Topology

TLDR

The lit review documents what the literature says about how AI restructures work, relationships, and meaning. This topology asks the structural question underneath: what depends on what? The field collapses to roughly fifty load-bearing claims across eight types (assumptions, empirical findings, mechanisms, guardrails, strategic recommendations, frameworks, open questions, distortions — see the legend on the graph). The graph encodes their dependencies so that the cruxes, the contested mechanisms, and the practical handles can each be read off cleanly without conflating them.

Scope. This topology takes the individual decision-maker as the unit of analysis — someone navigating career, relationships, and meaning architecture in a world where AI capability is accelerating. Institutional design (how organizations should structure AI integration), policy and governance (regulation, antitrust, AI safety), and aggregate-society dynamics (labor-share macro, geopolitics, civilizational risk) are out of frame. They affect the structural backdrop the individual operates in, but the strategic recommendations here are addressed to people, not to firms or governments. Population caveat: the empirical anchors (Brynjolfsson-Chandar-Chen ADP, OpenAI-MIT chatbot RCT, Therabot, Common Sense adolescent survey, Anti-Social Century, Acemoglu-Humlum-Vestergaard Danish nulls) are almost entirely US or Western. The strategic recommendations may not generalize cleanly across cultures with thicker baseline community, different family structures, or different work-identity expectations — readers in those contexts should treat the recommendations as hypotheses, not findings, until tested locally.

The single most useful conceptual move in this topology is separating the foundational cruxes from the reframer mechanisms from the logical guardrails. The foundational cruxes (A1 timelines, A2 relationship irreplaceability, A3 cumulative cognitive offloading, A4 structural relational depletion, A5 telic/atelic mapping) are the assumptions that, if falsified, would force rebuilding regions of the picture; A2 and A5 are most contestable, A3 has the most active falsification window (O3), A1 is the live-discourse fault line. The reframer mechanisms (G3 engagement-optimized substitution, G4 cognitive offloading bridge, G5 apprenticeship-ladder break, G7 telic exhaustion) don’t break the picture if they reverse — they change what the empirical findings mean, and their precise magnitudes are where the field is moving fastest. The logical guardrails (L1 finding/forecast/interpretation, L2 material-floor primacy, L3 substitution-vs-complement-is-the-wrong-binary) cannot be falsified — they can only be ignored, which is how most surface-level discourse on this topic proceeds.

The structural finding of the topology is that cognitive offloading (G4) is a cross-domain bridge node — the only mechanism that produces effects across all three domains under a single generating function. Practice atrophy when an effortful task is delegated produces deskilling at work (E16, E17), drives the competence-frustration that destabilizes meaning (G10), and erodes the relational depth that comes from sustained attention to another person’s emotional state. This is why S6 (maintain effortful practice) shows up as load-bearing in three different parts of the synthesis even though it looks like a workplace-productivity tip on its surface.

The field’s weakest links are not where public discourse focuses heat. The settled findings (E1 productivity gains, E14 Anti-Social Century baseline, E11 OpenAI-MIT dose-response, E10 Therabot RCT) are robust enough to bet on. The actual fragile zones in 2026: the cumulative-vs-calculator-analogue question for cognitive offloading (O3 falsifies A3 if it resolves toward calculator); whether asymmetric-adoption couples produce the relational-erosion patterns the dose-response findings predict (O2, the single largest empirical gap in the literature); whether AI-augmented atelic activities feel less meaningful (O4 falsifies A5 / S3 if it resolves yes); the cross-sectional-only nature of A3’s evidence base; the regulation-contingent stability of G3 (engagement-optimization could be moderated by policy); and the scope ambiguity of “AI relationships” (companion apps vs. general-purpose assistants used relationally vs. fine-tuned therapy bots are three different objects with different trajectories). On labor specifically, the optimist Ricardian story (G2) is currently less defensible than the apprenticeship-break story (G5) given E4 and E5 — not “contested” in the symmetric sense but actively losing ground as evidence accumulates. This topology is the input to model formalization (Stage 3); the cleanest formalization target is the competence-frustration / atelic-ballast trade-off — a quantified model of how much meaning a person loses per unit of telic work AI absorbs, partitioned by how much atelic infrastructure they had built before the shock.

