AI openness is not all or nothing; regulators should govern openness component-by-component. A 'differential openness' framework advises treating compute, data, models and labor as separately configurable levers so policymakers can balance innovation, democratized access, oversight and national security without resorting to blunt open/closed binaries.

Main Finding

AI openness is not a binary choice. Instead of “open” versus “closed,” AI systems are best understood as stacks of discrete components (compute, models, weights, data, training code, deployment interfaces, labor, etc.), each of which lies on its own spectrum of openness. Policymaking should adopt a “differential openness” approach that configures the relative openness of specific components to balance competing goals—innovation, oversight, safety, competition, and national security—rather than applying one-size-fits-all openness rules.

Key Points

• The open/closed framing inherited from software is misleading for AI because AI is a composite technology with multiple interdependent components controlled by diverse actors.
• Each component’s openness produces distinct benefits and harms:
  • Openness can enable independent oversight, inclusive collaboration, and faster diffusion of innovation.
  • Openness can also accelerate malicious actors, lower entry costs for strategic competitors (including adversaries), and help incumbent firms entrench power.
• Trade-offs are both within components (e.g., open weights speed replication but increase misuse risk) and across components (e.g., closed compute with open models shifts commercialization incentives).
• Effective governance requires an analytical taxonomy—“differential openness”—that (1) identifies the stack’s components, (2) maps openness spectra for each, and (3) evaluates how configurations align with policy objectives.
• Law and regulation should be designed to target component-level openness (e.g., conditional licensing, access controls, targeted disclosure requirements) rather than blanket mandates for full openness or secrecy.
• The article advances a research agenda to operationalize the taxonomy: clarifying legal tools, modeling economic impacts, measuring component openness empirically, and evaluating international coordination needs.

Data & Methods

• Conceptual taxonomy: Decomposes AI into core components (examples: hardware/compute, training data, model architectures, trained weights/checkpoints, training code and recipes, evaluation datasets and benchmarks, deployment APIs, human labeling and labor processes, maintenance/updates). For each component, identifies a spectrum of openness (from fully public to fully proprietary) and typical stakeholders.
• Analytical trade-off mapping: For each component, maps the primary benefits (e.g., innovation diffusion, transparency, auditing) against primary harms (e.g., misuse risk, national-security leakage, rent extraction), showing intra- and inter-component interactions.
• Illustrative case analysis: Uses examples (e.g., adversarial rapid catch-up like “DeepSeek-R1”) to show how component openness choices affect international competitiveness and risk diffusion.
• Policy and legal analysis: Surveys regulatory instruments (disclosure rules, licensing, export controls, antitrust interventions, standards-setting, conditional funding or procurement) and evaluates how they can be tuned to adjust component-level openness.
• Research agenda methodology: Proposes mixed methods—economic modeling, empirical measurement of openness and effects, legal doctrinal work, and simulation/incident analysis—to operationalize the taxonomy and inform policy calibration.

(Note: the article is primarily conceptual and normative rather than an empirical study; the “data” are illustrative cases and institutional examples used to motivate the taxonomy and trade-off analysis.)

Implications for AI Economics

• Market structure and entry:
  • Differential openness affects barriers to entry. Opening weights or datasets reduces upfront costs for entrants; closing compute or large-scale training pipelines preserves incumbents’ advantages.
  • Regulators seeking competition should consider which component to open (or mandate access to) to lower effective entry costs without unduly increasing risks.
• Innovation incentives and investment:
  • Openness in middle layers (e.g., model architectures, evaluation metrics) promotes cumulative innovation and shared standards, while proprietary control over deployment and APIs preserves commercialization incentives.
  • Policy must balance IP and public-good provision: subsidies, conditional funding, or prize structures could steer which components are made public.
• Knowledge diffusion vs. rent extraction:
  • Selective openness can maximize social learning (research transparency, reproducibility) while limiting private rent extraction by incumbents; conversely, blanket openness may erode returns on risky investments, undermining long-horizon R&D.
• Safety and externalities:
  • Component-level openness choices determine the diffusion speed of both benign innovations and harmful capabilities. Economic models of externalities should treat openness as an endogenous policy lever influencing the rate and distribution of capability diffusion.
• International competition and security:
  • Openness regimes interact with geopolitics: releasing certain components internationally can accelerate competitors’ capabilities. Economists should model strategic disclosure as part of national industrial policy and deterrence calculus.
• Policy design and empirical priorities:
  • Empirical work is needed to measure component-specific impacts: how does openness of weights vs. training data vs. deployment APIs affect startup formation, model abuse incidents, pricing, and welfare?
  • Recommended empirical approaches: constructing openness indices by component, difference-in-differences around policy changes or corporate disclosure shifts, causal inference on entry and innovation outcomes, and structural models embedding openness as a choice variable.
• Regulatory instruments to consider:
  • Targeted disclosure mandates (e.g., model cards, provenance for datasets), conditional public funding (tie openness of some components to grant terms), controlled-access mechanisms (licenses, vetted researcher access), and antitrust remedies focused on component-level bottlenecks (e.g., compute markets).
• Research agenda for AI economics:
  • Quantify trade-offs: estimate the marginal social benefit vs. harm of opening each component.
  • Model strategic behavior: include firms’ disclosure choices, competitor responses, and state actors in dynamic economic models.
  • Evaluate policy tools: simulate outcomes under differential openness regimes to inform cost–benefit and distributional analyses.

Overall, treating openness as a configurable policy variable across AI’s components enables more precise economic interventions that can better align innovation, safety, competition, and national-security objectives.