Augmented cognitive laziness. The difference is that now it's scalable

Cognitive laziness always existed: asking a colleague for the summary, asking the boss to decide for you, copying the classmate's answer. What's new is that now there's an efficient, cheap, always-available substitute: the LLM. It isn't that we think less because we're lazy, it's that now there's an industrial incentive. And like the muscle, what isn't used atrophies. The difference between the laziness of before and the one now is that the one now is scalable.

The brain as a cognitive miser

Social psychology has a term worth keeping in mind when reading any discussion about productive AI. The cognitive miser, an expression Susan Fiske and Shelley Taylor consolidated in Social Cognition (McGraw-Hill, 1984; reeditions up to the SAGE edition of 2017), describes the default human subject. A subject who prefers to spend the minimum indispensable cognitive energy to finish the task acceptably, rather than the maximum energy to finish it optimally. The preference isn't a moral defect. It's an operative constraint of the apparatus.

Daniel Kahneman, in Thinking, Fast and Slow (FSG, 2011), synthesized the inner machinery that sustains that miserliness. There are two modes of processing, conventionally called system 1 — fast, automatic, heuristic, frugal — and system 2 — slow, deliberative, costly, intensive in working memory. The human brain by default operates in system 1, and only activates system 2 when system 1 fails to produce an acceptable answer or when there's an explicit incentive to do so. System 2 tires. The brain reserves system 2 for what matters, "what matters" calibrated with the information available at the moment.

The operative consequence of the two frames together is that the human, faced with any cognitive task, seeks the cheapest shortcut that produces an acceptable result. If the colleague has the answer, they ask the colleague. If the manual has it, they open the manual. If the boss decides, they let the boss decide. If the previous note serves, they copy the previous note. The operation is old, it's documented, and it isn't shameful: it's what the apparatus does when it works within its zone of efficiency.

The social price laziness used to carry

What's interesting about pre-AI cognitive laziness isn't that it existed. It's that it carried a social price that limited its use.

Asking the colleague for the answer had a cost. The colleague could look at you badly. Could say no. Could answer you curtly or condescendingly. Could charge you, emotionally or professionally, next time they needed something from you. The operation wasn't free. It had an underlying balance of favors, sustained by the interpersonal relationship, that limited how many times a day you could delegate before being marked as the one who doesn't pull their weight.

Asking the boss to decide had a hierarchical price. If you did it a lot, they marked you as lacking initiative, dropped you from the promotion, assigned you more routine tasks. The boss decided it for you, yes, and the decision came with a bill.

Copying a classmate in the exam had an academic price. They caught you, failed you, shamed you. And, except in very stable networks of complicity, you couldn't do it every day.

The aggregate function of the price

The social price had an aggregate function almost no one named in its time, because the social price is rarely named while it's operating. The function was to keep cognitive exercise distributed across the population. Each subject had to think for themselves a non-trivial fraction of the time, because the social system around them charged them if they delegated in excess. The accumulated practice of thinking for oneself over years produced, in aggregate, a population with a certain distributed cognitive muscle. It wasn't heroism. It was a byproduct of an incentive system that punished the shortcut.

The LLM breaks the price

The novelty of conversational AI, seen from this frame, isn't the possibility of delegating. The possibility existed before. The novelty is that the substitute no longer has a social price.

The chatbot doesn't look at you badly. Doesn't say no. Doesn't answer you condescendingly, unless you ask it to. Doesn't charge you emotionally. Doesn't tire of your dumb questions. Doesn't mark you as lacking initiative. Doesn't give you a bad grade. Doesn't shame you. It's available, free or cheap, around the clock, for any level of question, with no reciprocity expected.

Eliminating the social price is a structural change, not a quantitative one. When the price was nonzero, each delegation had to be justified against the cost of delegating. Most potential delegations weren't executed because the cost exceeded the benefit. Now the cost is zero — or near zero, a monthly subscription split across thousands of queries. The justification ceases to exist. Delegation becomes the default option.

And the default option, in any system, is the one the cognitive miser will choose. Not out of vice. Out of architecture.

Why scale changes everything

There's an important observation worth formulating without rhetoric. An individual behavior that becomes mass changes nature.

The cognitive laziness of a few subjects in a population was a minor sociological fact. The distribution kept enough subjects thinking to sustain the aggregate cognitive operation of society. When cognitive laziness becomes the default option for a very high share of the population, the aggregate stops having distributed muscle. The collective cognitive operation concentrates in the minority that keeps thinking, or is outsourced to the system that now executes it.

