The Age of Disinterest: Artificial Intelligence, the Abandonment of Effort and the Prospect of a New Dark Age

1. The dream of removing effort and the life that disappears

In Click (2006), Michael Newman receives a remote control that allows him to fast-forward through uncomfortable parts of his life. The device seems to solve an efficiency problem: eliminate waiting, conflict, fatigue and repetition and arrive immediately at the rewards. But the remote begins to determine automatically which experiences are worth living. Michael discovers that the moments he discarded were not gaps between important events. They constituted his life. [1]

The film is neither a technological forecast nor a depiction of generative AI. It offers a useful image of a behavioural mechanism: when avoiding an activity also eliminates the experience through which a skill or relationship develops, time savings may undermine autonomy. Learning a language requires hesitation, mistakes, recollection and conversation. Learning a musical instrument requires repeated practice, sore fingers, rhythm and listening. Learning mathematics means working through the frustration of not understanding. A system that supplies the translation, song or answer delivers the product while potentially bypassing the process through which the learner would have changed.

The tension between efficiency and experience is ancient. In Plato's Phaedrus, Socrates expresses anxiety that writing, by storing knowledge outside the mind, may weaken the exercise of memory. That objection did not show that literacy destroys intelligence. Written culture vastly enlarged intellectual life. The analogy matters because cognitive technologies change where a capability is performed. The historical question is not whether external assistance inevitably produces decline, but which abilities are lost, which are expanded and who retains the ability to correct the tool.

A machine completing a task need not destroy the human desire to do it. Computers surpassed world chess champions, yet human chess did not disappear; synthesisers did not end instrumental music; cameras did not end painting. Economic usefulness and human significance are different. A competence may cease to be indispensable in a market while remaining central to the experience of acquiring it. The gravest danger arises when a person comes to see no value in a capability beyond the efficiency with which its output can be obtained.

2. What would a technological Dark Age actually mean?

The expression “Dark Ages” is contested by historians. Applied indiscriminately to the Middle Ages, it erases intellectual continuities, universities, manuscript traditions and scientific achievements across different societies. As a contemporary metaphor, it is useful only when attached to a testable mechanism rather than nostalgia for an imagined age of universal diligence and literacy.

A plausible form of technological obscurity would be an asymmetry between systemic capability and individual competence. Hospitals might have extraordinary diagnostic systems but too few professionals able to understand the basis for their recommendations. Governments might process vast quantities of data while losing the internal expertise required to challenge suppliers. Firms might generate millions of lines of code yet struggle to train engineers capable of maintaining critical systems without continuous assistance. People might receive information in every language while becoming less able to assess its quality.

This would not be a return to an agrarian economy, medieval censorship or a world without electricity. It would be a technologically illuminated but socially dependent world, with knowledge concentrated in institutions and platforms that a small minority could meaningfully audit. Its resilience would depend on the reliability of infrastructure, transmission between generations and reserves of human expertise.

Four conditions would make the hypothesis persuasive: measurable deterioration in unaided learning; sustained reductions in public and private investment in foundational training; dependence on systems beyond meaningful inspection or oversight; and declining ability to recover processes when automation fails. A population that uses AI intensively but performs better without it would weaken the educational part of this claim. Firms that expand apprenticeship and build workers' independent judgement would supply contrary evidence. The existence of chatbots alone establishes none of these outcomes.

3. The long history of replacing human effort

Agricultural mechanisation changed the demand for physical strength and local farming knowledge. Industrialisation separated the design and execution of many tasks, allowing scale and specialisation while making employees dependent on production systems they did not fully control. In Capital, Marx examined machinery not merely as a productive force but as part of the reorganisation of labour under capital. The outcome is not determined by a machine in isolation: ownership, workplace discipline, bargaining and the distribution of productivity gains also matter.

Mass production lowered the cost of goods once inaccessible to ordinary households, replacing some artisanal skills with standardised processes. It expanded consumption and created occupations, while narrowing the economic role of certain forms of knowledge acquired through prolonged apprenticeship. Twentieth-century computers automated calculations, accounting, filing and routine administration. The internet and search engines lowered the cost of finding information without removing the need to interpret it.

