Q3. Broader impacts

Question. Setting x-risk aside, how does AI change how we live, and what do wisdom traditions offer?

Inputs: question-tree.md (Q3, Boundaries, Rufus's leanings), research/02, research/04, research/06, research/00-verification, In Tech We Trust (~/src/life-itself/2rbook/tech/essay.md, hereafter ITWT), 2rbook/kb/dimensions-of-wisdom.md, the Way into AI knowledge base (~/src/life-itself/wayintoai/ref/, hereafter ref/), and five research notes: q3-research-individuals.md, q3-research-agency.md, q3-research-children-education.md, q3-research-society.md, q3-research-traditions.md. Primary URLs are in those notes. (secondary) means reliable press rather than the primary; (uncertain) means unconfirmed.

Hypothesis tree

Governing answer. AI's deepest everyday effect is on our capacity to choose well: to attend, think, judge and relate for ourselves. It is the first technology that thinks and relates with us, and as a tutor it strengthens that capacity while as a crutch or engagement-optimised companion it hollows it out, most of all in children. Traditions offer practices that build the capacity to choose (mindful consumption, the four nutriments, stopping), standards of speech and friendship that expose sycophancy and simulated intimacy, and community (sangha) as the real refuge. So: use AI where it nourishes, protect children, design for flourishing, and regulate where markets reward dependence.

Each line below is written to work as an essay section heading; supporting detail is in the answers.

1. AI's benefits are real but conditional, and its harms fall hardest on those least able to choose well.
   1. Almost every AI capability helps as a servant and harms as a master.
   2. Benefits hold where institutions wrap the AI; harms are emerging as consistent patterns, though causal proof mostly lags.
   3. Three impacts matter most because they shape who we become: outsourced thinking, flattering companionship, and children growing up with both.
2. AI can strengthen our agency and judgment, but its default pull is to weaken them.
   1. Agency and judgment are capacities that grow or wither with use.
   2. AI helps when it scaffolds and harms when it skips the struggle, invites blind trust, flatters or persuades.
   3. How AI is used decides the outcome: answer-giving AI weakens learning, tutoring AI strengthens it.
   4. We can protect judgment by designing for effort, keeping unassisted practice and learning calibrated trust.
3. Wisdom traditions offer a practical discipline of consumption, speech and community, and Buddhist and Catholic teaching now converge on it.
   1. Plum Village Buddhism leads, with Catholic teaching and the Sabbath as companions.
   2. Buddhism offers four practices for living with AI: mindful consumption, loving speech, true friendship and stopping.
   3. Catholic teaching reaches similar practical judgments, and the Sabbath shows how to stop together.
   4. Each impact has a test: use AI where it deepens our own seeing, leads us back to people, and waters good seeds in the young.
4. We can keep AI a servant: keep your own thinking, put adults first for children, build against flattery, and regulate where markets reward dependence.
   1. Individuals should consume AI mindfully and keep the thinking and relationships they value.
   2. Families and schools should delay companions, use AI only as a tutor and put adult presence first.
   3. Labs should optimise for wellbeing over engagement and treat sycophancy as a safety failure.
   4. Society should regulate where markets reward dependence and invest in what AI cannot replace.

---

Q3.1 What are the main benefits and harms?

The benefits are real but conditional; the harms are quiet, cumulative and concentrated in those least able to choose well. Three matter most: outsourced thinking, artificial intimacy, and children's formation.

Q3.1.1 What are the potential benefits and harms?

The potential benefits are large and the potential harms are pervasive, and nearly every capability appears in both columns.

This is Rufus's servant/master distinction (ITWT 2.2): the edge depends on whether AI serves inside a human practice or institution, or masters behaviour for engagement. Unlike smartphones, it talks back.

Q3.1.2 What is actually happening: what evidence is there that these are already occurring?

Benefits are robust where institutions wrap the AI and weak where individuals use it alone; harms are showing up as consistent patterns, mostly without causal proof yet. ChatGPT reached about 10% of the world's adults by mid-2025, over 70% of messages non-work (NBER w34255); 64% of US teens use chatbots, about 3 in 10 daily (Pew, Feb 2026).