The graph

Click a node for its claim, status, and load-bearing weight; hover an edge to see the relation type. Drag nodes to rearrange, drag empty space to pan, scroll to zoom. The variant toggles read the same graph through five lenses (full / vulnerability / flow / minimal / decision-leverage).

How to read this graph

Every claim in the lit review collapses to one of eight node types. Edges between them carry one of eight relations. Together they make the structure inspectable.

Node types

• A — Assumption. A foundational claim the rest of the picture leans on. Falsification would force rebuilding regions of the topology. The five A nodes are this stage’s cruxes.
• E — Empirical finding. A claim with direct empirical support — usually a number from a specific study or convergent meta-finding.
• G — Generating mechanism. The underlying dynamic that produces the empirical pattern. G nodes are the “why” answers behind E nodes; they’re where individual decisions can intervene because they expose the leverage point.
• L — Logical guardrail. A definitional or framing constraint that holds regardless of the empirical situation. L nodes cannot be falsified — they can only be ignored. They constrain how downstream claims can be interpreted.
• S — Strategic recommendation. The synthesis of empirical findings + mechanisms + frameworks into something a decision-maker can act on. S nodes form the high-leverage cluster in the leverage variant.
• F — Framework. A philosophical or conceptual tool the topology imports from outside the AI-discourse literature (Setiya, SDT, Arendt). F nodes are the lenses that make the mechanisms legible.
• O — Open question. A claim that would be load-bearing if resolved. O nodes carry the falsification windows for the cruxes — they show what evidence would change the picture.
• D — Distortion. A selective reading that ignores parts of the evidence base in service of a prior commitment. D nodes are not “wrong arguments” — they’re motivated readings that target specific E / S nodes.

Edge types

• dep (depends on): X requires Y to hold for X’s claim to be valid.
• sup (supports): Y is empirical or logical evidence corroborating X.
• gen (generates): Y produces X as its empirical signature.
• imp (implies): Y logically constrains how X can be interpreted.
• conf (confounds): Y creates an artifact or counter-pressure that complicates X.
• mod (moderates): Y changes the magnitude or direction of X without negating it.
• bridges (cross-domain): The same mechanism produces effects in two or more domains simultaneously. The blue dashed edges are this topology’s structural finding.
• attacks (selectively reads): a distortion vector targets X by ignoring or dismissing it.

Cruxes

The five A-nodes are a faithful transcription of the lit review’s own “Load-bearing assumptions” section — these are the assumptions the literature itself rests on, not a curated subset chosen for narrative shape. The five foundational assumptions are not equally uncertain, however. Naming the difference matters, because public discourse routinely conflates “this is a working assumption that has held so far” with “this is contested and a coin flip” — which leads to either complacency or fatalism depending on which assumption is in question.

A1 — AI advances without full substitution by ~2030. The live-discourse fault line. The slow-camp evidence (Acemoglu, Humlum-Vestergaard, the J-curve dynamics, Anthropic Economic Index aggregate task shares) supports A1; the fast-camp forecasts (Aschenbrenner, AI 2027, Korinek-Suh) target it. A1 has the most discourse heat, but falsification has asymmetric consequences across the S cluster. Under fast-AGI scenarios: S7 (career bet) collapses entirely — its targets (judgment, taste, AI-managerial framing) all become AI-substitutable. S2 (identity diversification), S3 (atelic ballast), S4 (in-person relationships), S5 (dose-limit), S6 (effortful practice) all survive — but their meaning shifts: they become more important, not less, because humans need their own practice ground when the entire knowledge-work surface is AI-saturated. The asymmetry is not “planning is wasted under fast timelines” — it is “skill-investment recommendations collapse, meaning-architecture and relational-infrastructure recommendations strengthen.”