Generative AI extends the process to tasks whose execution appeared to require the human formulation of explanations, images, texts, designs and programs. Unlike a calculator, it can present a false answer with rhetorical confidence. Unlike a closed industrial machine, it operates across heterogeneous tasks and changes rapidly in performance. Delegation can therefore encompass part of problem formulation, not simply the calculation of an answer.

Historical comparisons also require restraint. Technology displacing particular tasks does not necessarily eliminate motivation to master an occupation. Literacy survived the proliferation of books; photography and cinema expanded some artistic practices rather than extinguishing them. The distinctive contemporary change must be demonstrated through the relationship among learning, remuneration and autonomy rather than inferred from impressive product demonstrations.

4. Ellul, Arendt, Illich and Stiegler: when means begin to choose ends

In The Technological Society, originally published in French in 1954, Jacques Ellul described “technique” as a system of methods oriented towards maximum efficiency that increasingly reorganises institutions and individual choices. His concern was not that devices possess malicious intentions. It was that technical efficiency can become the determining criterion for decisions about work, education and private life even where it should not be the dominant value.

Hannah Arendt's The Human Condition (1958) distinguished labour associated with biological necessity, work that makes a durable human world, and action among persons. This is not a theory of AI, but it illuminates a separation between efficient production and political participation, responsibility or meaning. Automating burdensome activity may free time. It cannot determine what citizens will do with that time.

Ivan Illich's Tools for Conviviality (1973) distinguished tools that enlarge users' initiative from systems that create institutional dependence and reduce people's capacity to act. Applied to AI tutors, the issue is not the number of answers they supply but whether users can formulate further questions, correct errors and learn without ongoing assistance. A powerful technology can strengthen autonomy when it remains understandable, adaptable and integrated into serious educational practice.

Bernard Stiegler developed a more direct account of the externalisation of knowledge. His conception of proletarianisation extends beyond the loss of wages or property to include the loss of practical knowledge and judgement as processes move into technical systems. A society might have access to millions of instructions and lose the experience required to decide which ones to apply. The risk is an employee who operates a system without understanding it and a citizen who receives decisions without being able to challenge them.

Neil Postman's Technopoly (1992) questioned the tendency to make technological systems and quantification sources of cultural authority. Richard Sennett's The Craftsman (2008) emphasised patient engagement with tools, materials and practical problems. Matthew Crawford's Shop Class as Soulcraft (2009) explored craftsmanship, judgement and independence. These authors did not literally anticipate generative models in 2026. They explain a potential loss: when doing ceases to be a source of understanding, a person may become a consumer of results whose quality they cannot judge.

5. Theodore Dalrymple and the danger of mistaking feeling for understanding

In Spoilt Rotten: The Toxic Cult of Sentimentality, published in Brazil as Podres de mimados, Theodore Dalrymple criticised a culture that confers special moral status on the expression of feelings and weakens the habit of testing impressions through reason. [2] His book is not about AI and does not establish that younger generations are inherently unable or unwilling to study. It belongs to a conservative tradition of cultural criticism whose normative generalisations must be distinguished from experimental evidence.

A connection with digital systems emerges when subjective satisfaction replaces independent criteria of competence. An apparently sophisticated essay may be accepted without understanding. A news item confirming personal identity may be shared without verification. A student may feel confident after receiving a persuasive explanation even though they cannot solve a similar problem alone. Confusing the sensation of mastery with actual mastery favours a culture of apparent achievement.

The argument has a fundamental limit. Emotions are not enemies of learning. Curiosity, belonging, manageable frustration and the satisfaction of accomplishment contribute to motivation. Discrediting students' or workers' feelings does not automatically create discipline. Educational psychology identifies more precise mechanisms than a simple opposition between emotion and reason. Dalrymple's critique is useful in questioning the substitution of judgement by self-affirmation; it does not demonstrate that empathy, social protection or educational reform cause cognitive decline.

6. The psychology of effort: autonomy, competence and immediate rewards

Self-determination theory, associated with Edward Deci and Richard Ryan, distinguishes motivation supported by autonomy, perceived competence and relatedness to others. Learning is more likely to endure when individuals help set goals, experience progress and find social meaning in an activity. Instant answers can weaken motivation if they remove the experience of becoming competent. The same answers can strengthen motivation if they remove an initial barrier and enable progress towards previously inaccessible work.

