irrelevant in substance — a smoker can accurately describe what smoking does to lungs. Personal involvement in the phenomenon being critiqued doesn't invalidate the observation. It complicates it, yes. It doesn't annul it.
But structural hypocrisy exists at a different, more subtle level: I wrote a text asking the reader to resist the tool, published on a platform that exists precisely because digital tools exist. That's the real tension — not whether or not I used AI to write it.
"Sit with this. Think about it." — guilty.
Yes. It's a motivational seminar technique. I used it because it works rhetorically and because I believed it produced the intended effect — a pause, a resistance. The critique is right that it sounds hollow when it appears too many times. By the third repetition, it's no longer an invitation to reflect. It's a tic.
What remains after one month
The text is partially true, partially exaggerated, well written, and honest in intention — but not equally honest in execution. I'm leaving it as it is, with all its addenda, because deleting and rewriting from retrospective comfort would be another form of hypocrisy. The article exists in time. Its limitations exist in time. And the reader who makes it this far has seen what a text looks like when it tries to look itself in the face — not perfectly, but more than most texts do.
A harsh critique is a rare gift. The worst response to it is defensiveness. The second worst is total capitulation. The useful response is to separate what is true from what merely hurts, and work with the first.
I tried that. I don't know if I succeeded.