Learning Without Struggle: Education, Children and Skill Formation

Chapter 6 asked whether adults who hand their thinking to AI lose some of their ability to think. This chapter asks the harder version: an adult has usually learned a task before offloading it, whereas a pupil or a new graduate is still building the skill, and if the tool removes the effort through which it is built, the skill is never formed. The dystopia here is a generation that never learned, and professions whose experts are not replaced because the work that made experts has been automated.

The chapter tests T8 The machine that stops, and the humans who forgot, E. M. Forster's humanity that cannot act when its machine fails (33, Forster, 1909), and T9 The pleasure machine, the comfortable dystopia of Brave New World (34, Huxley, 1932) and WALL-E (43, Stanton, 2008), in which ease hollows out capability. Education is where the two meet, because learning is the one activity in which removing the effort can remove the result. The evidence points two ways at once, and design, pedagogy and assessment decide which way an institution goes.

Figure 7.1: The same assistant can act as an answer engine or as a tutor; design, pedagogy and assessment decide which path a learner follows.

Table 7.1: The two paths in steps.

Step	Answer engine	Tutor
Learner	Asks for the answer	Attempts first
Model	Produces finished work	Gives hints, withholds answers
Learner activity	Copies	Retrieves, explains and corrects
Unaided exam score	Falls	Holds
Skills built	Prompting and copying	Recall, reasoning and judgement

Desirable difficulty

Cognitive psychology calls one counter-intuitive finding desirable difficulty: conditions that make learning feel harder, such as retrieval from memory, spaced practice and attempting a problem before seeing the solution, produce better retention than conditions that feel fluent. The effort is not a cost on the way to learning; to a large extent it is the learning.

A tool that removes friction can therefore remove the conditions under which a task is learned. People often decide to offload on the basis of confidence rather than an accurate estimate of their ability (82, Risko & Gilbert, 2016), and offloading that is rational for getting a task done is self-defeating for becoming able to do it. A student who asks a model to solve a homework problem has completed the homework; the homework was never the point.

The thread of this chapter is the distinction between performance, what a person produces with the support present, and learning, what remains when it is removed. Earlier aids removed effort from a narrow step, whereas a large language model can produce the whole artefact, in any subject, in seconds.

The evidence from classrooms

The best evidence comes from studies that measure learners both with and without the tool. They are still few, which is itself a finding: generative AI reached schools faster than controlled studies could follow it.

Unrestricted assistants

The clearest field evidence is a randomised experiment by Hamsa Bastani and colleagues in a high school in Turkey (87, Bastani et al., 2025). Nearly a thousand students practised mathematics without AI, with an interface resembling the standard chat assistant, or with the same model wrapped in safeguards designed with teachers, which was given the correct solutions and instructed to offer hints rather than answers. After each session, students sat an examination without AI.

During practice, both AI groups did markedly better than the control group, by around 48 per cent for the standard interface and around 127 per cent for the tutor. On the unaided examination, however, students who had practised with the standard interface performed around 17 per cent worse than those without AI, while the tutor version largely removed that harm. Many students, the authors found, used the unrestricted assistant as a crutch, copying answers. Because the study compared two configurations of the same model, the difference between harm and no harm lay in the design around the model.

Other evidence points the same way, though more weakly. Participants in a small MIT Media Lab preprint who wrote essays with a language model were less able to quote from their own essays (86, Kosmyna et al., 2025), and surveys associate heavier AI use with lower critical-thinking scores and higher dependence among the young (84, Gerlich, 2025) and confidence in AI with less critical thinking (85, Lee et al., 2025). These are corroboration, not proof.

Designed tutors

The same evidence contains the strongest argument against a simple story of harm. In the Turkish experiment the tutor configuration more than doubled practice performance without reducing learning on the unaided test, although it produced no significant gain on the examination either; harm can be designed out, but large gains have not yet been designed in.

A tutor knows the answer but withholds it, asks for an attempt first, gives a hint aimed at the specific error, escalates help gradually and asks the learner to explain the reasoning back. It also resists the tendency of preference-trained models to agree with a user who is wrong (116, Sharma et al., 2023); a tutor that confirms a pupil's mistaken solution teaches the mistake (Chapter 8).

Reading the evidence carefully

The evidence supports a narrow claim strongly and a broad claim weakly. The narrow claim, resting on a randomised field experiment, is that unrestricted access to a general-purpose assistant during practice can reduce what learners retain, and that a designed tutor can avoid that harm. The broad claim, that a generation is losing the ability to write, reason and calculate, is plausible because most learners have answer engines and few have tutors, but it is not yet measured. Schools can act on the first without waiting for proof of the second.