A2 — Human relationships provide irreplaceable goods. The most philosophically defensible but functionally contestable. The structural argument (G8 embodied co-regulation + G9 mutual vulnerability + Vallor mirror-not-other) is strong; the contestation runs through whether most humans will care about the structural difference. If subjective satisfaction with AI relationships ends up functionally sufficient for a large population, the substitution-risk framing of E11 / E13 overstates the problem. This is the assumption most likely to be falsified by changing user behavior rather than by new evidence.

A3 — Cognitive offloading is cumulative, not transient. The crux with the most active falsification window. The cross-sectional evidence (E16 Gerlich-Stadler-Kosmyna, E17 Ehsan intuition rust) is suggestive but cross-sectional and short-duration. O3 — calculator-analogue versus cumulative atrophy — is the single best-defined empirical question in this topology, and resolving it within ~3 years is plausible. If A3 falsifies, S6 (maintain effortful practice) loses its evidential basis — though it would still hold as a hedge.

A4 — Relational depletion is structural. Carries the depleted-baseline that makes E11 / E13 / E14 substantially more dangerous than they would be in a thicker relational environment. If Gen Z’s high loneliness produces a counter-movement toward in-person community, A4 weakens. The Thompson Anti-Social Century data is the strongest evidence; it would take ~5 years of trend-reversal data to overturn.

A5 — Telic/atelic maps onto AI’s effects on meaning. The most novel-application crux. Setiya’s distinction was developed for midlife crisis; importing it to AI-meaning-disruption is a load-bearing hypothesis, not an empirical finding. O4 — whether AI-augmented atelic activities feel less meaningful — is the falsification window. If atelic activities are also degraded by AI proximity (AI companions changing the phenomenology of friendship, AI art changing aesthetic contemplation), S3 (atelic ballast) loses its structural basis.

Reframer mechanisms

Below the cruxes sit four mechanisms whose magnitudes — not whether they exist — determine what the empirical findings actually mean. They are this topology’s most active research frontier. Reversing them would not falsify the picture; it would recolor it.

G3 — engagement-optimized AI selects for substitution. Muldoon-Park’s structural argument: companion-app commercial incentives are the same as engagement-optimized social media — profit by prolonging the loneliness they purport to alleviate. If this mechanism dominates (default trajectory), E11 (OpenAI-MIT dose-response) and E12 (identity-discontinuity causal harm) get worse, not better, as the technology improves. If regulation, consumer demand, or shipped product changes weaken G3, the same models could shift toward complementarity rather than substitution. G3’s stability is contingent on commercial design choices, not on the underlying capability — which makes it the mechanism most directly affected by policy.

G4 — cognitive offloading via practice atrophy. The cross-domain bridge. Same mechanism, three downstream domains: work-deskilling (E16, E17), relational-depth erosion (S4), meaning-architecture disruption via competence-frustration (G10). If A3 falsifies in the calculator-analogue direction, G4’s downstream causal weight collapses across all three domains simultaneously — which is why O3 (the falsification window for A3) is the single most consequential open question in this topology.

G5 — apprenticeship-ladder break. AI absorbs entry-rung tasks → no rung-1 → expert pipeline collapses. Distinct from full-occupation substitution: only the bottom rungs get automated, but the bottom rungs were the training ground for everyone above them. This is the mechanism behind E4 (Brynjolfsson-Chandar-Chen entry-level disruption) and Shulman’s friction-cost argument against simple Ricardian comparative advantage. G5 is currently strengthening, not weakening, as E4 and E5 evidence accumulates — making it the reframer mechanism with the most directional momentum.

G7 — telic exhaustion. Setiya: telic activities self-annihilate on completion. When AI completes telic projects in seconds, the share of meaning staked on telic completion shrinks. This mechanism has the cleanest theoretical structure (the telic/atelic distinction either maps or it doesn’t — see A5) and the murkiest phenomenology (do atelic activities really stay untouched?). O4 is the falsification window. If G7 is correct, S3 (atelic ballast) is the highest-leverage strategic move available; if G7 fails because atelic is also degraded, the entire meaning-architecture half of this topology needs reconstruction.