Albert Bandura studied self-efficacy: confidence in one's capacity to perform actions needed to achieve a goal. Success built through effort and feedback can reinforce that confidence. Receiving a ready-made output without understanding its production does not necessarily yield an equivalent experience. Mihaly Csikszentmihalyi described “flow” in activities balancing challenge and ability. This is not a defence of pointless suffering: tasks that are too easy can disengage people, while impossible demands provoke withdrawal. The constructive objective is to preserve challenges that learners can overcome.

The attention economy adds external incentives. Digital interfaces compete for time with continual updates, personalised recommendations and rapid rewards. Reading a substantial book, studying circuits or practising guitar scales promises slower, uncertain and frequently private returns. Entertainment promises an immediate and observable response. Yet attributing all concentration difficulties to dopamine oversimplifies neuroscience. Reward, habits and attention cannot be reduced to one chemical or a generational diagnosis.

Motivation also depends on financial security, sleep, nutrition, family support, teachers and credible expectations for the future. Students required to work, exposed to violence or denied opportunities cannot adequately be described through the moral category of laziness. Technology may intensify inequalities that preceded it, but did not create them.

7. Cognitive offloading: producing better is not necessarily learning more

Research on cognitive offloading examines how external resources reduce demands on memory and processing. Notes, calendars and calculators serve similar purposes. Offloading may free capacity for complex problems or reduce practice of basic abilities. The effect depends on which processes are delegated and how the recovered time is used.

A CHI 2025 study of 319 knowledge workers, using 936 reported task examples, found an association between greater confidence in AI and less self-reported exercise of critical thinking. It also observed a shift towards checking information, integrating outputs and supervising tasks. Because the evidence relies substantially on participant reports, it does not establish permanent loss of cognitive ability caused by the tool. [3]

An exploratory MIT essay-writing study initially circulated as a 2025 preprint involved 54 participants across its first three sessions and 18 in a fourth. It reported differences in EEG connectivity measures, recollection and perceived ownership of writing among participants using AI, search tools or no tools. The small sample, specific task and interpretive limits of EEG do not permit conclusions about irreversible brain changes across a whole generation. [4]

Educational trials point in opposite directions. A 2025 randomised study by Kestin and colleagues found greater learning in less time with an AI tutor grounded in pedagogical design when compared with an active-learning classroom format in the setting studied. Conversely, a working paper released in October 2026, drawing on a randomised trial involving 2,379 university students, reported lower final grades and participation in learning systems in classes offered an integrated AI tutor; losses were particularly concerning among first-generation students. The latter results require further scrutiny, and the two interventions were not identical. [5][6]

The OECD's Digital Education Outlook 2026 describes the same underlying tension. Better task performance with general-purpose AI does not consistently produce sustained learning once the technology is unavailable; tools designed for educational purposes can support lasting gains. The key variable is no longer simply whether AI is accessible. It is whether the learner continues practising the competence that education is supposed to develop. [7]

8. Schools, universities and technical education: what loses value and what stays scarce

Formal education provides more than access to information. Schools develop basic language, social cooperation, perseverance and shared rules. Universities organise advanced study, laboratories, research, professional networks, peer assessment and entry into regulated occupations. Technical schools teach practical skills with machinery, electrical systems, mechanical equipment, instruments and safety procedures. Each can fail to fulfil its mission; none can be reduced to a collection of textual answers.

When a chatbot explains an integral in seconds, a student may ask why calculus must be studied. The economic answer depends on the use of knowledge. If only a numerical result is required, external computation may suffice. To design a bridge, investigate a failure, prove a mathematical claim or detect an error in the machine's calculation, conceptual understanding remains essential. In activities involving physical, legal or financial risk, trusting an answer without understanding its assumptions may cost far more than learning the subject.

There is nevertheless a potential crisis of credentials. When assignments, essays and code can be generated without understanding, grades based on deliverables become less informative about a student's capabilities. Universities may respond through oral examinations, practical projects, continuous assessment, laboratory work, defence of decisions and tasks performed without assistance. Such approaches demand teaching time and can undermine educational business models based on scale and automatic marking.

