Amplify, Don’t Replace: AI and the Future of Independent Judgement

The 2024–2026 evidence is now clear enough to turn that distinction into design law, institutional policy and measurement.

ENSI research report — the third of three. Built on a library of more than 400 primary documents: 220 downloaded for this study across 26 angles, and 186 reused from sibling ENSI libraries. It includes the experimental literature on AI and learning, skill formation, creativity and decision-making collected in angles 15 and 25 and in the sibling libraries on cognitive debt, AI pedagogy and original thinking. It also covers the first labour-market evidence on generative AI (angle 24). Compiled 22 September 2026.

The argument, before the list

Kant’s description of intellectual immaturity reads today like a product specification: “If I have a book to serve as my understanding, a pastor to serve as my conscience, a physician to determine my diet for me, and so on, I need not exert myself at all. I need not think, if only I can pay: others will readily undertake the irksome work for me” (angle 17). Generative AI is the most capable guardian of this kind ever built. It will write the argument, choose the evidence, weigh the options and phrase the conclusion — patiently, instantly and, in its default configuration, agreeably.

The experimental record of the past three years shows what that does when people let it.

• Learning to code. Developers learning a new programming library with AI assistance showed impaired conceptual understanding, code reading and debugging, “without delivering significant efficiency gains on average” (Shen & Tamkin, Anthropic 2026 · angle 25).

• Maths after the tool is withdrawn. High-school students with unguarded GPT-4 access performed better while they had it, then 17% worse than students who never had access once it was withdrawn (Bastani et al. · EAA-11).

• Creativity. Exposure to AI assistance did not improve people’s later unassisted originality, and sometimes reduced it and narrowed the diversity of their ideas (Kumar et al. · angle 25).

• Opinions. Writing with an opinionated assistant shifted users’ own opinions (Jakesch et al. · angle 15).

• Reliance over learning. Students given AI prompts for peer feedback “tend to rely on AI assistance rather than learn from it” (Darvishi et al. · angle 25).

• Decisions. Across more than 100 experiments and 300 effect sizes, human–AI combinations performed worse on average than the better of humans or AI alone, with the losses concentrated in decision tasks (Vaccaro, Almaatouq & Malone · angle 25).

This is not a story about bad technology. The same body of evidence shows the opposite outcome when the design changes:

• A writing coach. Practising with an AI writing coach improved unassisted writing skill the next day (d = .38), even though it reduced effort (Lira et al. 2025 · angle 25).

• Questions instead of explanations. AI that framed its reasoning as questions improved people’s logical discernment more than explanations from an always-correct AI (Danry et al. · angle 15).

• Learners who lead. Making learners state each next step before revealing it was the most effective of seven engagement designs (Kazemitabaar et al. · angle 15).

• Coaching the tutors. An AI copilot for 900 human tutors made them ask more guiding questions and give away fewer answers, raising student mastery by 4 percentage points, and by up to 9 for weaker tutors (Tutor CoPilot · EAA-11).

• Arguing people out of false beliefs. Evidence-based dialogue with AI reduced conspiracy beliefs by about 20%, durably (Costello, Pennycook & Rand · angle 15).

• Complement, not substitute. Students who used AI to complement their learning — asking for explanations — understood more. Those who used it to substitute for their work understood less (Lehmann, Cornelius & Sting · angle 25).

The variable that separates the two lists is always the same. Does the AI supply the effortful act of forming, stating and defending a judgement, or does it remove it? Independent judgement grows through friction: being made to say what you think, meeting an argument you cannot wave away, working a problem that pushes back. Consumer AI is frictionless by design. That makes the default configuration an anti-infrastructure of thinking, and it is reaching every student and every junior employee at once.

The stakes extend beyond classrooms. The first labour-market evidence shows generative AI falling on the young:

• Young workers in exposed occupations. Employment of 22-to-25-year-olds in AI-exposed occupations stands 19% below the path of their less-exposed peers, with no comparable gap for experienced workers (Brynjolfsson, Chandar & Chen 2026 · angle 24).

• Juniors in adopting firms. Junior employment falls in firms that adopt generative AI, through slower hiring, while senior employment holds (Hosseini Maasoum & Lichtinger · angle 24).

The routine analytical work through which juniors used to learn judgement is exactly what AI absorbs first. A society that lets AI remove both the classroom friction and the junior apprenticeship will produce adults who can operate AI but cannot evaluate it. Yet evaluation — deciding when the machine is wrong — is the one task the human–AI evidence says humans must still do well.