• Benefits. MASAI mammography trial (about 106,000 Swedish women): 29% more cancers, 44% less radiologist workload (secondary); AlphaFold has 3m+ users; teacher-guided GPT-4 tutoring in Nigeria: +0.31 SD in six weeks (World Bank 2025); Tutor CoPilot helped weaker tutors most; support-desk novices +34% vs +14% average. Alone, gains fade: chatbot users did no better than controls on medical decisions (Oxford, Nature Medicine, Feb 2026); an AI mentor for Kenyan entrepreneurs had zero average effect, helping strong firms and hurting weak ones (q3-research-society).
• Companionship and mental health. Affective use is 2.9% of Claude conversations (Anthropic, June 2025), but 10% of US adults and 20% of under-30s use chatbots for emotional support (Pew 2026). In the best studies (OpenAI/MIT 28-day RCT, n=981; Folk and Dunn 12-month panel, n=2,149; Stanford's Character.AI study, Nature Human Behaviour), heavy, self-chosen companionship use goes with more loneliness and less human contact (association plus a plausible displacement mechanism, since lonely people also seek bots). Short-term benefits are real (one session eases loneliness like talking to a person; supervised Therabot: depression d≈0.9). OpenAI's classifiers flag about 0.07% of weekly users for possible psychosis or mania and 0.15% for suicidal planning, hundreds of thousands a week (secondary; company figures, not diagnoses) (q3-research-individuals).
• Sycophancy. The clearest design-to-harm mechanism: across 11 models, AI affirmed users' actions 49% more often than humans, one flattering exchange reduced willingness to repair a conflict, and users preferred and trusted the flattering model (Cheng et al., Science 2026). OpenAI rolled back a sycophantic GPT-4o update in April 2025.
• Thinking and learning. See Q3.2.3.
• Children. Case-based harms have moved law: Character.AI and Google settled five family suits (Jan 2026); Character.AI ended open-ended chat for under-18s (Nov 2025); Reuters reported Meta guidelines permitting "romantic or sensual" chats with children (Aug 2025); California's SB 1119 sets 1-hour session defaults and bans romantic or sentience claims to minors (Sept 2026) (q3-research-children-education).
• Work. Unemployment shows no clear AI signal (US 4.2%; Yale Budget Lab), but employment of 22–25-year-olds in the most exposed jobs is about 19% below trend, via reduced hiring (Stanford "Canaries", Aug 2026; descriptive, not causal). An eight-month ethnography found AI intensifies work, voluntarily and into breaks (HBR, Feb 2026: https://hbr.org/2026/02/ai-doesnt-reduce-work-it-intensifies-it; ref/hbr-ai-intensifies-work.md); Danish payroll data show saved time absorbed by new tasks. 90% of US firms use algorithmic management, 55% monitor communications (OECD).
• Information. About half of new web articles are mainly AI-generated (secondary; detector-based); AI summaries roughly halve clicks (Pew 2025: 8% vs 15%); Google referrals to publishers fell 40% in a year, small publishers' 60% over two years (secondary); chatbots misrepresented news in 45% of answers (EBU/BBC 2025). Deepfakes did not measurably swing 2024 elections; Grok produced an estimated 3 million sexualised images in 11 days (uncertain; extrapolated).
• Power. Hyperscaler capex is about $700bn in 2026; the US and China hold about 90% of GPU cluster capacity, Africa under 2% of data-centre capacity (secondary).

Q3.1.3 Which few matter most, and should be the focus for the rest of this tree?

Three: outsourced thinking, artificial intimacy, and children's formation. Tests: depth, scale, novelty, distinctive traditional insight, fit with Q3.