The four reframers cluster the topology’s active uncertainty. The cruxes are where the discourse heat is; the reframer magnitudes are where the actual research is moving.

Weakest links

Six pressure points where the topology would crack or shift if pushed:

1. G2 — comparative-advantage equilibrium under binding constraints. Smith’s optimist case (AI faces some binding constraint → comparative advantage holds → Ricardian wage stability) is contested by Shulman’s friction-cost argument (humans introduce contamination / insurance / coordination friction → equilibrium wage to zero). G5 (apprenticeship-ladder break) is the empirical mechanism Shulman names — and G5 is currently strengthening as E4 / E5 evidence accumulates. The optimist Ricardian story is currently less defensible than the apprenticeship-break story given the available evidence — not “contested” in the symmetric sense but actively losing ground. Anyone betting on Autor-style middle-class-jobs reconstitution should require G5 to weaken first.

2. A3’s evidence base is cross-sectional only. The cognitive-offloading evidence cluster (E16 Gerlich-Stadler-Kosmyna, E17 Ehsan intuition rust) is convergent but methodologically thin: cross-sectional correlations and short-duration interventions. The 2+ year longitudinal study that would actually test A3 (calculator-analogue versus cumulative atrophy) doesn’t exist yet. If S6 (effortful practice) is the highest-leverage recommendation in the topology because of its bridge position, then the highest-leverage recommendation is being grounded by the methodologically weakest evidence layer. The hedge logic survives — practice is cheap insurance — but the strong claim that AI offloading durably erodes capacity remains a working hypothesis.

3. O2 — asymmetric-adoption couples. No peer-reviewed quantitative evidence yet exists on outcomes for couples where one partner uses AI heavily and the other does not. The single largest empirical gap in the literature. The technoference literature (McDaniel-Coyne 2016+) is the closest analogue but doesn’t capture the third-party-AI structure. Whoever runs the first such study will produce the most decision-relevant relational-AI finding of the next three years.

4. G3’s stability depends on commercial design choices. The engagement-optimized-substitution mechanism is currently the default trajectory of consumer AI products, but it is contingent — regulation (EU AI Act companion-app provisions, FTC enforcement), consumer demand shifts, or competitor product design could moderate it. Unlike G4 (which is grounded in cognitive psychology) or G7 (grounded in philosophy), G3 is a feature of the current commercial environment. The topology’s E11 / E12 / E13 cluster gets worse under G3-dominance and better under G3-moderation — these are not fixed empirical facts about AI but path-dependent outcomes.

5. “AI relationships” scope ambiguity. The empirical literature treats companion apps (Replika, Character.AI), general-purpose assistants used relationally (ChatGPT for emotional processing), and fine-tuned therapy bots (Therabot) as one object. They are three different objects with different generating mechanisms, different commercial incentives, and probably different long-run trajectories. The topology’s relationship-domain claims are sharpest for companion apps, fuzzier for general-purpose assistants, and possibly inverted for therapy bots — Therabot’s clinical-grade benefit (E10) coexists with companion-apps’ dose-response harm (E11) because they are not the same kind of thing.

6. D3 — productivity-only optimization. The most common distortion among the demographic most likely to be reading this topic (knowledge workers in AI-adjacent fields). Treats relational and meaning consequences as out-of-scope; ignores G4’s cross-domain bridge structure. The topology’s strongest defense against D3 is the bridge-edge cluster (G4 → S2 / S4 / S5) — if the reader cannot dismiss the bridge structure, they cannot consistently read this topology as a productivity-optimization frame. But the bridge claim leans on A3, which is fragile (see #2 above) — so D3’s defense is downstream of A3’s evidential strength.

Cross-domain bridges — the structural argument

The Reframer-mechanisms section names G4 as a cross-domain bridge. This section spells out the three channels and what is and isn’t supported by direct evidence in each.