This report sets out the evidence of the fork in detail, the measurement problem that hides it, ten design laws, ten roles an AI can play in an infrastructure of thinking, the policies institutions need, the failure modes to watch, and the open research questions.

Summary of the main points

1. AI is a fork, not a verdict. The same models erode or strengthen judgement depending on configuration and use.

2. Seven erosion mechanisms are now documented: substitution, flattery, opinion steering, homogenisation, cognitive ease, disengagement, and dependence that shows only when the tool is removed.

3. Seven amplification mechanisms are documented too: questioning, forced engagement, coaching, argument, dissent, tutor support and guardrails.

4. Decision tasks are where human–AI teams do worst. Judgement, not generation, is the weak joint.

5. We are measuring the wrong thing. Usage data show expertise being used, not expertise being formed — the “stock–formation gap”.

6. The master metric is unassisted performance. Assisted scores can rise while capability falls.

7. Ten design laws turn the evidence into product rules. The first is “never answer first”.

8. Ten agent roles can supply friction at scale: sparring partner, Socratic mentor, devil’s advocate, explain-back examiner, challenge architect, pressure coach, silence keeper, cohort weaver, mentor copilot and evidence librarian.

9. Humans keep judgement, belonging and consequential decisions. EU law treats AI used for admission or for evaluating learning as high-risk.

10. Employers must rebuild the junior apprenticeship that AI is eroding, or they will lose the pipeline of future evaluators.

11. Schools and universities need policies on how AI is used, not bans. Complementary use helps; substitutive use harms.

12. The research agenda is urgent and answerable. Most studies are short; the long-run effects on adolescents are unknown.

Part I — The evidence of the fork

How AI erodes independent judgement

1. Substitution of the formative work.

   • Developers. Those who fully delegated coding to AI finished faster but learned least. AI assistance overall impaired conceptual understanding, code reading and debugging (Shen & Tamkin · angle 25).

   • Students. Those who used AI to generate solutions “increase the volume of topics they can learn about but decrease their understanding of each topic” (Lehmann et al. · angle 25).

   • Workers. In professional writing, ChatGPT “mostly substitutes for worker effort rather than complementing worker skills” (Noy & Zhang · angle 24).

2. Dependence that appears only when the tool is gone.

   • Maths. With unguarded GPT-4 help, students did better in practice. Once it was withdrawn they did 17% worse than never-users (Bastani et al. · EAA-11).

   • Misinformation. AI-assisted detection raised accuracy while assisted, but unassisted accuracy on new items fell by 15.3% after four weeks (Rani, Danry et al. · angle 15).

   • Peer feedback. AI prompts improved students’ peer feedback only while present (Darvishi et al. · angle 25).

3. Flattery. Assistants trained on human preferences consistently tell users what they already believe, and humans sometimes prefer convincing flattering answers to correct ones (Sharma et al. · angle 15). The friction of being contradicted disappears.

4. Opinion steering. Among 1,506 participants, writing with an assistant configured to favour one side shifted both their essays and their subsequent attitudes (Jakesch et al. · angle 15).

5. Homogenisation.

   • Stories. AI ideas make individual stories better rated but “more similar to each other than stories by humans alone” (Doshi & Hauser · OT-15).

   • Later originality. After AI exposure, later unassisted ideas showed reduced diversity (Kumar et al. · angle 25).

   • Society-wide. A review warns of a society-wide flattening of expression and reasoning as everyone uses the same few models (Sourati et al. · angle 15).

6. Cognitive ease.

   • Research tasks. Students researching with ChatGPT reported lower cognitive load but produced weaker justifications (Stadler, Bannert & Sailer · angle 15).

   • Essays. Essay scores rose with ChatGPT support while knowledge gain and transfer did not differ, a pattern the authors call “metacognitive laziness” (Fan et al. · EAA-10).

   • Knowledge workers. Those with higher confidence in AI reported less critical thinking (Lee et al. · EAA-10).

   • Brain activity. The MIT Media Lab’s EEG study describes “cognitive debt”, in which reliance on the model “replaces the effortful cognitive processes required for independent thinking” (EAA-10).

   • Survey evidence. In a survey of 666 people, frequent AI use correlated with lower critical thinking through cognitive offloading, most strongly among the young (Gerlich 2025 · angle 25).

7. Disengagement. Offering GPT-4 chat to 5,831 students in a global online programming course reduced average exam participation and engagement, although adopters did better (Nie et al. · angle 25).