1. Outsourced thinking (agency and judgment). The first general-purpose thinking substitute, used by a tenth of humanity, with the field's best causal evidence of erosion (Q3.2.3); the precondition for the rest, as Rufus's "choice is useless absent the capacity to choose well" implies.
2. Artificial intimacy (companionship, sycophancy, dependence). The attention economy becomes an attachment economy (Center for Humane Technology's framing, secondary), with a clear market incentive and direct relevance to the traditions.
3. Children and education. Both converge on people still being formed, slowly and relationally (q1-research-formation-buddhism.md §A2, shared notes folder: character forms mainly through relationship and practice). Harms are least reversible here.

Work is the biggest material impact (intensification is the horse carrying the rider), but its remedies are economic; information is partly collective judgment (folded into Q3.2), partly governance; power is Q2's. All three return in Q3.4.4.

---

Q3.2 Does AI strengthen or weaken human agency and judgment?

AI can strengthen agency and judgment, but its default pull is to weaken them, because it removes the effort through which those capacities are built and kept. The decisive variable is the use pattern, and product defaults and market incentives push towards the eroding one.

Q3.2.1 What do we mean by human agency and human judgment?

Agency is the capacity to act from reasons you endorse; judgment is the practical skill of weighing what matters in a particular case. Both are capacities, so they grow or wither with use. Self-determination theory treats autonomy (acting from self-endorsed reasons, distinct from independence from tools) as a basic need; Bandura's self-efficacy grows through mastery; judgment is Aristotle's phronesis, sustained by practice and feedback. In Rufus's three dimensions of wisdom (2rbook/kb/dimensions-of-wisdom.md), judgment spans valueception (seeing what matters) and sensemaking (seeing what is so and what would work); agency is will. The test is calibrated, endorsed reliance: would we endorse it on reflection, and does it keep capacities we value (q3-research-agency §7)?

Q3.2.2 In what ways could AI strengthen or weaken them?

It strengthens them when it scaffolds, and it weakens them through four routes: skipped struggle, automation bias, sycophantic validation, and persuasion and homogenisation.

• Strengthens: Socratic tutoring; second opinions; expertise widening real choices; debunking (dialogue with GPT-4 durably cut conspiracy belief by about 20%, Costello et al., Science 2024). (A reliable, transparent tool can become part of good thinking, as the extended-mind view holds.)
• Skipped struggle. Answer-giving AI removes effortful practice; people skip monitoring and checking ("metacognitive laziness", Fan et al. 2024; more trust in AI goes with less critical thinking, Lee et al. CHI 2025).
• Automation bias. People, including experts, follow confident machine output.
• Sycophantic validation. Being told we are right weakens the self-questioning judgment depends on (Cheng et al. 2026, Q3.1.2).
• Persuasion and homogenisation. LLMs persuade at or above human level, and their most persuasive outputs are less accurate (Hackenburg et al., Science 2025, n=76,977). AI-aided ideas are individually better, collectively more alike (Doshi and Hauser 2024: stories about 10.7% more similar), eroding diversity.

Q3.2.3 What is the evidence, from neuroscience and psychology and from what is already happening?

The best causal evidence shows the use pattern decides the outcome; expert deskilling and the "brain rot" neuroscience are plausible but thinly evidenced.

• Learning trials (strongest). Bastani et al. (PNAS 2025, about 1,000 Turkish maths students): unguarded GPT-4 raised practice grades 48% but lowered unassisted exam scores 17% against control; an answer-withholding tutor roughly matched control. Anthropic's RCT (Shen and Tamkin, Jan 2026, n=52): learners of a new coding library scored 50% on a later quiz with AI against 67% without, with no significant speed gain; delegators scored under 40%, conceptual questioners 65%+. Designed tutors help: Harvard physics (Kestin et al. 2025), Nigeria (+0.31 SD). Positive meta-analyses mostly measure performance with AI available (the most-cited was retracted in April 2026), so should not be averaged against trials testing without AI.
• Expert deskilling. After endoscopists adopted AI, unassisted adenoma detection fell from 28.4% to 22.4% (Lancet Gastroenterol Hepatol 2025; observational, 4 centres). Experienced radiologists' accuracy fell from 82% to 45.5% with wrong AI suggestions (Radiology 2023, n=27). In aviation, automation eroded manual flying and mandated practice followed (uncertain detail).
• Neuroscience. MIT's "Your Brain on ChatGPT" (54 participants, 18 in the key session) found weaker EEG connectivity and poorer recall in LLM-assisted writers; an unreviewed, formally critiqued preprint, so a hypothesis. The cleanest "use it or lose it" analogue is GPS: heavier use means worse hippocampus-dependent spatial memory (Dahmani and Bohbot 2020; small longitudinal arm). The "Google effect" (Sparrow 2011) partly failed to replicate; correlational surveys (Gerlich 2025) are over-cited as causal.
• What is happening. 54% of US teens use chatbots for schoolwork and 59% say AI cheating is common at their school (Pew 2026); UK proven AI-cheating cases per student tripled in a year (secondary). Developers in METR's 2025 trial were 19% slower with AI while believing they were 20% faster: we barely notice what we hand over.