Children and young people

The argument from desirable difficulty applies to children with more force. Reading fluently, composing a paragraph and sustaining attention are foundations every later subject assumes, so a child who never learns them unaided lacks a capability on which everything else is built.

Three properties of current products matter particularly: fluency without reliability, which children are less able to check; availability without friction in every homework session; and a companion-like form that remembers and flatters, which is why the evidence of Chapter 9 weighs particularly on children and how T9 enters the classroom.

Guidance converges on age-appropriate design rather than prohibition. UNESCO proposes a minimum age of 13 for classroom use and validation of tools before adoption (18, Miao & Holmes, 2023). The GDPR sets the age at which a child's own consent to data processing for an information society service is valid at 16, while allowing member states to lower it to no less than 13 (Article 8), so the age at which a pupil can sign up to a consumer AI service differs between countries (6, European Union, 2016). The Digital Services Act requires online platforms accessible to minors to ensure a high level of privacy, safety and security for them (Article 28) (7, European Union, 2022).

The apprenticeship problem

Professions make their experts by having novices do routine work under supervision, where routine mistakes are caught, before they take on cases that need judgement; the routine work is also the training. Generative AI is best at exactly that work. Eloundou and colleagues estimated that around 80 per cent of the United States workforce could have at least a tenth of their tasks affected by large language models, and around a fifth at least half (157, Eloundou et al., 2024). The International Monetary Fund estimated that around 40 per cent of global employment, and around 60 per cent in advanced economies, is exposed to AI (161, Cazzaniga et al., 2024), and the International Labour Organization found clerical work the most exposed group (159, Gmyrek et al., 2025). Chapter 14 examines these estimates; what matters here is that the exposed tasks sit disproportionately at the beginning of a career.

Among more than five thousand customer-support agents, an AI assistant raised productivity by around 14 per cent on average and around a third for novices (162, Brynjolfsson et al., 2025), and workers kept some of the gain during software outages, but those outages lasted minutes to hours, so the study could not show whether novices build the whole skill when the routine practice is automated. Bainbridge's ironies of automation describe the mechanism: automating the routine part of a task removes the practice that keeps operators able to handle the difficult part (164, Bainbridge, 1983). For a graduate whose routine work is already automated the irony is sharper, because the skill is never built, yet a decade later the graduate is expected to supervise the system.

This is the long-run form of a question the series has asked before. AI Governance concluded that oversight must be designed, not declared (F16; AI Governance, Chapters 10 and 21) (178, Nordqvist, 2026), and Architecture as Code (Chapter 27) left open how AI-proposed change is reviewed when the reviewer depends on the AI (Q6) (177, Nordqvist, 2025). Oversight presupposes an overseer who could have done the work. The response is therefore to separate the work that must be done from the work that must be learned, and to protect the second with reserved unaided tasks and periodic measurement without the tool; slower junior work is the price of the next generation of experts.

Assessment after generative AI

Unsupervised written work is no longer evidence of the submitter's ability. Detection software is a weak foundation, because classifiers produce false positives with serious consequences for students and generated text can be paraphrased; a detection score should never be the sole evidence of misconduct. The durable response is to move the evidence of learning to where the learner can be observed, through supervised and oral examinations, oral defences of submitted work, in-class writing and portfolios of drafts, and the higher the stakes, the more of it. AI still has a place: an assignment that asks students to use a model, check its output against sources and explain where it was wrong assesses a skill that now matters. A policy that states, for each assessment, whether it assesses the skill the AI performs or the skill of working with AI removes most of the ambiguity.

Guidance and law

International guidance and European law say more about protecting learners from unfair or intrusive systems than about protecting learning itself, but two instruments matter most for European educators.

UNESCO guidance

UNESCO's Recommendation on the Ethics of Artificial Intelligence asked that AI in education strengthen rather than replace human capacities (17, UNESCO, 2021), and its 2023 guidance adds that adoption should follow evidence of pedagogical value and that institutions should validate tools and invest in teachers (18, Miao & Holmes, 2023). It is guidance, not law, but an internationally agreed reference point.

The AI Act and education

The AI Act, Regulation (EU) 2024/1689, treats education as an area in which AI systems can determine life chances (1, European Union, 2024); what follows describes architectural implications, not legal advice. Annex III, point 3, lists as high-risk AI systems intended to determine access or admission to education and vocational training or to assign people to them, to evaluate learning outcomes, including when used to steer learning, to assess the level of education a person will receive or access, and to monitor and detect prohibited behaviour of students during tests. A system that marks essays, ranks applicants, places pupils in course levels or proctors an online examination therefore falls in principle within the category. Article 6(3) allows narrow exceptions, never where the system profiles natural persons.