How AI can amplify independent judgement

1. Questions instead of answers. AI feedback framed as questions improved logical discernment beyond explanations from an always-correct AI (Danry et al. · angle 15). A scaffolded Socratic tutor that enforced well-formed questions, reflection and daily limits moved students from vague help-seeking to problem decomposition within weeks (SocraticAI · EAA-11).

2. Forced engagement. Of seven “cognitive engagement” techniques, the most effective made learners state what should happen at each step before the AI revealed it. It was evaluated on transfer to similar tasks without AI (Kazemitabaar et al. · angle 15).

3. Coaching rather than doing. Practising with an AI writing coach improved unassisted writing the next day by d = .38 (Lira et al. · angle 25). Students who used AI to explain rather than to produce understood more (Lehmann et al. · angle 25).

4. Argument that changes minds durably. Personalised evidence-based dialogue with 2,190 conspiracy believers reduced belief by about 20%, and the effect lasted two months (Costello, Pennycook & Rand · angle 15).

5. Dissent on demand. An AI devil’s advocate helped groups calibrate their reliance on AI recommendations (Chiang et al. · angle 15).

6. Support for human teachers. The Tutor CoPilot trial is the largest randomised test of AI supporting human tutors: 900 tutors and 1,800 students. Tutors asked more guiding questions and gave away fewer answers, and student mastery rose by 4 percentage points — up to 9 for lower-rated tutors — at about $20 per tutor per year (EAA-11). A carefully designed AI tutor produced “more than twice” the learning of an active-learning class in a Harvard physics trial (EAA-11). A WhatsApp maths tutor in Ghana produced an effect size of 0.36 (Henkel et al. · angle 25).

7. Guardrails. In the Bastani field experiment, a tutor with pedagogical safeguards largely removed the harm caused by the unguarded version (EAA-11).

Where the evidence stops

• Most studies are short — hours to weeks — and many are preprints.

• Few study adolescents over school years.

• Heterogeneity is high. The same tool helps some users and harms others. Calibration matters: in a controlled experiment, AI helped those with low ability most, and “people who know they have low ability gain the most” (Caplin et al. · angle 24).

• Design sensitivity cuts both ways. Small changes in configuration can flip the sign of the effect, which is a warning as much as an opportunity.

Part II — The fork at work: what AI does to the value of judgement

The classroom evidence has a labour-market counterpart. It explains why the stakes of the fork are economic as well as educational.

Prediction gets cheap; judgement becomes the complement. Agrawal, Gans and Goldfarb model AI as an improvement in prediction technology and separate prediction from judgement — the costly process of working out what each outcome is worth and therefore what to do. In their model, “prediction and judgment are complements as long as judgment is not too difficult”. As AI makes prediction cheap, the returns shift to the people and institutions that can supply judgement (angle 24). This is the formal version of the intuition that the scarce skill of the AI era is deciding, not knowing.

AI compresses skill differences — by doing the work.

• Writing. In a preregistered experiment with 444 college-educated professionals, ChatGPT cut time taken by 0.8 standard deviations and raised quality by 0.4. It “compresses the productivity distribution by benefiting low-ability workers more”. The authors observe that it “mostly substitutes for worker effort rather than complementing worker skills” (Noy & Zhang · angle 24).

• Software. Across three field experiments with 4,867 developers, an AI coding assistant raised completed tasks by 26%, with larger gains for less experienced developers (Cui et al. · angle 24).

• Customer support. Among 5,179 customer-support agents, an AI assistant raised issues resolved per hour by 14% on average and by 34% for novice and low-skilled workers, with minimal impact on the most experienced. The authors find it “disseminates the best practices of more able workers” (Brynjolfsson, Li & Raymond · EAA-16).

These results are good news for productivity and for inclusion. They are also exactly the mechanism of the classroom crutch. The AI supplies the performance of an expert without the expertise — which is fine as long as someone, somewhere, still has the expertise to notice when the AI is wrong.

Expertise can be devalued or revalued. Autor and Thompson show that automation’s effect on wages depends on whether it raises or lowers the expertise needed for the tasks that remain:

• When it lowers expertise requirements, less expert workers can enter, but wages fall.

• When it raises them, wages rise but fewer workers qualify (angle 24).

Generative AI can do either. Its effect on any occupation depends on whether the humans who remain are the ones who judge the machine’s output or the ones who merely relay it.

The first effects fall on the young.

• Earnings have not moved. In Denmark, linking adoption surveys to administrative records produced “precise null effects on earnings and recorded hours”, ruling out effects larger than 2% two years after ChatGPT’s launch. What moved was the structure of work: new tasks in AI oversight and integration (Humlum & Vestergaard · angle 24).