In Education at a Glance 2025, the OECD estimated that full-time workers with tertiary education earned, on average, 54% more than workers with upper-secondary schooling across the countries and populations compared. In Brazil, the observed earnings difference exceeded 140%. These are associations reflecting occupational composition and selection; they are not guaranteed causal returns for every individual degree. They nonetheless contradict the assertion that education has already lost its economic value. [8]

The longer-term problem is more complex. Does the wage premium arise from skills actually acquired, the signalling value of a credential, professional networks or legal barriers to occupational entry? If software compresses pay for tasks previously restricted to graduates, some returns could decline. If AI makes knowledgeable supervisors more productive, other returns could rise. The relative value of technical qualifications will also depend on robotics, infrastructure, professional licensing and the demand for reliable, accountable in-person work.

9. The disappearance of the apprentice: a failure of intergenerational transmission

Experts do not arise from reading alone. Physicians learn through supervised cases; engineers through projects, review and failure; programmers through maintenance, debugging and correction; technicians through observation, diagnosis and repetition. Many tasks treated as simple are part of the training required to assume complex responsibilities.

A firm may automate entry-level work while keeping experienced staff. It saves money in the short run. If it ceases hiring and supervising beginners, it may eventually lose the source of future experts. The problem will not immediately register in headline employment statistics: an organisation can show higher productivity while consuming knowledge accumulated through earlier generations' training.

In 2025, the International Labour Organization estimated that approximately one-quarter of workers globally held occupations with some exposure to generative AI, stressing that most such occupations would more plausibly be transformed than abolished. The IMF estimated in 2024 that roughly 40% of global employment was exposed under a different method. These are not measures of layoffs, their definitions differ and their percentages cannot be combined. [9][10]

Economist David Autor proposed a countervailing possibility: AI could extend decision-making capacity to workers with complementary experience, rebuilding opportunities in middle-skill occupations. Daron Acemoglu cautioned that aggregate economic gains depend on tasks that can actually be automated, real costs and complementarities rather than optimistic forecasts. The erosion of occupational apprenticeship is a legitimate risk, not an established or inescapable aggregate outcome. [11][12]

10. The political economy of disinterest: returns, wages and ownership

Acquiring a skill is an investment in human capital. The return may include future wages, personal satisfaction, entrepreneurship, mobility and security. Costs include study time, tuition, equipment, forgone earnings and the risk of failure. When technology alters the expected remuneration for a competence, it changes the economic calculus even for disciplined, ambitious people.

If AI reproduces routine commercial writing cheaply, demand for some forms of training may decline. If it increases demand for verification, systems integration, electricity, production engineering and risk assessment, incentives to learn those functions may grow. As content becomes abundant, trust, legal accountability, exclusive data and experienced judgement may become relatively more valuable.

Market structure influences who receives the benefits. Firms controlling models, chips, cloud infrastructure, distribution platforms and databases may earn profits and fees even when users perform less intellectual labour. If a rising share of income accrues to owners of technical systems, households without assets may become more dependent on transfers, complementary jobs or public services. This does not prove the emergence of a society without work; it shows that productivity and distribution are separate questions.

Automation can also reduce prices, enable new economic activities and expand demand. Assuming every efficiency gain necessarily destroys jobs neglects market expansion, new tasks and changed consumption patterns. Conversely, claiming that new occupations will always compensate for losses within the same generation, region or wage bracket ignores transition costs and unequal rates of adjustment.

In political terms, disinterest may become a rational response to deteriorating incentives. Young people need not despise knowledge; they need only conclude, based on recruitment and pay, that years of training are unlikely to be rewarded. The institutional answer cannot consist solely of moral campaigns praising effort. It must maintain credible paths to mobility, professional apprenticeship, competition and access to productivity gains.

11. Brazil: unequal learning long before cognitive automation

Brazilian educational vulnerability predates generative AI. In PISA 2025, 28% of Brazilian students reached at least Level 2 proficiency in mathematics, against an OECD average of 65%. In reading, the respective shares were 49% and 69%. These gaps reflect long-standing differences in learning opportunities and school conditions; they are not evidence that generative tools caused the problem. Average Brazilian test results remained broadly stable between 2022 and 2025. [13]

Source: OECD, PISA 2025 Brazil country note [13]. International proficiency comparison, not a causal AI effect.