1. G4 → work-deskilling (E16, E17). Practice atrophy in analytical work → reduced critical thinking, intuition rust. Direct evidence: Gerlich N=666 (cross-sectional correlation), Stadler-Bannert-Sailer (controlled comparison, lower-quality arguments under ChatGPT-aided research), Kosmyna MIT neuroimaging (reduced neural engagement under LLM-assisted writing), Ehsan year-long cancer-specialist field study (intuition rust in expert judgment). All cross-sectional or short-duration; longitudinal evidence does not yet exist (Weakest link #2).
2. G4 → relational-depth erosion (S4). Practice atrophy in emotional processing — outsourcing venting, sense-making, conflict rehearsal to AI — would erode capacity for sustained attention to another’s emotional state, tolerance of ambiguity in live conflict, sitting with unresolved tension. No direct empirical evidence yet exists for this channel. The mechanism is parallel to analytical deskilling, and the technoference literature (McDaniel-Coyne 2016+) shows phone-presence harms in-person interactions — but technoference is attention-split, not delegation, so it is not direct evidence for the offloading-eroding-emotional-capacity claim. This channel is currently a parallel-mechanism inference, not a finding. O2 (asymmetric-adoption couples) is the first study that would actually test it.
3. G4 → meaning architecture (G10, S2). Competence-frustration in SDT terms is threatened actually (not just symbolically) when offloading erodes the underlying skill — but this channel piggybacks on the work-deskilling evidence (channel 1). Identity foreclosure becomes more brittle when the foreclosed identity is no longer a domain you can practice. The phenomenology is well-described in interpretive sources (Balwit-Cowen, Mollick); empirical studies operationalizing it directly are scarce.

Honest summary of the bridge claim’s evidence base: Channel 1 has direct (if cross-sectional) evidence; channels 2 and 3 are parallel-mechanism inferences from channel 1. The bridge structure is the topology’s most consequential conceptual move and also its most empirically thin link — the strategic recommendation S6 hedges all three channels under one practice precisely because the bridge is plausibly load-bearing despite the evidence gap.

Variant readings

Each variant toggle reads the same graph through a different lens.

Full — all 50 nodes, all 80+ edges. The default view. Useful for orienting and for finding specific nodes you want to inspect.

Vulnerability — highlights the 5 cruxes plus weight-5 load-bearing nodes. Useful for thinking about what evidence would shift the picture and what advice survives even under crux-falsification. The takeaway is asymmetric: most strategic recommendations survive even if A1 falsifies fast (timelines collapse but convergent advice holds); strategic recommendations contingent on A3 / A5 are more fragile because they lean on hypotheses the field is actively testing.

Flow — A → G → E → S cascade with cross-domain bridges. Reads the topology causally: assumptions ground mechanisms, mechanisms produce empirical findings, findings drive synthesis, bridges connect the domain-specific stories. This is the variant that makes G4’s bridge role most visible.

Minimal — 15 nodes that recover the qualitative integrated picture. Removing any one breaks the qualitative shape. The minimal set is the answer to “if I had to fit this on one page, which 15 nodes?” — it concentrates on the cruxes, the bridge node, the load-bearing empirical findings, the material-floor guardrail, and the high-leverage strategic recommendations.

Decision-leverage — saturated nodes are high individual leverage (the seven strategic recommendations — direct decisions you can make about how to live). Mid-tone are medium leverage (mechanisms G4 / G6 / G7 / G10 and the discipline L1, plus the three frameworks — mental models that make the strategic recommendations applicable to your specific life). Faded nodes are structural or outside individual control (cruxes you can’t move, empirical findings you can’t change, mechanisms operating at population scale, distortions that affect public discourse). The variant answers “what should I actually do, and what do I need to understand to do it?” — the S cluster is the doing layer; the medium-leverage cluster is the understanding layer; everything else is the world the doing happens in.