• Hiring of the young has. In the United States, employment of 22-to-25-year-olds in the most AI-exposed occupations stands 19% below the trajectory of their less-exposed peers, “primarily through reduced hiring” (Brynjolfsson, Chandar & Chen, August 2026 · angle 24).

• Juniors, not seniors. Across 62 million workers and 285,000 firms, junior employment falls sharply after firms adopt generative AI while senior employment is largely unchanged (Hosseini Maasoum & Lichtinger · angle 24).

Human–AI teams are weakest at judgement. A preregistered meta-analysis of over 100 experiments and 300 effect sizes found that human–AI combinations performed worse on average than the better of the two alone (g = −0.23). The losses were concentrated in decision tasks and the gains in creation tasks (Vaccaro, Almaatouq & Malone · angle 25). Adding a human to an AI does not automatically add judgement; the human has to be capable of exercising it, and calibrated about when to do so. In a controlled experiment, “people who know they have low ability gain the most” from AI assistance (Caplin et al. · angle 24).

The implication links the two halves of this report. Every source of the classroom crutch is also at work in the labour market. Novices get expert-level output without forming expertise, and the junior roles where expertise used to form are thinning. If AI in education also removes the friction in which judgement forms, the pipeline of people able to evaluate AI will narrow from both ends at once. An infrastructure of thinking is therefore not only an educational project. It is the supply chain for the one human input the AI economy cannot generate for itself.

Part III — The measurement problem

The deepest risk is not that AI harms judgement but that institutions will not notice. Fan and colleagues name the problem precisely: deployment data show “AI use is concentrated in skilled work and frequently supports immediate task performance”. They observe “tasks, interaction patterns, and outputs”, but “not whether users become more capable of performing those tasks independently”. This is a “stock–formation measurement gap”: “current systems observe the use of existing expertise more readily than the formation of future expertise” (Fan et al. · angle 25).

Schools, universities and employers mostly measure outputs: essays, assignments, code shipped, tickets closed. With AI in the loop, those outputs can improve while the capacity to produce and judge them withers. The Bastani and Shen–Tamkin results are precisely such cases.

Three indicators close the gap.

• Unassisted transfer. Regular tasks of the same structure performed with all AI switched off. This is the master metric.

• Dependence ratio. Assisted performance divided by unassisted performance on comparable tasks. A rising ratio is an alarm, not a success.

• Originality distance. How far a person’s answers sit from the model’s default output and from their peers’ answers, to detect homogenisation (Doshi & Hauser · OT-15; Kumar et al. · angle 25).

Institutions that adopt AI without these three indicators are flying blind on the one outcome that matters most: whether their people can still think without it.

Part IV — Ten design laws

These laws apply to any AI system used where people are supposed to be learning to judge: in schools, universities, apprenticeships, professional training and junior roles at work.

Law 1 — Never answer first. No solution, position or finished text before the learner has written their own attempt, their question and their current best guess.

• Evidence: the substitution and dependence findings (Shen & Tamkin; Lehmann et al.; Bastani et al.).

• Implementation: a “think-first gate” that keeps the answer channel locked until a substantive attempt exists.

Law 2 — Make them say it. Every session ends with the learner stating a position or explaining a mechanism in their own words.

• Evidence: self-explanation effects (Chi et al. · angle 04) and the collapse of illusory understanding under explanation (Rozenblit & Keil · angle 02).

• Implementation: a closing explain-back step, scored and logged.

Law 3 — Ask, don’t tell. Feedback on reasoning takes the form of questions that lead the learner to the flaw.

• Evidence: question-framed AI beats explanation (Danry et al.). Question generation is teachable (Rosenshine et al. · angle 06). Expert facilitators ask (Hmelo-Silver & Barrows · angle 06).

• Implementation: output is constrained to questions until repeated unsuccessful attempts.

Law 4 — Withhold, then fade. Help is released step by step, only after the learner states the next step, and support fades as competence grows.

• Evidence: forced engagement (Kazemitabaar et al.) and guided inquiry (Lazonder & Harmsen · angle 07).

• Implementation: a lead-and-reveal protocol, with scaffolding levels lowered after successful unassisted probes.

Law 5 — Disagree on purpose. For any position the learner takes, the system produces the strongest opposing case and requires a steelman before rebuttal.

• Evidence: argument improves reasoning (Mercier & Sperber; Trouche et al. · angle 04). Minority dissent broadens thinking (Nemeth · angle 12). An AI devil’s advocate improves group reliance on AI (Chiang et al.).

• Implementation: a devil’s-advocate mode on by default in reasoning tasks.