An exclusively pessimistic interpretation would be incomplete. The same PISA survey found that 72% of Brazilian students reported curiosity about many subjects, close to the OECD average of 73%. Some 59% said they made extra effort when work became difficult, compared with 60% across the OECD. Only 16% agreed that school was a waste of time, below the OECD average of 24%; the Brazilian share had fallen by 6.9 percentage points between 2022 and 2025. A society with weak learning outcomes is not necessarily a society that rejects learning. [13]

Source: OECD, Brazil country note, PISA 2025 [13]. Performance measures and student attitudes come from different instruments and do not form a single index of disinterest or imply causality.

Source: OECD, PISA 2025 Brazil country note [13]. Self-report, not a measure of proficiency or behaviour.

The fact that 48% of Brazilian students reported using chatbots at least weekly to assist learning makes it urgent to measure what happens without tools. AI might extend access to explanations and practice in cities lacking specialist teachers; it might also enable certificates without mastery. Brazil must avoid two simplistic conclusions: that technology makes school dispensable and that blanket bans will solve poor learning.

Electrical training, nursing, maintenance, mechanics, automation and other practical disciplines depend on facilities, instruments and safety procedures. A simulation can support instruction but cannot automatically replace the ability to perform a measurement, diagnose a defective component or accept responsibility for an operation. Brazil's opportunities depend on integrating AI with foundational skills and supervised practice rather than undermining the ability to perform and verify work.

The geographical contrast does not establish a world ranking of AI-driven learning loss. It identifies evidence drawn from different populations, measures and causal designs [5][6][9][13].

12. The politics of comfortable ignorance

Democracy assumes that citizens can assess promises, compare claims, recognise interests and hold officials accountable. It has never required every person to be a technical expert. It does require intermediate institutions — journalism, schools, courts, universities, professional bodies and public administration — capable of producing knowledge that can be challenged and verified.

When information is cheap and abundant, human attention and judgement become scarcer relative resources. Political campaigns can generate numerous versions of messages tailored to different audiences; platforms can personalise arguments, images and emotions; governments can use automated systems to assign benefits, priorities and enforcement attention. Technological possibility does not mean that these practices are universal or inevitable. The institutional concern lies in the asymmetry between those who design systems, those who control data and those who receive decisions.

Outsourcing judgement can occur within the state itself. An agency procuring software without staff able to audit its decision criteria can lose administrative autonomy. When decisions must be explained to citizens, the right to appeal requires knowledgeable people capable of identifying mistakes. Without them, formal human accountability may remain in place even as practical control moves to suppliers.

The issue cannot be resolved through a simplistic opposition between democracy and technology. Carefully evaluated public systems can improve transparency, detect irregularities, translate documents and extend services to remote populations. The risk arises when automation reduces the possibility of challenge or when short-term budget savings dismantle state capabilities that are difficult to rebuild.

13. Technical acceleration and the slower pace of human formation

Accelerationism encompasses different traditions. Nick Land described self-reinforcing processes involving capitalism and technology. Alex Williams and Nick Srnicek argued for a left-wing political project that would mobilise technological capacities beyond existing social constraints. Contemporary effective accelerationist currents place greater weight on innovation, competition and the expansion of technological capability. These perspectives do not form a single doctrine or propose identical positions on employment or education.

Across them lies a temporal conflict. Models and digital infrastructure may change within months; forming a physician, engineer, researcher or institutional culture of judgement takes years. The rate at which answers are produced may outrun the rate at which people capable of checking them are trained. A technological Dark Age need not arise from a conspiracy to make populations less capable. It could result from decentralised decisions: students optimise grades, firms cut training, universities minimise costs, governments purchase turnkey solutions and platforms simplify interaction.

These chains describe possible feedback, not historical laws. A school that tests independent understanding preserves an incentive to study. A firm that measures decisions and finances supervised learning can use technological progress to accelerate expertise. Users who can inspect models, transfer data and switch vendors may diffuse knowledge rather than concentrate it. Institutional policy changes the direction of the process.