Distortions

Four distortion vectors selectively read the same evidence base:

D1 — AGI-cancels-planning fatalism. Targets the entire S cluster. Move: assume A1 falsifies fast, conclude planning is wasted. The defense: the convergent S recommendations (S2 identity diversification, S3 atelic ballast, S4 in-person relationships, S6 effortful practice) survive even fast timelines. They are about meaning architecture and relational infrastructure, not about specific skill bets. Only S7 (career bet) is sensitive to A1.

D2 — slow-camp dismissal. Targets E4, E5, E16. Move: cite Acemoglu-Humlum-Vestergaard aggregate nulls (E8) as evidence that “nothing is happening.” Ignores that disruption is concentrated, not diffuse — entry-level (E4), freelance (E5), and creative work are leading indicators while aggregate effects lag. The defense: the coexistence of E4/E5 with E8 is the actual pattern, not their contradiction. G1 (Hulten/task-exposure aggregation) explains why aggregate effects are slow; G5 (apprenticeship-ladder break) explains why entry-level effects are sharp. Both can be true.

D3 — productivity-only optimization. Targets E11, E14, S4, S5. Move: optimize for AI-augmented output throughput; treat relational and meaning consequences as out-of-scope. The defense: G4 cross-domain bridges. The reader cannot consistently dismiss relational and meaning consequences while accepting the productivity gains, because the same offloading mechanism produces both.

D4 — material-blind class-position bias. Targets L2, E4, E5. Move: assume the decision-maker has economic runway; treat labor disruption as marginal versus relational/meaning advice. Inverts which advice is decision-relevant for whom. The defense: L2 (material-floor primacy). The framing in this topology is most useful for the knowledge worker with stable employment; for those whose floor is crumbling, S7 precedes S2-S6 and the labor economics is the existential question.

Adversarial + steelman

Distortions are how others read the picture badly. The honest move is to also ask: is my own picture wrong? The strongest reasonable critique of this topology — not a motivated misreading but a serious objection — is that the relationship and meaning sections are systematically over-weighted toward harm and under-weighted toward upside.

Objection. This topology’s relationship-domain cluster (E10–E15) leans heavily on companion-app pathology (Replika identity-discontinuity, OpenAI-MIT dose-response, adolescent companion uptake) and the depleted-baseline framing (Anti-Social Century, Surgeon-General loneliness mortality). It treats Therabot’s clinical-grade benefit (E10) as the lone counterweight rather than as the leading edge of a much larger therapeutic-AI category. Symmetrically, in the meaning domain, G7 (telic exhaustion) and G10 (competence-frustration) are framed as threats, but the same mechanism that compresses telic-completion meaning also dramatically expands the accessibility of telic projects to people previously excluded by skill or capital constraints (the “novice-skill compression” finding E2 hints at this but the topology doesn’t carry it forward). On cognitive offloading, the topology treats the cumulative-atrophy interpretation of A3 as the working hypothesis and the calculator-analogue as the contestation, when an honest reading of the evidence base could equally well treat them as a coin-flip. Net result: a reader following this topology’s strategic recommendations would over-invest in defensive moves (S5 dose-limit, S6 effortful practice, S2 identity diversification as defense) and under-invest in offensive moves (using AI to attempt projects previously out of reach, building genuinely new capabilities, leveraging novice-skill compression to enter new domains). The frame is “navigate the threats” rather than “exploit the asymmetry between current capability and current price” — and the asymmetry between current capability and current price may be the largest individual-leverage opportunity in the entire AI transition, larger than any of the harms the topology catalogues.