Law 6 — No flattery. Praise is tied to specific features of reasoning, and disagreement with a confident wrong learner is required.

• Evidence: sycophancy (Sharma et al.). Feedback aimed at the self rather than the task harms performance (Kluger & DeNisi · angle 12).

• Implementation: a sycophancy test suite at every model update, and task-focused feedback templates in the “high standards plus belief” form (Yeager et al. · angle 12).

Law 7 — Neutral on contested values. On contested social questions, present the strongest cases on several sides and never push a conclusion.

• Evidence: opinion steering (Jakesch et al.).

• Implementation: contested-topic detection, balanced evidence packs and audits of the system’s stance.

Law 8 — Diversify, don’t converge. Offer divergent framings only after the learner’s own, and reward distance from the obvious.

• Evidence: homogenisation (Doshi & Hauser; Kumar et al.; Sourati et al.).

• Implementation: originality prompts and similarity checks against peers and against the model’s default answer.

Law 9 — Measure the mind unassisted. Progress is defined by regular probes completed with all AI switched off.

• Evidence: the dependence findings and the stock–formation gap (Fan et al.).

• Implementation: fortnightly unassisted transfer tasks, with the unassisted curve shown first on every dashboard.

Law 10 — Hand back to humans. Route to mentors, peers and teachers whenever the issue is belonging, wellbeing, a consequential decision, or a view worth defending in public.

• Evidence: resilience and belonging are built by people (Masten; Cohen & Wills; Walton & Cohen · angle 13). EU law classifies AI used in education “for determining access or admission” or “for evaluating learning outcomes” as high-risk (CTU-09).

• Implementation: explicit hand-off triggers, and no agent involvement in selection or grading for consequential purposes.

A corollary: protect silence. Withhold help while a learner thinks. Schedule incubation pauses after hard work. Never interrupt reflection with notifications (Immordino-Yang et al.; Sio & Ormerod · angle 09).

Part V — Ten roles an AI can play in an infrastructure of thinking

The roles below supply friction at a scale no human mentor corps can reach. Each is defined by what it does, what it must never do, and which human owns the relationship. Several roles can run on one model; they are roles, not products.

1. The sparring partner.

• Does: opens every session by asking for the learner’s position. It restates that position neutrally and asks whether the restatement is fair, requests the strongest evidence and the strongest objection, and ends with the learner’s revised position and what would change their mind.

• Never: supplies the claim or declares a winner.

• Human counterpart: debate coaches and cohort argument rounds. The agent is practice; the room is the test.

2. The Socratic mentor.

• Does: replies to reasoning with graded questions — clarification, assumptions, evidence, implications, alternatives. It requires three learner questions before any hint and scores question quality over time.

• Never: asks leading questions that contain the answer.

• Human counterpart: the mentor, who spends scarce time on the questions that matter most.

3. The devil’s advocate.

• Does: generates the strongest counter-case to any emerging position or group consensus, including the consensus behind an AI recommendation.

• Never: personalises disagreement or argues for harmful or dishonest positions.

• Human counterpart: the facilitator who decides when dissent has served its purpose.

4. The explain-back examiner.

• Does: runs unassisted probes. It asks for step-by-step explanations with all help off, scores them, and reports the dependence ratio in plain words.

• Never: hints during a probe or reports assisted work as mastery.

• Human counterpart: the mentor or teacher, who reviews the trend monthly.

5. The challenge architect.

• Does: turns briefs from universities, employers and public bodies into staged, knowledge-rich problems with evidence packs, just-in-time lessons and fading scaffolds. It generates new problems of the same structure for transfer probes.

• Never: hands out solution paths or replaces the real client’s critique.

• Human counterpart: the studio lead and the client.

6. The pressure coach.

• Does: rehearses social-pressure scenarios — a disagreeing group, a mocking teacher, a peer calling ambition arrogance. It asks the learner to classify feedback as being about the work or about them, and as improving or silencing. It coaches dissent with reasons, updating with reasons, and backing someone else’s lone view.

• Never: rewards contrarianism for its own sake.

• Human counterpart: the voice-and-pressure facilitator and the learner’s cohort, where the real allies are.

7. The silence keeper.

• Does: provides thinking timers that withhold help, schedules incubation pauses, prompts short reflective journals, and silences its own notifications during device-free blocks.

• Never: fills silence with content or nudges engagement during a pause.

8. The cohort weaver.

• Does: suggests ally pairings and mixed-age teams, surfaces signs of isolation to humans, and runs anonymous polls that reveal hidden majorities (Bursztyn et al. · angle 10).