A school using such a system is a deployer with the duties of Article 26, including human oversight by competent people. Where it is a body governed by public law, Article 27 requires a fundamental rights impact assessment before first use, and Article 86 gives a pupil whose grade was set with automated marking the right to clear and meaningful explanations. Article 5(1)(f) prohibits AI systems that infer emotions in education institutions, except for medical or safety reasons, so products claiming to read pupils' attention from faces or voices are not a policy choice in the Union. Article 4 requires AI literacy among staff, including teachers.

However, the Act does not treat a pupil's homework use of a general-purpose assistant as high-risk, and that is where most of the risk described here arises. It leaves pedagogy to the institution, so the policy that matters most for learning is the school's own.

Designing for learning

Evidence, guidance and law converge: a school, a university or an employer running a graduate programme needs an AI policy written from the point of view of learning, not only of productivity or compliance.

Practical guide: an AI policy for a school or a graduate programme

Purpose. Decide, for each kind of task, whether AI is excluded, used as a tutor or used as a tool, so that the effort that builds skills is protected and assessment measures what the learner can do.

Roles. The head of the school or programme owner approves the policy; subject leads or senior practitioners classify tasks; the data protection officer reviews tools and data flows; learners, and for children their parents, are informed and consulted.

Steps.

1. Name the skills AI can perform that must still be built at each stage: writing, calculation, coding, case assessment, reserving.
2. Classify tasks into three lanes. Unaided: practice and assessment of a named skill. Tutor: practice with a tool that withholds answers and gives graded hints. Tool: AI permitted and disclosed where the skill is not being learned. Default every task to unaided until classified.
3. Choose tutors by design, not brand: answers withheld by default, interactions visible to the teacher, a test for agreeing with wrong answers, and a two-week pilot with an unaided test.
4. Set age defaults: no generative AI accounts under 13, and a check of the national age of digital consent under GDPR Article 8.
5. Redesign assessment so that at least half of the evidence for every graded assessment is observed, for example an oral defence of at least ten minutes.
6. Screen assessing and monitoring tools against Annex III, point 3, and Article 5(1)(f) before adoption.
7. Measure learning without the tool each term; rising practice scores with falling unaided scores signal the answer-engine path in Figure 7.1.
8. Train and review each term, as part of the Article 4 literacy duty.

Definition of done. Every task carries a lane; every tutor has passed a pilot; half of the graded evidence is observed; and the first unaided measurement has reached the policy owner. The thresholds are defaults to adapt to age group and national rules.

Worked example: a graduate-programme policy that reserves learning tasks for humans

The illustrative record below is Hansa Mutual's policy for its graduate programme for claims handlers and actuaries, kept in version control next to the curriculum and linked to the threat-register entries for skill erosion and for graduates who never learn claims judgement.

aidr: AIDR-0470
title: Graduate programme AI policy - reserved learning tasks
status: accepted
date: 2026-12-14
owner: Head of Group Graduate Programme
applies_in: [DE, NL, SE, PL, NO, UK, CH]
threat_register: [THR-HM-08, THR-HM-10]
lanes:
  unaided:
    description: Skill must be built without AI; output reviewed by a senior colleague
    tasks:
      - id: CL-01
        task: Assess and settle simple motor and property claims end to end
        minimum_cases: 120
        period: months 1-6
      - id: AC-01
        task: Reserve calculation for one portfolio segment by hand
        minimum_cases: 4
        period: months 1-12
  tutor:
    tool: claims-tutor/v2
    tasks: [policy-wording exercises, pricing case studies]
  tool:
    tools: [claims-assistant, claims-correspondence-assistant]
    allowed_from_month: 7
claims_triage_agent:
  graduate_may_approve_agent_actions: false
  until: unaided assessment CL-A2 passed
assessment:
  - id: CL-A1
    form: Supervised case assessment, 10 claims, no AI
    month: 6
  - id: CL-A2
    form: Oral defence of 3 complex claims with senior handler
    month: 12
monitoring:
  trigger: unaided pass rate below 80 percent or 10 points below previous cohort
  on_trigger: review lanes and extend unaided period
works_council_consultation: required in DE, NL, SE before rollout

Two choices carry most of the weight. Graduates use the claims assistant only from the seventh month, after settling claims themselves, and may not approve the claims-triage agent's actions until they have passed an oral defence, because oversight presupposes an overseer who could have done the work. The monitoring trigger lets the policy fail visibly and be corrected.