Steelman. This is partially correct and worth integrating. Three specific points survive scrutiny: (a) the topology’s relationship-domain framing genuinely is calibrated to companion-app pathology more than to the broader landscape of AI in relationships (assistants used in healthy ways within human relationships, AI as scaffolding for hard conversations, therapy-bot benefit at clinical scale) — the strategic recommendation S5 (dose-limit) is correct as stated for high-engagement companion use, but it does not generalize to “use less AI for emotional things” as a blanket rule. (b) On meaning, the topology under-develops the opportunity side of telic compression: when AI absorbs the rote bottom of telic work, the ambitious top of telic work becomes accessible to far more people. This is the “novice can now ship a real product / write a real book / publish real research” effect, and it is not currently encoded in the topology except as a productivity gain (E1, E2). (c) The strategic recommendations (S2 identity diversification, S3 atelic ballast, S4 in-person relationships) implicitly require social and structural conditions — community, time affluence, intact local relational infrastructure — that are themselves contingent and unevenly distributed. L2 (material-floor primacy) names economic security as a precondition; the topology does not name atelic-infrastructure primacy as a parallel precondition. For someone whose neighborhood has no third places, no inherited community, and no time-affluent peers, “build atelic ballast” is not just unhelpful advice — it is advice that presupposes resources the disrupted populations don’t have. This is the symmetric counterpart to D4 (material-blind class bias) but for atelic infrastructure rather than economic floor.

The version that survives. Three structural points hold even after the steelman is fully accepted:

1. The dose-response in E11 (OpenAI-MIT) is real, prospective, and not predicted by the upside framing. Heavy daily use predicts loneliness, dependence, and less in-person socialization regardless of modality. This is not “framing” — it is a well-powered RCT showing a specific harm pattern at high dose. The strategic recommendation S5 (dose-limit) survives as evidence-based even if the broader relationship-domain picture is uncertain.
2. The cross-domain bridge (G4) survives. Even if A3 turns out closer to calculator-analogue, the bridge structure holds at lower magnitude — practice atrophy across analytical, emotional, and meaning-architecture domains is the same mechanism, and S6 (effortful practice) is cheap insurance regardless of whether the magnitude is large or small.
3. The Waldinger and Surgeon-General-baseline findings (E15, E14) are not contestable in the upside-framing way. Whatever AI does to relationships, the baseline that human relationships predict health outcomes more than cholesterol is durable. The topology’s strategic recommendation S4 (embodied in-person relationships as non-negotiable) follows from this baseline, not from the companion-app harm cluster.

What changes if the objection is fully accepted: the topology should add an explicit “AI as exploitable asymmetry between capability and price” track in Stage 3 — a model of when AI-assisted ambitious-telic projects are net-positive vs. net-deskilling. Currently this is implicit in S7 (career bet) but deserves its own formalization target in a future stage. Adding it as an explicit handoff to Stage 3.

Open questions

The four open questions are the falsification windows for the assumptions:

The time-to-resolution column is my own estimate based on what kind of study would have to be designed and run to produce a defensible answer; treat as rough ordering, not calibrated forecasts. O1 has the most variance in resolution time and the largest discourse footprint; O3 has the cleanest empirical operationalization; O2 is the lowest-hanging fruit for an empirical study; O4 is the hardest to operationalize but would have the deepest impact if it resolves yes.

Next moves — preparing for model formalization (Stage 3)

The topology hands off one integrated formalization target to Stage 3, with two simpler alternatives if the integrated version proves untractable.

Recommended target — Net meaning-budget under AI absorption: ΔNet = ΔV − ΔM. Stage 3 should produce a single dashboard that quantifies both halves of the strategic question — defensive (meaning-loss from telic absorption) and offensive (value-gain from ambitious-telic access) — so the user can see the net across configurations of their own life. Variables:

• Defensive side (ΔM, meaning loss): telic share of identity (T), atelic ballast (B), AI-absorbable fraction of telic work (φ), competence-frustration sensitivity (κ). Form: ΔM = −κ·φ·(T − B). Boundary: B ≥ T → ΔM ≈ 0 (atelic-ballast hypothesis holds); T ≫ B → catastrophic ΔM (foreclosure path).
• Offensive side (ΔV, value gained): pre-attempt skill stock (s), AI capability on the task (a), feedback-loop richness (f), retained effortful practice (ρ). Form: ΔV = α·a·(1 − s) when f and ρ are above thresholds (novice-skill compression + access expansion); ΔV ≤ 0 when f or ρ falls below thresholds (self-automator trap E7).
• Net: ΔNet = ΔV − ΔM, with explicit interaction terms — high B raises both ΔV (more identity surface that survives) and ΔM-resistance.

