Reframing AI alignment as an economic problem, the authors argue, would shift policy from narrow technical fixes to systemic reforms—slowing growth incentives, capping resource use, and governing AI as a commons—to reduce social, environmental and existential risks; they call for prioritising tool-like, autonomy-enhancing systems over agentic AI and for new economic theories rooted in post-growth thinking.

Main Finding

The paper argues that the AI “alignment problem” is fundamentally an economic alignment problem: if advanced AI (including AGI) is developed within a growth-first, market-driven economy, incentives will push AI toward outcomes that increase social, environmental, and even existential risk. Post-growth economic frameworks (degrowth, Doughnut, wellbeing economy, steady-state) offer concrete concepts and policy levers that can reduce those risks and should shape AI R&D, deployment, and governance.

Key Points

• Economic alignment problem: AI alignment requires aligning the underlying economic system with human wellbeing and planetary boundaries; building AI inside a growth-maximising system will bias AI toward growth-centric goals.
• Five facets of the problem:
• Exponential capability growth — multiple AI metrics (training compute, minimal model size, task difficulty solved) have shown exponential improvement; the authors estimate task-difficulty doubling ≈ every 7 months (implying >200% annual capability growth).
• Unconstrained development & accelerationism — dominant norms (effective accelerationism, market competition) promote fast, unconstrained AI development that neglects social innovation and safety.
• Growth–environment tension — explosive AI-driven GDP growth risks grossly increasing resource use and emissions given rebound effects; efficiency gains alone likely won’t avoid environmental harm.
• Inequality & employment — AI may concentrate gains (profits, power), displace labour unevenly, and amplify Global North / Global South asymmetries; post-growth redistributive measures (wealth taxes, UBI) are proposed to manage distributional risks.
• Meaning, identity & purpose — newer automation exposure data show AI threatens jobs with high satisfaction/meaning; broad labour displacement could have serious social-psychological effects.
• Policy / design recommendations (post-growth AI roadmap, high-level):
  • Replace optimisation-for-growth objectives with satisficing (sufficient wellbeing) and adopt Doughnut-style social + planetary boundary constraints.
  • Apply resource caps and policies to curb systemic rebound.
  • Treat AI as a commons: new ownership/governance forms to decentralise control and limit concentration.
  • Prioritise tool-like, autonomy-enhancing systems over agentic, goal-setting AI.
  • Use selective, democratically governed adoption of AI; integrate social innovation with technological change.
  • Prepare new economic theories and models for AGI that incorporate non-growth goals and planetary limits.
• Empirical and attitudinal signals: surveys of AI researchers show a substantial minority attach non-trivial probabilities to catastrophic outcomes; forecasting platforms and model-performance trends indicate timelines for AGI are moving earlier in some forecasts.

Data & Methods

• Mixed-methods approach combining:
  • Literature review across AI safety, social science, and ecological economics.
  • Empirical trend-fitting: exponential fits to historical AI indicators (training compute, minimal model size, task difficulty) to estimate doubling times (authors report: training compute doubling every ~5–6 months, minimal viable model size halving every ~3 months, task-difficulty doubling every ~7 months). Projections to 2030 are extrapolated from 2019–2025 data (authors caution real-world applicability is uncertain).
  • Scenario construction for GDP per capita (1950–2024 historical baseline + five alternative 2025–2050 trajectories derived from existing studies ranging from “modest growth” to “intelligence explosion” and “human extinction”).
  • Occupational analysis: compared occupational automation exposure datasets (Frey & Osborne, 2013; Eloundou et al., 2023) against job-satisfaction / meaning measures from Payscale to assess how exposure relates to subjective job meaning over time (finding the negative association in 2013 weakens or reverses with 2023 automation estimates).
  • Use of surveys and forecasting platform outputs (e.g., computer scientist surveys on AGI timelines; Metaculus median forecast shifts) to document expectation heterogeneity.
• Documentation: figures (e.g., capability-doubling plots, GDP scenarios, job-satisfaction scatterplots) and supplementary information detail data sources and fitting procedures.
• Limitations acknowledged by authors: short historical windows for some metrics, large uncertainty in mapping benchmark trends to real-world task capability, and diversity of estimates across research communities.

Implications for AI Economics

• Model assumptions must change: macroeconomic and AI-impact models should incorporate planetary limits, non-growth wellbeing objectives, rebound effects, and distributional outcomes rather than assuming unconstrained productivity-led growth.
• Policy design must internalise ecological constraints: carbon/resource caps, Doughnut-guided targets, and mechanisms to prevent scale-driven environmental overshoot are essential complements to AI regulation.
• Incentives & ownership matter: economic incentives driving rapid, competitive AI deployment can misalign safety and fairness goals; treating AI as a commons and promoting alternative business models (public, cooperative, licensing restrictions) can decentralise power and reduce risky concentration.
• Labour & social policy: prepare redistributive and social-support policies (e.g., progressive taxation, UBI, shorter working time, re-skilling oriented to meaningful work) that account for the possibility of pervasive cognitive automation.
• Research agenda: economists should develop post-growth–compatible models of AGI effects (endogenous innovation under planetary constraints; demand saturation; ecological macroeconomics with AI-driven productivity) and empirically test rebound and resource implications of large-scale AI deployment.
• Governance implications: regulatory frameworks should prioritize limiting agentic, self-directed AI deployment; require impact assessments that include ecological and social system coupling; and enable democratic oversight and participatory governance of AI development.
• Uncertainty management: given wide scenario dispersion (from modest gains to explosive growth or catastrophe), policy and economic modelling need to be robust to tail risks and to asymmetric, non-linear outcomes (not just point forecasts).

Caveats: many quantitative projections (doubling times, capability multipliers, AGI timelines) are highly uncertain and sensitive to short-run data and modelling choices; the paper blends normative post-growth prescriptions with empirical trend analysis, so follow-up empirical work is needed to operationalise proposed policies and to test key assumptions (e.g., real-world mapping of benchmark performance to economic impact).