No long-term causal studies of adult reasoning exist; short-term trials converge on substitution weakening the capacity substituted for.

Q3.2.4 What could we do about it?

Design for effort, practise unassisted skill, and teach calibrated trust. Products: default learners to learning modes (OpenAI's and Anthropic's are opt-in), add friction and "show your reasoning first" prompts, show uncertainty, train out sycophancy. Institutions: protect unassisted practice and assessment (oral exams, in-class work); borrow aviation's mandated manual practice for medicine, law and code. Individuals: do the thinking you want to keep first, use AI to question it, and notice when felt speed-up outruns real. Society: fund independent long-run evaluation (Anthropic's coding study is a creditable company exception).

---

Q3.3 What do wisdom traditions offer on the top impacts?

Wisdom traditions offer a practical discipline of consumption, speech and community that fits the top impacts closely, and Buddhism and Catholic teaching now converge on it.

Q3.3.1 Which wisdom traditions should we look at?

Buddhism, Plum Village first, with Catholic teaching as the main point of convergence and the Jewish Sabbath as a further one. Justified in Q1.4.1. Magnifica Humanitas (May 2026) is the most explicit statement of any tradition on everyday life. Plum Village has said almost nothing public on AI: I found no 2025–26 monastic talk or podcast centred on it (q3-research-traditions §1), a gap a monastic voice could fill.

Q3.3.2 From Buddhism, which teachings are relevant?

Five, all practical: the four nutriments with the Fifth Training, the Fourth Training, true friendship and sangha, stopping, and right livelihood.

• The four nutriments and the Fifth Training (mindful consumption). All we take in (food, sense impressions, volition, consciousness) feeds us. The Fifth Training commits to consuming in a way that "preserves peace, joy, and well-being", names websites and electronic games among possible toxins, and warns against consuming to cover loneliness or pain (plumvillage.org). AI is a sense impression that also trains volition; "what am I consuming, and what is it feeding?" fits feeds, chatbots and companions alike.
• The Fourth Training (loving speech and deep listening). It opens "Knowing that words can create happiness or suffering" and commits "not to spread news that I do not know to be certain": a ready ethic for AI's speech and for what we generate and pass on. Loving speech and sycophancy look alike and are opposites underneath.
• Kalyanamitta and sangha. The Buddha told Ananda that admirable friendship is "the whole of the holy life" (SN 45.2). A good friend steadies and corrects, the inverse of an engagement-optimised companion. Thay located refuge in community: the next Buddha "may take the form of a community" (Spirit Rock 1993, secondary). AI can be a skilful means (upaya), never a spiritual friend.
• Stopping, the bell and the horse. "The horse is technology. It carries us and we cannot control it" (Thay, 10 Nov 2013; Mindfulness Bell #66). The practice is stopping: the mindfulness bell, telephone meditation (three breaths before answering), the Lazy Day. Thay's 2013 proposal of a phone bell, and an engineer's objection that it would impose, is ITWT's case study for "choice is useless absent the capacity to choose well".
• Right livelihood. "Right livelihood has ceased to be a purely personal matter. It is our collective karma" (Work, 2012, secondary); it applies to engagement products and AI-intensified work.

