Who owns the inputs owns AI: scaling laws in machine learning have made compute and data the decisive assets, driving costly concentration in a few hands; market-based property over data is inadequate, so democratic control over AI objectives is needed.

Main Finding

Control over what AI systems optimize for—i.e., their objective functions—is ultimately determined by who controls the means of prediction (data, compute, expertise, energy). The production function that maps data and compute into predictive performance (the “scaling laws”) explains why industry incentives have driven a costly scramble for scale, producing concentration of power. Market-based governance (e.g., individual property rights over data) cannot adequately address the resulting harms because machine learning intrinsically involves data externalities and maintained platform network effects. The paper argues for democratic control of the means of prediction (e.g., sortition, liquid democracy) so those affected by algorithmic decisions have a say over AI objectives.

Key Points

• AI as optimization: Modern AI systems are best seen as systems constructed to maximize measurable objectives (rewards). Social conflicts about AI are largely conflicts over which objectives are chosen and who chooses them, not merely misalignment or optimization failure.
• Means of prediction: The principal inputs for contemporary AI are data, compute, expertise, and energy. Control over these inputs gives practical control over AI objectives and deployment.
• Production function / scaling laws:
  • Theory: Bias–variance tradeoffs imply that more data reduces variance (overfitting) while more compute/model complexity reduces bias (underfitting). Optimal performance requires trading these off; model size, training steps and data interact.
  • Empirics: Industry studies (e.g., Kaplan et al. 2020; Hoffmann et al. 2022; Muennighoff et al. 2025) document empirical scaling laws: predictive performance improves with data and compute but with diminishing marginal returns and a nonzero irreducible loss.
  • Policy consequence: Because performance can be improved by scaling data and compute, firms have incentives to invest massively in both, leading to very costly infrastructure and concentration among a few players.
• Limits to unbounded progress: Data scarcity (e.g., finite available text) constrains future performance gains for LLMs; in many domains the amount and nature of available data fundamentally limit how well statistical learning can perform (e.g., genomics, macroeconomics, some real-world tasks).
• Market failure and externalities:
  • Machine learning is characterized by data externalities: individual data rights/transactions do not internalize social spillovers from aggregation, reuse, and model training.
  • Platform network effects and artificial maintenance of dominance further weaken competitive solutions.
  • Private property in data thus cannot be expected to align AI objectives with broad social welfare.
• Governance proposal: Democratic control over the means of prediction—institutions that give affected stakeholders a voice in choosing objectives (examples: sortition panels, liquid democracy, public deliberative processes)—is recommended to rebalance power and specify societally appropriate optimization objectives.

Data & Methods

• Methods: Conceptual and theoretical analysis combining statistical learning theory (bias–variance, model complexity, early stopping) with a review and synthesis of empirical industry studies on scaling laws for deep learning and LLMs. The paper also uses political-economy reasoning about externalities, market structure, and governance.
• Empirical evidence cited:
  • Industry scaling-law studies (Kaplan et al. 2020; Hoffmann et al. 2022; Muennighoff et al. 2025) estimating functional forms L(N,D) that relate model size N, data size D (and compute C) to loss/performance.
  • Descriptive industry facts: exponential growth in model parameters and training tokens since ~2020, large fundraising and compute expenditures (e.g., OpenAI’s financing), and environmental/resource footprints of data centers.
  • Case examples illustrating socially consequential use of prediction systems (e.g., algorithmic management, targeted advertising, predictive policing, wartime targeting).
• Nature of contribution: The paper does not report original experimental or econometric estimation; it synthesizes theoretical results and existing empirical findings to draw normative and institutional conclusions.

Implications for AI Economics

• Distributional concentration: Scaling-law-driven investment patterns imply large fixed-cost, scale-sensitive returns that favor concentrated ownership of the means of prediction. Antitrust and industrial policy should account for the scale-dependent nature of AI production.
• Rethinking data policy: Standard market solutions (tradeable individual data rights) are unlikely to internalize social value/harms because data generates strong aggregate externalities. Policy should emphasize collective mechanisms (data trusts, public datasets, regulation of reuse) and account for spillovers from model deployment.
• Public investment and infrastructure: To reduce private concentration and enable socially beneficial uses, public actors might supply or subsidize common-pool inputs—curated public datasets, shared compute infrastructure, or regulated access regimes—while pairing such investments with democratic governance over objectives.
• Governance of objectives: Economic analysis of AI must shift part of its focus from efficiency of algorithmic performance to institutional mechanisms that determine objective choice. Mechanisms like sortition, participatory decision-making, or liquid-democracy-style processes could be evaluated as tools for aligning AI objectives with societal welfare.
• Environmental and labor externalities: The scramble for compute has environmental and human-cost consequences (energy and water use for data centers; precarious human labeling work). AI economic policy should internalize these externalities via taxation, regulation, labor protections, and environmental constraints.
• Research & evaluation priorities: For economists studying AI, the production-function perspective suggests research priorities: measuring returns to additional data vs. compute across domains; quantifying externalities from data aggregation and model deployment; evaluating institutional designs (data governance, public infrastructure, democratic processes) for shaping objective choice.

Limitations / caveats - The paper synthesizes existing theory and industry evidence rather than producing new empirical estimates. Some quantitative aspects (future attainable gains, exact shapes of scaling laws across domains) remain uncertain and domain-dependent. - Normative proposals (sortition, liquid democracy) are sketched at a conceptual level; operational design, feasibility, and potential unintended consequences require further empirical and institutional work.