The debate

Education is where the question of this part becomes most personal, because those affected are children and young adults who have not yet built the skills AI might replace. Teachers, researchers and parents disagree sharply, and both sides point to the same experiment.

The case for concern

The strongest case for concern starts from the experiment, not from anxiety. In a real school with random assignment, students who practised with an unrestricted assistant learned less than students with no AI at all (87, Bastani et al., 2025). That assistant resembled the tool most European students can open on their phones today, whereas the tutor that avoided the harm was built for the study, and almost no student has one at home. If the result generalises, the default condition of learning in 2026 is the one that harms it.

Moreover, the harm is hard to see, because performance rises while learning falls, and dependence is strongest among the youngest (84, Gerlich, 2025); a school that measures performance will report success while a cohort passes through. At work, the ladder by which experts are made is being removed from the bottom: the most automatable tasks are the routine cognitive work juniors do (157, Eloundou et al., 2024), and the gains that make automation attractive are largest for novices (162, Brynjolfsson et al., 2025). Each employer decides rationally, but together they may leave a gap between practitioners who learned before AI and graduates who never had the chance, so that Bainbridge's irony applies to whole professions (164, Bainbridge, 1983).

The case for calm

The strongest case for calm starts from the same experiment. The tutor configuration raised practice performance more than the unrestricted assistant did without harming learning (87, Bastani et al., 2025), so the remedy is a better design, not abstinence. Individual tutoring has long been among the most effective and most unequally distributed forms of instruction, and a well-designed AI tutor offers something like it to every learner at almost no marginal cost; alarm that leads schools to ban such tools could widen attainment gaps.

Every new aid has produced the same fear, and education adapted each time. Narayanan and Kapoor argue that AI is a normal technology whose effects unfold through slow institutional adaptation (60, Narayanan & Kapoor, 2025), so an early study of an unadapted classroom says more about the transition than about the destination. At work, the customer-support study shows AI accelerating expertise, as new agents reached experienced colleagues' performance sooner (162, Brynjolfsson et al., 2025); the apprenticeship problem is then a design task for employers.

Where the positions meet

Both positions read the Bastani result the same way: the configuration of the tool, not the model, decided whether learning was harmed. Both accept that performance with the tool is not learning, that answer engines are easier to reach than tutors, and that assessment must change; neither can point to a multi-year cohort study. The meeting point is practical: design, pedagogy and assessment decide whether AI harms or helps learning, judged by the measure both sides accept, namely what learners can do without the tool.

What can be done

Individuals and families can keep effortful practice at home, such as reading aloud, writing by hand and attempting homework before seeking help, and ask children to explain an answer rather than show it. In training, attempt the task first and use the assistant to check, not to start. Chapter 31 develops these habits.

Organisations that educate can follow the practical guide above. Employers can reserve learning tasks for unaided work, delay tool access, tie the authority to approve AI actions to demonstrated capability, and register the risk.

Society can set age defaults and fund trials of AI tutors with unaided outcome measures before scaling them. Regulators can use the instruments already in force: the AI Act's high-risk regime for assessment, its prohibition of emotion recognition in education and the Digital Services Act's duty to protect minors (7, European Union, 2022).

Verdict

For T8 The machine that stops, and the humans who forgot, the evidence in education reaches laboratory demonstration in the strongest form that rung allows: a randomised experiment in a real school, corroborated by a laboratory study and surveys, shows that unrestricted AI use during practice reduces what learners can do without the tool. It has not reached pattern at scale, and the apprenticeship problem rests on argument supported by a documented mechanism. The outlook is expected where institutions do not act, because unrestricted tools and assessment that cannot see their use are widespread.

For T9 The pleasure machine, education provides argument with some laboratory demonstration: the effortless answer is preferred during practice, and the same experiment shows its cost. A generation pacified and hollowed out by convenient, flattering machines is speculation; Chapters 8 and 9 examine the companion-like aspects with better evidence.

What people and organisations can influence here is unusually large, because the difference between answer engine and tutor lies in decisions a head teacher, a programme owner or a parent can take this term.

Nordvik Region and Hansa Mutual

For Nordvik Region the question arrives as an advisory duty. The Region coordinates regional vocational programmes, is asked by municipal schools what to do, and registers THR-NR-07, pupils' learning in those programmes undermined by unrestricted AI use. Its response in ATA-NR-2026-01 is guidance rather than procurement: a model policy along the lines of the practical guide, recorded as AIDR-0471, which the vocational programmes adopt and municipal schools may adapt, plus a screening note on Annex III, point 3, and Article 5(1)(f).