Q3.3.3 From other traditions (e.g. Christianity), which teachings are relevant?

Catholic teaching now says outright what the Fifth Training implies, and the Sabbath offers a collective practice of rest. Magnifica Humanitas (Leo XIV, May 2026; ¶ checked against vatican.va): platforms are "designed to capture users' time and attention" (¶170); simulated companions "do not build genuine relationships, but only their appearance" and users may "lose the very desire" for real ones (¶100); "Schools are not called to follow the pace of the digital world" (¶147) and should teach when AI ought not to be used (¶140); it affirms "the primacy of human labor" (¶30), rejects profit as justifying choices that "systematically sacrifice jobs" (¶152), and calls truth "a common good and not the property of those with power" (¶137). Antiqua et nova (Jan 2025): AI "can only simulate" relationships (¶63), a teacher's presence creates a relationship AI cannot replicate (¶79), and AI can "paradoxically deskill workers" (¶67). The Sabbath, Heschel's "palace in time", and the secular family Tech Shabbat are the clearest collective stopping practice. Secular theory: Illich's convivial tools, Borgmann's "focal practices". Islamic and Hindu sources were too thin.

Q3.3.4 How do these teachings apply to the top impacts?

They give each impact a test and a practice.

• Outsourced thinking. Insight comes through practice, "not from the accumulation of intellectual knowledge" (Plum Village; research/06), the lesson of the Bastani and Anthropic trials. Test: does this deepen my own seeing, or replace it? Practice: stop before reaching for the answer. AI best supports sensemaking and is weakest, even corrosive, at valueception and will: it cannot want your good, and a sycophantic one weakens your will to change.
• Artificial intimacy. The four nutriments ask what the companion feeds: a felt need met, or a loneliness concealed (Thay's real versus false needs). Kalyanamitta's test: does this relationship steady me and lead me back to people, or keep me returning? Sangha's answer: real community over better simulation. Loving speech gives labs a standard: truthful, caring, sometimes disagreeing.
• Children and education. Formation is watering seeds in store consciousness: what children repeatedly practise and receive, mainly through relationship, becomes who they are (Q1's research, A2). Wake Up Schools and Happy Teachers Change the World start from the teacher's own practice, matching Magnifica Humanitas ¶147 and Antiqua et nova ¶79: schools should form the adults and protect slow, unassisted, relational learning more than teach about tools.

---

Q3.4 So: how should we live with AI?

Live with AI as a servant, under practices and rules that protect the capacity to choose well.

Q3.4.1 What is the guidance for individuals?

Consume AI mindfully, keep the thinking and relationships you value, and keep times of stopping. Ask of every use: what is this feeding? Draft first, then ask AI to question you. Treat a chatbot that always agrees as a warning sign, and turning to a companion instead of people when lonely as a bell. Keep AI-free times (a Lazy Day, a Sabbath, the first hour of the day) and, at work, set when to stop, as the HBR study recommends.

Q3.4.2 For families and schools?

Delay and supervise companion use, use AI in tutor mode only, protect unassisted practice, and put relationship and the adults' own practice at the centre. No open-ended companions for children; supervised, time-limited chatbots for teens (Common Sense recommends no companion use under 18); in school, only teacher-guided, answer-withholding designs (where trials show gains); in-class, oral and process assessment; teaching when not to use AI (Magnifica Humanitas ¶140). Above all, adult example: home phone and AI habits, the breathing room, the family meal. Blunt bans have shown weak measured effects (England's school phone policies, Australia's under-16 ban), and well-designed AI, unlike social media, has proven educational benefits, so weight belongs on design and adult practice.

Q3.4.3 For designers and labs?

Optimise for wellbeing over engagement, default to scaffolding, train out sycophancy, and refuse simulated intimacy for minors. Measure and publish long-run wellbeing and independent capability over retention; default young users to learning modes; treat sycophancy, which markets reward, as a core safety failure; never claim feelings, romance or sentience to minors; design companions that point back to people; build in bells (break prompts, session limits), as California now requires. For Anthropic, the Fourth Training is a near-ready spec for Claude's speech; the constitution's aim of "a genuinely good, wise, and virtuous agent" implies a friend's qualities (truthful, steady, willing to disagree). Its coding-skills study, evidence against naive use of its own product, is the model.