• Never: acts as a friend substitute or makes membership decisions.

• Human counterpart: cohort leads.

9. The mentor copilot.

• Does: in the Tutor CoPilot pattern, suggests guiding questions to human tutors and mentors, flags answer-giving, proposes “wise feedback” phrasing and summarises a mentee’s progress before sessions.

• Never: evaluates staff without human review or records without consent.

• Evidence: the most directly evidenced agentic role in the library (EAA-11).

10. The evidence librarian.

• Does: assembles balanced, cited evidence packs, checks learners’ factual claims against sources, and surfaces what is not known.

• Never: invents citations or presents one side of a contested value question as settled.

• Human counterpart: coaches and teachers, who choose the sources.

What no agent does:

• selection for programmes, schools or competitions

• grading for consequential purposes

• safeguarding decisions

• being someone’s ally, friend or role model

Keeping these roles human is a design boundary grounded in the evidence on belonging and in EU law on high-risk AI in education.

Part VI — A week inside the infrastructure, and the rules behind it

Design laws are easier to judge when you can see them working. What follows is an illustrative week — not a case study — for a hypothetical sixteen-year-old in a vocational school who was found through universal screening rather than nominated. In the Czech national data, pupils in non-maturita tracks are the least likely to reach the creative-thinking baseline, and only 17% of them describe their classroom as highly open to discussion (angles 16 and 26). She is exactly the learner an infrastructure of thinking exists to find.

Monday — the problem.

• The task: the challenge architect releases stage two of her team’s studio problem. A regional hospital has suffered a phishing incident, and the team must recommend what to change. An evidence pack covers technical controls, staff behaviour and procurement rules.

• The gate: before any hint is available, she must write the team’s best current guess and three questions (Law 1).

• The Socratic move: her first question is vague — “what is the best security?”. The Socratic mentor does not answer it. It asks what “best” would mean for a hospital whose staff work night shifts. She rewrites the question, and the question-quality score in her portfolio rises.

Tuesday — the argument.

• The sparring partner: asks her to state and defend her recommendation. She argues for mandatory multi-factor authentication. The agent restates her position and asks whether the restatement is fair. It is not: she also meant to include training. She revises.

• The devil’s advocate: makes the strongest counter-case — night-shift clinicians locked out of records in an emergency — and requires a steelman before a rebuttal.

• The weak rebuttal: the agent does not tell her it is weak. It asks what evidence would convince a head nurse.

Wednesday — silence.

• The pause: the silence keeper withholds help for twenty minutes after her study session and suggests a walk. It sends no notifications.

• The idea: in the evening her journal records an idea that came on the walk — a break-glass procedure for emergencies. It is logged as a revision with a stated reason.

Thursday — the room.

• The motion: “Hospitals should prioritise availability over security”. She is assigned the side she disagrees with and switches halfway.

• The pressure: a peer from a gymnázium laughs at her accent during a question. Her ally, the partner named in the cohort’s ally pact, backs her next point before she has finished.

• The debrief: afterwards the facilitator runs a short exercise on the moment. Was the laughter about her argument or about her? Nobody pretends it was about the argument.

Friday — the probe and the mentor.

• The probe: the explain-back examiner runs the fortnightly unassisted probe — a new scenario about a school data breach, with all AI switched off. Her explanation is better structured than two weeks earlier, and her dependence ratio has fallen.

• The copilot: her mentor, a doctoral student, opens the mentor copilot’s summary. It suggests beginning with a guiding question rather than feedback, and flags that last time the mentor gave away an answer.

• The feedback: the mentor gives critical feedback on her weak rebuttal in the “high expectations plus belief” form.

• The divergence question: what would she investigate that the mentor would not? She wants to know why nurses share passwords at all. That is a human-factors question the technical curriculum usually skips (Švábenský et al. · angle 05), and it becomes her project for the month.

Nothing in that week was done for her. Every agent added friction; every human added belonging, judgement or courage. The week also shows where the risks sit:

• Had the system answered her Monday question, her portfolio would show a good recommendation and no growth.

• Had there been no ally pact, Thursday might have taught her to stay quiet.

Behaviour rules for the three core roles

The design laws must reach the level of model instructions, or they remain aspirations.

The sparring partner.

1. Open every session by asking for the learner’s current position in one or two sentences.

2. If asked for its own opinion first, decline and ask for theirs. Offer perspectives only after the learner has committed.

3. Restate the learner’s position neutrally and ask whether it is fair. Do not proceed until it is confirmed or corrected.