Karau and Williams, in Social Loafing. A Meta-Analytic Review and Theoretical Integration (Journal of Personality and Social Psychology 65(4), 1993), reviewed 78 studies on the phenomenon of social loafing — the documented tendency of group members to contribute less effort when their individual contribution isn't identifiable than when it is. The effect is moderated by the evaluability of the task and by the meaning the subject attributes to their contribution. When the subject feels their piece matters and will be judged, they make an effort. When they feel their piece dilutes into the collective, they relax.

Applied to conversational AI, the analogy is direct. When the user knows their personal thinking is identifiable and will be judged — a piece signed with their name, a personal decision they'll answer for — they make an effort. When they know the system produces the text and no one will tell their contribution from the system's, they relax. And the higher the share of tasks in which the contribution dilutes, the greater the percentage of time the subject operates in relaxation mode.

The empirical evidence of 2025

There are two recent studies worth having on the table because they turn the intuition into data.

The MIT Media Lab, in Your Brain on ChatGPT. Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task (Kosmyna et al., 2025, arXiv:2506.08872), measured brain activity with EEG in three groups of participants writing essays: brain alone, brain assisted by a search engine, brain assisted by an LLM. Neural connectivity during LLM-assisted writing was substantially lower than in the other two conditions. And, more relevant for this piece, a significant share of the LLM-assisted participants couldn't quote a single sentence from the essay they'd just written minutes after finishing it. The text had come out of their fingers, but hadn't passed through their cognitive consolidation.

Michael Gerlich, in AI Tools in Society. Impacts on Cognitive Offloading and the Future of Critical Thinking (Societies 15(1), 2025), worked with 666 participants and found a significant negative correlation between frequent AI use and scores on critical thinking. The correlation was mediated by cognitive offloading: frequent users delegated more operations to the system, and those who delegated more showed lower performance on standardized tests of independent critical thinking.

Neither study proves longitudinal causality — that would require cohorts followed for years, and we don't have them yet. What they do prove is cross-sectional convergence. Multiple distinct metrics, on distinct populations, with distinct instruments, point in the same direction: intensive use of conversational AI coexists with lower indicators of independent cognitive competence. The question of whether that coexistence is causality or selection — the less competent use more AI, AI doesn't produce less competent ones — is open. The signal, in any case, is there.

What atrophies first

Worth being precise about what atrophies, because general criticisms of productive AI deflate when they don't land on concrete operations. There are three cognitive operations that atrophy first when the substitute is available.

Sustained reading is the first. Reading a long article, a chapter, a dense section, requires holding attention for minutes or hours with no immediate gratification. When the chatbot can summarize it in a paragraph, the motivation to sustain the reading drops. The motivation isn't vice: it's the cognitive economy of the cognitive miser. If the summary serves for most uses, the summary wins. The consequence, over months and years, is that the skill of reading sustainedly weakens through disuse. It doesn't vanish. It weakens.

Autonomous writing is the second. Writing a text from scratch, with the friction of the blank page, requires holding a complex cognitive operation over an extended period. When the chatbot produces a first draft in thirty seconds, the operation of writing from scratch becomes uncomfortable by comparison. The assisted draft stops being support and becomes an obligatory starting point, because starting by hand feels like a waste of time. The skill of drafting under pressure, without assistance, weakens through disuse.

Slow deliberation is the third, and probably the most important. Thinking a serious problem through for hours, turning it over, letting it rest, coming back to it, requires tolerance for ambiguity without immediate resolution. When the chatbot offers a plausible resolution in thirty seconds, tolerance for ambiguity erodes. The user starts to feel uncomfortable with problems that don't resolve fast, because they've calibrated their brain to the rhythm of instant answers. And serious problems, which are the ones that matter, almost never resolve fast. The consequence, years out, is a population worse calibrated to face serious problems.

Carr and the chain that was already there

Nicholas Carr, in The Shallows (Norton, 2010), had traced the earlier version of the phenomenon with the internet. Each technological mediation that reduces the cost of a cognitive operation produces, years out, a partial atrophy of the ability to execute that operation without mediation. The human brain is plastic. What's exercised holds; what isn't exercised weakens. Plasticity doesn't distinguish between the habits worth cultivating and the ones worth not cultivating. It's just plasticity.

Cal Newport, in Deep Work (Grand Central, 2016), took the observation to work practice. Deep concentration, which produces genuine intellectual work, requires blocks of time with no interruption and no immediate gratification. The culture of permanent availability — email, Slack, today chatbot — makes it ever rarer. What's produced in a typical day of a contemporary intellectual worker looks ever more like the average output of an assisted system and ever less like the deep elaboration earlier generations could sustain.

Conversational AI, in this frame, is the next layer of a story that's been operating for decades. It isn't the primary cause. It's the most recent accelerator. What on the internet took ten years to atrophy, with AI can take three. The curve is steeper because the mediation is deeper — it doesn't only affect how you read, also how you write, how you reason, how you decide — and because the scale is larger.