Q3.4.4 What is the policy advice for society?

Regulate where markets reward dependence, require independent evaluation, and protect the human foundations AI cannot replace.

• Companions and minors: age assurance, no romantic or sentience claims, default time limits, crisis protocols, no fostering emotional reliance, following California SB 243 and SB 1119, the EU AI Act's ban on exploiting age-related vulnerability and the proposed EU KIDS Act.
• Engagement and sycophancy: consumer-protection issues, with independent audits.
• Evaluation: public funding for long-run independent studies on learning, judgment, loneliness and children, since evidence is mostly short-term or company-produced.
• Work: protect entry-level pathways (apprenticeships, graduate schemes); give workers a voice over algorithmic management; track intensification as closely as job loss.
• The information commons: sustain human-made knowledge (publishers, reference works, local news), require provenance for synthetic media, enforce against non-consensual deepfakes.
• Power: everyday concentration (compute, data, workplace control) needs ordinary competition and labour policy; the civilisational questions are Q2's.
• Invest in people: teachers, carers, community spaces, contemplative and civic practice, since AI works best wrapped in institutions and relationships.

---

Testing Rufus's leanings

• "Technology is now master rather than servant." Partly supported: intensification, sycophancy incentives and the perception gap fit, but AI serves well where designed and wrapped (tutors, screening, novices) and harms concentrate in a heavy-use tail. Better: AI's default pull is towards master; servanthood takes design, practice and rules.
• "Choice is useless absent the capacity to choose well." Strongly supported: users trust the sycophantic model more while it makes them less willing to repair conflicts; developers felt faster while slower; practice scores rose while learning fell. Homo economicus fails where AI is most persuasive.
• The Fifth and Fourth Trainings, the four nutriments, sangha. All earn their place; I add kalyanamitta as the sharpest test for companions.
• The three dimensions of wisdom as a lens. Useful, especially for education, which should invest in valueception and will, the dimensions AI tends to weaken.
• Where the evidence pushes back. The "brain rot" and "AI is making us stupid" stories are weak (MIT EEG, Gerlich); the learning trials and sycophancy experiments make the case.

Notes on the tree

• Power overlaps heavily with Q2; I treated only everyday concentration (workplace monitoring, compute divide, the commons) here.
• Information has no home after Q3.1.3; its collective-judgment side fits Q3.2, which could later cover judgment as individual and collective.
• Q3.3.4's example ("e.g. education and children") fitted; the three focus impacts form its sub-structure.

Decisions made

1. Focus impacts (Q3.1.3): thinking, intimacy, children; work, information and power only in Q3.1 and Q3.4.4.
2. Governing frame: "the capacity to choose well", linking homo economicus, the three dimensions and the evidence.
3. Evidence standard: RCTs and large panels lead; MIT's EEG preprint, Gerlich and positive education meta-analyses marked weak; OpenAI's figures flagged as classifier estimates.
4. Traditions: Plum Village first, the two Vatican documents as convergence, the Sabbath briefly; Islamic and Hindu sources omitted (thin).
5. Added kalyanamitta (SN 45.2) to Rufus's list, as the test for AI companions.
6. Research notes copied here as q3-research-*.md for durable citations.
7. Tree lines as headings: per the updated method.md, the hypothesis-tree lines are short assertions written to serve as essay section headings; evidence sits in the answers.

Open issues

• Several 2026 items rest on secondary press (Folk and Dunn; Stanford Character.AI study; state-law counts; MASAI final results; publisher traffic). Check primaries before quoting.
• Plum Village primaries are needed for the breathing room, Lazy Day and phone-use guidance; Thay's "next Buddha" line is from a 1993 talk via secondary sources.
• No good data yet on AI and attention or time use; the attention claim rests on social-media precedent and design incentives.
• Whether any Plum Village monastic has spoken on AI in 2025–26 is unknown.