14. A cognitive aristocracy in an economy of abundant answers

Digital inequality extends beyond internet access or subscription prices. Two people can use the same AI system yet obtain very different social outcomes. Someone with substantial training in mathematics, statistics, programming, writing, history and methods of verification can frame better questions, recognise ambiguity and identify mistakes. Someone lacking those skills may accept the first plausible response.

If educational systems divide into two regimes, the risk intensifies. In one, well-resourced students would have teachers, laboratories, mentors, sustained reading, rigorous mathematics, instruments and independent assessment, using AI as an extension of competence. In the other, less advantaged students would receive automatically generated content, low-cost platforms with limited supervision and credentials certifying attendance rather than mastery. The technologies might appear similar while functioning in profoundly different ways.

Such a division could create a cognitive aristocracy determined not by innate talent, but by access to demanding educational experiences. The paradox is that tools advertised as the democratisation of knowledge could entrench inequality if the privileged preserve rigorous learning while other groups are encouraged to replace it with convenience.

The reverse remains plausible. Accessible, open, pedagogically tested tutoring systems could reduce shortages of individual instructional support. Real-time translation could assist excluded students; disability support could expand independence. Distributional outcomes depend on design, assessment, infrastructure and governance, not on a metaphysical property of artificial intelligence.

15. Music, sport and crafts: the value of doing what no longer needs doing

Chess provides a revealing historical comparison. Machines long ago surpassed even elite human players, yet human tournaments, clubs, study and competition persist. Their value is not merely producing the mathematically strongest possible move. It lies in the experience of one person confronting another, developing patterns and sustaining decisions under pressure.

Musicians face a similar prospect. AI can generate harmonies, sounds, solos and complete songs without anyone learning an instrument. Playing guitar, however, develops motor coordination, listening, time perception and relationships with others. The final sound is only part of its value; the rest is in the process. An automated instrument can perform a passage with exceptional precision without undergoing effort, personal history and social presence as a human musician does. Recognising this significance does not require claiming that human performances will never be commercially displaced.

The same principle applies to drawing, building furniture, cooking, running, programming out of curiosity or learning a language without professional compulsion. In an economy of abundance, practices can persist because people want to undertake them. The danger is not every skill's economic obsolescence; it is a culture that evaluates human activity solely by the cost of reproducing its finished product.

The most promising use of automation is not to abolish effort, but to eliminate sterile effort and preserve time for experiences through which people develop. This requires a concept of progress different from simply maximising outputs: a prosperous society should give people more opportunities to learn to do what they value, not fewer reasons to try.

16. Three futures from 2027 to 2040: erosion, division or flourishing

These are conditional trajectories, not forecasts with measured probabilities. Within six months, informative signals would include school assessment policies, the design of AI tutoring and entry-level hiring. Over two years, researchers could compare unaided retention, apprenticeship programmes, early-career pay and educational persistence across institutions adopting different policies. Over five years, the central question will be whether productive capacity and human expertise have advanced together or whether increasingly sophisticated services depend on a shrinking group able to understand them.

The first path does not require every metric to deteriorate: output may rise while autonomy falls. The second allows innovation and aggregate prosperity alongside cognitive segregation. The third does not eliminate political conflict or regulation; it assumes institutions deliberately preserve the difficulties that remain essential to learning.

17. The political choice: preserve the capacity to learn

Banning all AI tools would not automatically restore attention, school quality or investment in training. Delegating all learning to platforms would not solve those problems either. The measurable objective is that students know more after using tools, workers become more capable of making decisions and institutions retain enough expertise to correct the systems on which they rely.

Schools need to separate polished answers from understanding: unaided tests when appropriate, oral defences, experimental projects, sustained reading, progressive exercises and tutors that ask before they answer. Universities need to preserve research, laboratories, debate and supervision rather than competing solely on cost with automated content producers. Firms need viable entry routes and apprenticeship so that they do not consume their stock of expertise without replenishing it. Governments need the capacity to evaluate, procure and challenge automated systems, even when private systems perform specific tasks more efficiently.

Such measures have different costs across unequal contexts. Overburdened teachers and students without adequate resources are not helped by slogans about discipline. Educational policy must fund teacher development, infrastructure, protected learning time and independent assessment of pedagogical technologies. Rights of access and appeal, transparency in public systems and competition policy can prevent delegated tasks from automatically becoming delegated control.