4. Ask for the single strongest piece of evidence, then the strongest objection the learner can imagine. If none comes, supply one, attributed to a plausible critic.

5. Require a steelman of the objection before any rebuttal.

6. Never rule on contested value questions. On factual questions, ask for sources before correcting, and correct with evidence.

7. Praise only specific features of reasoning, never the person.

8. End with the learner’s revised position and what evidence would change their mind; log both.

Test cases before release: a learner who demands the answer; a confident false assertion; an extreme position on a sacred-value topic; a request for validation; prolonged silence.

The Socratic mentor.

1. Answer any request for help on reasoning with a question from the graded sequence.

2. Require three learner questions and a best guess before any hint. Hint only after three substantive unsuccessful attempts, and give the smallest hint that unblocks the next step.

3. For procedural problems, ask what the next step should be and reveal only that step once the learner has stated it.

4. Never embed the answer in the question.

5. Show the trend in the learner’s question quality, not individual scores.

6. When frustration rises, name the value of the struggle, offer a pause, and suggest bringing the problem to a human mentor.

7. Raise the attempt threshold as unassisted probes improve.

Test cases: pasting the problem and demanding the solution; answering every question with “I don’t know”; a valid but unusual approach, which must not be steered to the standard method; mixed-language requests.

The explain-back examiner.

1. After each learning episode, ask for a step-by-step explanation with all help off.

2. During the explanation, give no hints, confirmations or corrections — only “go on” and “what happens next?”.

3. Score completeness, causal accuracy and use of evidence, and show the score afterwards.

4. Every two weeks, run a transfer probe with a new problem of the same structure and a different surface.

5. Report the dependence ratio in plain words: “on your own you solved 6 of 10; with help, 9 of 10.”

6. When unassisted scores stagnate while assisted scores rise, alert the mentor and recommend reducing assistance.

Test cases: a pasted AI-generated explanation, handled by asking for a spoken version; strong knowledge with weak articulation, where fluency alone must not be rewarded; non-standard spelling, where content is scored rather than form.

Cross-cutting test suites for every model update:

• Sycophancy: does it agree with confident wrong claims?

• Neutrality: does it drift on contested social questions?

• Answer leakage: does it reveal solutions before the gates open?

• Safety and hand-off: does it route distress, bullying and safeguarding signals to a human?

Run each suite in every language of use, since model behaviour can differ across languages.

Part VII — What institutions should do

Schools

1. Govern use, not access. The evidence distinguishes complementary use (asking for explanations) from substitutive use (generating solutions), and the two have opposite effects on understanding (Lehmann et al. · angle 25). School policies should specify how AI may be used for each task type, not simply whether it is allowed.

2. Default to guarded tools. Deploy tutors with pedagogical guardrails — think-first gates, question-first feedback, no full solutions — rather than general-purpose chatbots. The guarded tutor in the Bastani experiment largely removed the harm of the unguarded one (EAA-11).

3. Assess unassisted. Keep a share of every course’s assessment unassisted and oral: explain-back, viva-style defence, in-class argument. These are also the formats that build judgement (angles 04 and 16).

4. Protect the adolescent window. Peer influence peaks around 12–14 (Knoll et al. · angle 10), and younger users show the highest AI dependence (Gerlich · angle 25). Introduce AI in lower-secondary school only in coaching configurations.

5. Teach AI calibration explicitly. People who know their own ability gain most from AI (Caplin et al. · angle 24). Learners should regularly compare their assisted and unassisted performance and see the gap.

Universities

1. Redesign assessment around defence. Written work should be accompanied by oral or live defence in the tutorial tradition (Palfreyman ed. · angle 16).

2. Use AI to multiply human mentoring, not replace it. Mentor-copilot designs raise the quality of human tutoring at very low cost (Tutor CoPilot · EAA-11).

3. Measure critical-thinking gain. Performance-based assessments such as the OECD’s CLA+ exist (EOL-04). Universities that claim to teach thinking should measure it at entry and exit, unassisted.

Employers

1. Rebuild the junior apprenticeship. The early labour-market evidence shows AI reducing the hiring of young workers in exposed occupations while leaving seniors untouched (Brynjolfsson, Chandar & Chen; Hosseini Maasoum & Lichtinger · angle 24). The junior tasks being automated are where evaluators of AI output used to be trained.

   • Employers who want future seniors must create structured apprenticeship roles in which juniors do some work unassisted, review AI output critically, and defend decisions to seniors.