This integrated form addresses both the original defensive crux-of-cruxes question (A2 + A3 + A5 all engaged) and the upside-side critique surfaced in the Adversarial + steelman section. It gives Stage 4 (data) clean targets — fit κ from existing time-use + meaning surveys; fit f and ρ from the BCG self-automators study; calibrate α from productivity-gain literature (E1, E2). It gives Stage 5 (build) a direct path to a shippable interactive dashboard where users tune their own (T, B, φ, κ, s, a, f, ρ) and see ΔNet.

Fallback alternatives if the integrated form proves untractable:

• Defensive-only: ship just the ΔM model. Cleanest derivation; engages A2 / A3 / A5 directly; carries forward the adversarial-section asymmetry but is still genuinely useful.
• Dose-response curve: fit OpenAI-MIT + Therabot + identity-discontinuity into a continuous curve mapping daily AI-emotional-engagement minutes to outcomes, conditional on baseline relational thickness. Empirically testable, narrowest scope, addresses E11 / S5 specifically. Useful as an embedded module inside the integrated model rather than as the standalone target.

Recommendation: pursue the integrated ΔV − ΔM target. If the interaction terms prove too speculative to ground in existing data, fall back to the defensive-only ΔM model and treat ΔV as a future Stage-3 expansion.

Glossary

• A1–A5 — the five foundational assumptions; see Cruxes section.
• Apprenticeship ladder break (G5) — AI absorbs entry-rung tasks → no rung-1 → expert pipeline collapses. Distinct from full-occupation substitution.
• Atelic / Telic (Setiya) — atelic activities are realized in the doing (durative — friendship, contemplation); telic are aimed at completion (self-annihilating — finishing a project).
• Bridge edge — a cross-domain dependency; the same mechanism produces effects in two or more of work / relationships / meaning. Visualized as blue dashed lines in the graph.
• Cognitive offloading (G4) — delegating an effortful cognitive task to AI; mechanism for deskilling via practice atrophy. The structural bridge of this topology.
• Competence frustration (G10) — Self-Determination-Theory mechanism whereby AI threatens the felt sense that one’s skill makes the difference, both symbolically and actually.
• Crux — a foundational-assumption node whose falsification forces rebuilding regions of the topology. The five A nodes are this topology’s cruxes.
• Decision-leverage — informal measure of how directly an individual can act on a node; high / medium / low. Used in the leverage variant.
• Distortion (D) — a selective reading that ignores parts of the evidence base in service of a prior commitment. Not “wrong” — motivated.
• Engagement-optimized substitution (G3) — companion-app commercial incentives select for prolonging the loneliness they purport to alleviate (Muldoon-Park 2025).
• Finding / Forecast / Interpretation (L1) — the three evidence-weight categories; ignoring the distinction is the most common analytical error in AI-transition discourse.
• Foreclosure (Marcia) — adopting a single-strand identity without exploration; predicts catastrophic disruption response when that identity is destabilized.
• Jagged frontier (Dell’Acqua-Mollick) — AI capability is uneven across superficially similar tasks; large gains inside the frontier, degraded performance outside.
• Material-floor primacy (L2) — for anyone whose material floor is insecure, labor disruption IS the existential question. Overrides relational/meaning framing for that population.
• Relational depletion (A4) — the 60-year decline in face-to-face sociality (Putnam → Thompson Anti-Social Century) that AI lands into.
• Self-automator (Randazzo et al.) — third class beyond centaur/cyborg; delegates both what and how to AI; 27% of consultants in BCG study; no skill development in either domain.
• Setiya midlife framework — telic activities self-annihilate on completion; atelic activities are realized in the doing; the most useful philosophical instrument in this lit review for strategic life decisions under AI.
• Working Alliance Inventory (WAI) — measure of patient–therapist therapeutic alliance; Therabot RCT scored 3.59, comparable to outpatient norms.