AI often raises short‑term wellbeing and convenience but can, through repeated use, erode skills, autonomy and social ties — producing long‑run welfare losses; the authors formalize this two‑horizon risk (DCDT) and propose an AI‑Happiness Impact Assessment to spot and prevent reversals.

Main Finding

The paper proposes Dynamic Capability-Dependency Theory (DCDT): AI-mediated happiness is a two-horizon process. Short-term gains (affective relief, convenience, enjoyment) can coexist with long-term harms if repeated use erodes human competence, autonomy, relatedness or creates dependency. The paper formalizes this as a dynamic state model, defines a Temporal Well-Being Reversal condition (initially positive effects become negative over time), derives testable propositions, and offers an AI‑Happiness Impact Assessment to evaluate both immediate and durable welfare effects. The central normative claim: AI should be judged not only by present satisfaction but by whether repeated use leaves people more capable, autonomous, connected and able to pursue meaningful lives.

Key Points

• Two-horizon framework:
  • Horizon 1 — immediate affective gains: convenience, symptom relief, productivity boosts, temporary reductions in loneliness.
  • Horizon 2 — accumulated stocks: competence, autonomy, human relatedness, meaning, and dependency that evolve with repeated use.
• Temporal Well-Being Reversal: a formal condition where initially positive AI impacts reverse as capability erosion, relational substitution, or dependency accumulates.
• Determinants of outcomes: system design, intensity/frequency of use, degree of user autonomy, relational context (human vs AI substitutes), and time horizon.
• Empirical evidence: randomized trials, longitudinal studies and workplace deployments show genuine near-term benefits but heterogenous longer-term outcomes.
• Practical tool: AI-Happiness Impact Assessment to evaluate both immediate effects and durable impacts on capability and dependency.
• Central conceptual shift: AI changes the production function of happiness by reallocating effort, attention, agency and social ties across time.

Data & Methods

• Theoretical formalization:
  • Dynamic state model capturing stocks (capability, autonomy, relatedness, dependency) that evolve with use and feed back into experienced well-being.
  • Formal condition (Temporal Well-Being Reversal) that identifies when short-term gains predict long-run losses.
• Empirical program outlined:
  • Randomized controlled trials for short-run causal effects on affect, symptoms and productivity.
  • Longitudinal and panel studies to observe stock accumulation/erosion and detect reversals.
  • Field deployments and workplace studies to measure productivity, skill trajectories and relational substitution.
  • Quasi-experimental designs, difference-in-differences, instrumental variables and natural experiments to estimate causal long-run effects where RCTs are infeasible.
• Measurement targets:
  • Momentary affect and utility (experience sampling, surveys).
  • Stocks: measures of competence/skill, autonomy, social connection, meaning, and behavioral dependence on AI.
  • Usage intensity, autonomy of decision-making, and relational context variables.
• Testable propositions (examples):
  • High-frequency, low-autonomy use increases dependency and raises reversal risk.
  • Designs that actively scaffold skill acquisition reduce long-run negative effects and may produce net durable welfare gains.
  • Social-AI substitutes decrease human relatedness stocks more than assistive-AI that augments social interactions.

Implications for AI Economics

• Welfare measurement: standard welfare analyses must incorporate dynamic, intertemporal effects on human capital, autonomy and social capital — not only instantaneous utility or productivity gains.
• Production function of happiness: AI shifts inputs and marginal returns (effort, attention, agency, relationships) across time; economists should model these intertemporal re-allocations when valuing AI.
• Cost–benefit and policy:
  • Evaluate deployments with discounted long-run stocks (capability, autonomy, relatedness) and account for potential Temporal Well-Being Reversal.
  • Design incentives (regulation, procurement standards, subsidies) to favor AI that augments user capability and autonomy rather than substitutes and creates dependency.
  • Consider disclosure, monitoring and post-deployment impact assessments (AI-Happiness Impact Assessment) as part of governance.
• Labor and human capital:
  • Employers and platforms should weigh short-term productivity gains against possible erosion of worker skills and long-run employability.
  • Policies may be needed to encourage complementary use (training, skill-sparing interfaces) to preserve human capital.
• Markets and product design:
  • Business models that monetize engagement might bias toward dependency-creating design; market and regulatory interventions may be necessary to realign incentives with durable welfare.
  • Standards and metrics for “durable value” (does repeated use increase users’ capabilities?) should be developed and adopted.
• Research agenda for AI economics:
  • Develop intertemporal welfare models incorporating capability and relational stocks.
  • Run longitudinal RCTs and quasi-experiments to quantify reversal risks and heterogeneity.
  • Create validated measures of dependency, capability erosion/growth, and social-capital substitution.
  • Quantify externalities (e.g., social capital loss) and second-order effects in labor markets and public goods.

Overall, DCDT reframes assessment of AI value: beyond immediate utility and productivity it calls for measuring whether AI use builds or depletes the durable foundations of human well-being.