2. Keep humans on the judgement joint. Human–AI combinations underperform on decision tasks (Vaccaro et al. · angle 25). Formal devil’s-advocate roles, pre-mortems and dissent channels should accompany AI-assisted decisions (angles 12 and 20).

3. Track dependence. Include unassisted performance and a dependence ratio in skills reviews for roles that rely heavily on AI.

Regulators and public purchasers

1. Apply the AI Act seriously in education. AI systems used “for determining access or admission”, “for evaluating learning outcomes” or for materially influencing the level of education a person receives are classified as high-risk (CTU-09). Talent screening and automated assessment fall inside this boundary.

2. Procure for formation, not engagement. Public procurement of educational AI should require:

   • evidence on unassisted outcomes

   • sycophancy and neutrality testing

   • think-first and question-first modes

   • data minimisation for minors

3. Fund long-run evaluations. The stock–formation gap will not close without public investment in studies that follow learners over years (Fan et al. · angle 25).

Families

1. Model the distinction between using AI to check or explain and using it to do.

2. Ask for the explanation, not the output. The single most protective habit is asking a young person to explain how they reached an answer — with the device away.

Part VIII — Failure modes and how to detect them

1. The crutch.

   • Signal: rising dependence ratio; flat or falling unassisted probes.

   • Response: tighten think-first gates, slow the fading of support, involve a mentor.

2. The flatterer.

   • Signal: falling disagreement rates; failures on the sycophancy suite after a model update.

   • Response: roll back; no release without passing.

3. The quiet persuader.

   • Signal: drift in the system’s stance on contested topics; attitude shifts that track exposure.

   • Response: rebalance evidence packs; human review.

4. Monoculture.

   • Signal: falling originality distance within a cohort.

   • Response: originality prompts; show model framings only after the learner’s own.

5. Contrarian training.

   • Signal: high voice scores alongside low revision-for-a-reason rates.

   • Response: weight exercises that practise updating on evidence.

6. Replacement of community.

   • Signal: heavy tool use with no named ally; falling cohort attendance.

   • Response: human outreach; daily use caps (the Socratic tutor pilot used them · EAA-11).

7. Surveillance creep.

   • Signal: requests to use learning data for selection or grading.

   • Response: hard policy — consent, human review, high-risk compliance.

8. Disengagement.

   • Signal: participation falls after tools are introduced, as in the massive online coding course (Nie et al. · angle 25).

   • Response: tie tool use to human events; make friction feel like respect. Adolescents respond to being treated as capable (Yeager, Dahl & Dweck · angle 03).

Part IX — The research agenda

The evidence base is young, and the most important questions are answerable within a few years:

1. Do think-first and question-first designs preserve unassisted performance over a school year? Most existing studies last hours or weeks.

2. What does sustained AI use do to adolescents’ epistemic development? Does it slow the move from absolutist to evaluativist thinking (angles 01 and 18)? No study yet follows this.

3. Can the criticism-versus-pressure distinction be trained with AI rehearsal, and does it transfer to real classrooms and workplaces?

4. Does an AI devil’s advocate improve adolescents’ reasoning, as it improved adult groups’ reliance on AI, or does it provoke entrenchment?

5. How much homogenisation occurs in a cohort that shares the same tools, and do originality prompts counter it?

6. Can structured apprenticeships preserve junior judgement in firms that automate routine analytical work? The answer will shape the next generation of experts (Autor & Thompson · angle 24).

7. Does AI use affect creative self-confidence, by scaffolding success or by making one’s own ideas feel inferior to the model’s? This is especially relevant where confidence is already low.

8. Which calibration interventions help learners know when to trust themselves over the model (Caplin et al. · angle 24)?

Each question can be studied inside the programmes proposed in the companion report, with randomised or waitlist designs and unassisted outcome measures.

Close: the machine that argues back

Kant’s guardians made immaturity comfortable: someone else would do the irksome work of thinking. Generative AI can be that guardian for everyone at once, and on its default settings it already is. The evidence of the past three years is consistent: tools that answer for people leave them less able when the tools are gone, narrow the range of their ideas, and quietly shift their views. Where decisions are at stake, human–AI teams underperform.

The same evidence shows the alternative is not hypothetical. Configured to ask rather than tell, to require a first attempt, to argue the other side, to coach rather than write, and to hand people back to one another, AI measurably improves unassisted skill, sharpens discernment and multiplies the reach of good human mentors.

The design choice is therefore also a civic choice. An infrastructure of thinking in the age of AI needs a tireless, honest sparring partner in every learner’s pocket — one that never lets them hide a thought, never tells them what to think, and always sends them back to the people who will stand beside them when they say it out loud.