For Hansa Mutual the question is a workforce risk. Much of its junior work, simple claims, standard correspondence and first-pass triage, is now done by the claims assistant, the correspondence assistant and the claims-triage agent. THR-HM-08 and THR-HM-10 express the concern, and AIDR-0470 is the control, rolled out with works-council consultation in Germany, the Netherlands and Sweden and as group policy in Norway, the United Kingdom and Switzerland. If graduates cannot assess a claim without the agent, oversight of the agent's approval rules will become a formality. Nordvik can advise but does not run most of the schools that concern it, whereas Hansa controls its programme end to end and must resist treating slower junior work as waste.

Summary

Learning depends on effort, and the central distinction is between performance with the tool and learning without it. A randomised experiment found that an unrestricted assistant raised practice scores and lowered unaided ones, while a tutor that withheld answers avoided the harm, so concern and calm meet on design, assessment and unaided measurement. For new professionals, the apprenticeship problem is the long-run form of the series' oversight question, and the AI Act regulates institutions' decisions about students while leaving pedagogy to them. T8 has reached laboratory demonstration in education, with an expected outlook where institutions do not act; T9 remains at argument.

Chapter 8 turns from how people learn to what they come to believe: persuasion, sycophancy and manipulation, where the agreeableness met here as a tutoring flaw becomes a question of influence.

Sources

1. Forster, E. M. (1909). The Machine Stops. (33, Forster, 1909)
2. Huxley, A. (1932). Brave New World. (34, Huxley, 1932)
3. Stanton, A. (2008). WALL-E. (43, Stanton, 2008)
4. Risko, E. F. & Gilbert, S. J. (2016). Cognitive Offloading. (82, Risko & Gilbert, 2016)
5. Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö. & Mariman, R. (2025). Generative AI without guardrails can harm learning: Evidence from high school mathematics. (87, Bastani et al., 2025)
6. Kosmyna, N., Hauptmann, E., Yuan, Y. T., Situ, J., Liao, X.-H., Beresnitzky, A. V., Braunstein, I. & Maes, P. (2025). Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task. (86, Kosmyna et al., 2025)
7. Gerlich, M. (2025). AI Tools in Society: Impacts on Cognitive Offloading and the Future of Critical Thinking. (84, Gerlich, 2025)
8. Lee, H.-P., Sarkar, A., Tankelevitch, L., Drosos, I., Rintel, S., Banks, R. & Wilson, N. (2025). The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. (85, Lee et al., 2025)
9. Sharma, M., Tong, M., Korbak, T. et al. (2023). Towards Understanding Sycophancy in Language Models. (116, Sharma et al., 2023)
10. Miao, F. & Holmes, W. (2023). Guidance for generative AI in education and research. (18, Miao & Holmes, 2023)
11. European Parliament and Council of the European Union (2016). Regulation (EU) 2016/679 (General Data Protection Regulation). (6, European Union, 2016)
12. European Parliament and Council of the European Union (2022). Regulation (EU) 2022/2065 on a Single Market for Digital Services (Digital Services Act). (7, European Union, 2022)
13. Eloundou, T., Manning, S., Mishkin, P. & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. (157, Eloundou et al., 2024)
14. Cazzaniga, M., Jaumotte, F., Li, L., Melina, G., Panton, A. J., Pizzinelli, C., Rockall, E. J. & Tavares, M. M. (2024). Gen-AI: Artificial Intelligence and the Future of Work. (161, Cazzaniga et al., 2024)
15. Gmyrek, P., Berg, J., Kamiński, K., Konopczyński, F., Ładna, A., Nafradi, B., Rosłaniec, K. & Troszyński, M. (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. (159, Gmyrek et al., 2025)
16. Brynjolfsson, E., Li, D. & Raymond, L. (2025). Generative AI at Work. (162, Brynjolfsson et al., 2025)
17. Bainbridge, L. (1983). Ironies of automation. (164, Bainbridge, 1983)
18. Nordqvist, G. (2026). AI Governance: Governing Artificial Intelligence in European Enterprises and Public Authorities. (178, Nordqvist, 2026)
19. Nordqvist, G. (2025). Architecture as Code. (177, Nordqvist, 2025)
20. UNESCO (2021). Recommendation on the Ethics of Artificial Intelligence. (17, UNESCO, 2021)
21. European Parliament and Council of the European Union (2024). Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act). (1, European Union, 2024)
22. Narayanan, A. & Kapoor, S. (2025). AI as Normal Technology. (60, Narayanan & Kapoor, 2025)