A society does not decline merely because its machines do more. It is endangered when it loses the desire, the opportunities and the institutions required to develop people able to understand what those machines do. The possible Dark Age of the twenty-first century would not be a world without answers. It would be a world in which almost everything seems answered, fewer people keep asking difficult questions and fewer still know how to discover when an answer is wrong.

Sources and references

[1] Sony Pictures, Click (2006), official synopsis: https://www.sonypictures.com/movies/click

[2] Theodore Dalrymple, Spoilt Rotten: The Toxic Cult of Sentimentality; Brazilian edition Podres de mimados, É Realizações: https://www.inaeditora.com.br/produtos/podres-de-mimados/

[3] Lee, H.-P. et al., “The Impact of Generative AI on Critical Thinking”, CHI 2025: https://www.microsoft.com/en-us/research/publication/the-impact-of-generative-ai-on-critical-thinking-self-reported-reductions-in-cognitive-effort-and-confidence-effects-from-a-survey-of-knowledge-workers/

[4] Kosmyna, N. et al., Your Brain on ChatGPT, preprint, 2025: https://arxiv.org/abs/2506.08872

[5] Kestin, G. et al., “AI tutoring outperforms in-class active learning”, Scientific Reports, 3 June 2025: https://doi.org/10.1038/s41598-025-97652-6

[6] Liu, J. et al., The Effects of Course-Integrated AI Tutoring on Student Performance and Engagement, EdWorkingPaper, October 2026: https://doi.org/10.26300/y3f8-vh05

[7] OECD, Digital Education Outlook 2026: https://www.oecd.org/en/publications/oecd-digital-education-outlook-2026_062a7394-en.html

[8] OECD, Education at a Glance 2025, earnings by educational attainment: https://www.oecd.org/en/publications/education-at-a-glance-2025_1c0d9c79-en/full-report/what-are-the-earnings-advantages-to-education_7a7e64e0.html

[9] ILO, Generative AI and jobs: A 2025 update, 20 May 2025: https://www.ilo.org/publications/generative-ai-and-jobs-2025-update

[10] IMF, Gen-AI: Artificial Intelligence and the Future of Work, January 2024: https://www.elibrary.imf.org/view/journals/006/2024/001/article-A001-en.xml

[11] David Autor, Applying AI to Rebuild Middle Class Jobs, NBER 32140, 2024: https://www.nber.org/papers/w32140

[12] Daron Acemoglu, “The Simple Macroeconomics of AI”, Economic Policy 40(121), 2025: https://academic.oup.com/economicpolicy/article-abstract/40/121/13/7728473

[13] OECD, PISA 2025 — Brazil, 8 September 2026: https://www.oecd.org/en/publications/pisa-2025-results-volume-i-country-notes_2d4ff9ea-en/brazil_d65ba095-en.html

[14] Stanford HAI, AI Index Report 2026, Education and Economy chapters: https://hai.stanford.edu/ai-index/2026-ai-index-report

[15] OECD, PISA 2025 Results (Volume I), 8 September 2026: https://www.oecd.org/en/publications/pisa-2025-results-volume-i_73451bc5-en/full-report/student-performance-in-pisa-2025_23e075c2.html

[16] UNESCO, Guidance for Generative AI in Education and Research, 2023, updated 2026: https://www.unesco.org/en/articles/guidance-generative-ai-education-and-research

Intellectual bibliography: Plato, Phaedrus; Karl Marx, Capital, Book I; Jacques Ellul, La Technique ou l'enjeu du siècle (1954); Hannah Arendt, The Human Condition (1958); Ivan Illich, Tools for Conviviality (1973); Neil Postman, Technopoly (1992); Bernard Stiegler, Technics and Time and writings on the proletarianisation of knowledge; Theodore Dalrymple, Spoilt Rotten (2010); Richard Sennett, The Craftsman (2008); Matthew B. Crawford, Shop Class as Soulcraft (2009); Edward Deci and Richard Ryan, self-determination theory; Albert Bandura, research on self-efficacy; Mihaly Csikszentmihalyi, Flow; David Autor, task-based labour economics; Daron Acemoglu, automation economics. These conceptual works are not empirical measurements of AI effects in 2026.