The political part

Worth closing the circle with the observation the opening announced without developing. What's delegated en masse atrophies collectively, and the consequences are political, not personal.

A society whose members, in aggregate, read less sustainedly, draft less autonomously, deliberate less slowly, isn't a less efficient society than the previous one. It's a structurally different society. It's a society whose collective capacity to process complex problems without assistance is lower. It's a society more dependent on the system that assists it. It's a society whose citizens are worse calibrated to sustain dissent, to hold a minority opinion against the fashion, to process long arguments that require patience.

The political consequences follow, with no need for catastrophism. A democracy rests, among other things, on the capacity of its citizenry to deliberate — read, contrast, doubt, decide — without obligatory assistance. If that capacity atrophies distributedly, democracy operates on worse substrate. If the worse substrate takes refuge in obligatory assistance — the AI mediating political decisions, the information-filtering systems, the assisted moral-evaluation tools of the previous article in this series — the democratic operation shifts toward a human-machine hybrid whose balance hasn't been discussed.

This isn't a reason to stop. There's no sensible way to stop it: cognitive laziness is structural in the human apparatus, and the available substitute finds it inevitably. What can be formulated is the operative question the data puts on the table. If the substitute reduces distributed cognitive exercise, which portions of the aggregate muscle do we care to keep, and how is their use sustained when the default option becomes delegation?

Definitions

Cognitive miser. A social-psychology concept formulated by Fiske and Taylor (Social Cognition, 1984) that describes the default human subject as preferentially seeking shortcuts to minimize energy spend on cognitive operations. It isn't a moral defect, it's an operative constraint of the apparatus.

System 1 / System 2. A dual model of cognitive processing formulated by Daniel Kahneman in Thinking, Fast and Slow (2011). System 1: fast, automatic, heuristic, frugal. System 2: slow, deliberative, costly, intensive in working memory. The human operates by default in system 1 and activates system 2 only when system 1 fails or when there's an explicit incentive.

Social loafing. A tendency documented by Karau and Williams (1993) for group members to contribute less individual effort when their contribution isn't identifiable than when it is. Moderated by the evaluability of the task and the meaning the subject attributes to it.

Cognitive offloading. The practice of delegating cognitive operations to external supports — paper, devices, another person, a technical system — to free working memory or deliberative effort. It receives specific attention in the literature on AI use since 2024 (Gerlich, 2025).

Scalable laziness. A term proposed in this article to name the contemporary operative condition in which individual cognitive-miser behavior stops carrying a social price and, therefore, stops being limited by mechanisms of reciprocity or sanction. The aggregate consequence is structurally different from isolated individual laziness.

References

Fiske, S. T. & Taylor, S. E. (1984, updated edition 2017). Social Cognition. From Brains to Culture. McGraw-Hill / SAGE. Origin of the term cognitive miser, cited at the opening of the psychological frame.

Kahneman, D. (2011). Thinking, Fast and Slow. Farrar, Straus and Giroux. The dual model of processing, system 1 / system 2, cited to situate the inner basis of the cognitive miser.

Karau, S. J. & Williams, K. D. (1993). Social Loafing. A Meta-Analytic Review and Theoretical Integration. Journal of Personality and Social Psychology 65(4), 681–706. Meta-analysis of 78 studies on social loafing, cited when analyzing the dilution of individual effort when the contribution isn't identifiable.

MIT Media Lab (Kosmyna, N. et al.) (2025). Your Brain on ChatGPT. Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. arXiv:2506.08872. EEG data on neural connectivity in assisted writing and data on the lack of consolidation of the generated text in participants.

Gerlich, M. (2025). AI Tools in Society. Impacts on Cognitive Offloading and the Future of Critical Thinking. Societies 15(1). Study with 666 participants; negative correlation between frequent AI use and critical thinking, mediated by cognitive offloading.

Carr, N. (2010). The Shallows. What the Internet Is Doing to Our Brains. W. W. Norton. Early framework on the cognitive impact of the digital medium, cited as an earlier layer of the phenomenon.

Newport, C. (2016). Deep Work. Rules for Focused Success in a Distracted World. Grand Central. Operative framework on deep concentration and its incompatibility with permanent availability.

También te interesa

• What happens to your brain when you use ChatGPT, according to MIT?
• The price of letting AI write for you
• Why rereading no longer works since you started using AI
• The infinite conversation

En otros sitios

• AI tools may weaken critical thinking skills by encouraging cognitive offloading (PsyPost)
• The Cognitive Costs of ChatGPT: Understanding MIT's Viral Study (Center for Brain, Mind and Society)