Some legal uncertainty is a feature, not a flaw: the paper argues that leaving room for interpretation in high-level AI laws enables the boundary negotiations needed for adaptive governance. Well-designed technical sandboxes and boundary artifacts translate abstract legal requirements into operational checks, but enforcing premature legal closure risks stifling the learning that makes regulation effective.

Main Finding

Legal uncertainty in high-level AI regulation (exemplified by the EU AI Act) cannot — and should not — be fully eliminated ex ante. Within reasonable bounds, uncertainty is an enabling condition for regulatory learning: it creates a space where socio-technical meaning is negotiated among stakeholders. Technical sandboxes and other “boundary negotiating artifacts” (BNAs) are crucial mechanisms for translating abstract legal requirements into operational, verifiable technical practices. Prematurely fixing technical rules risks closing off the adaptive learning processes the regulation is designed to elicit.

Key Points

• Problem framed: High-level, principle-based AI regulation (e.g., the AI Act) creates downstream legal uncertainty as stakeholders struggle to operationalize abstract obligations into concrete tests and technical requirements.
• Analytic lens: The paper uses Coleman’s “bathtub” model (macro–meso–micro levels) plus concepts from boundary-object theory to explain how regulatory pressure and learning propagate across levels.
• Three abstraction levels where learning occurs:
  • Legislative (macro): law and high-level obligations (AI Act).
  • Regulatory/standards (meso): harmonised standards, implementing/delegated acts, codes of practice.
  • Technical (micro): software requirements, assessments, tests implemented by developers and assessors.
• Actor mapping: The authors map AI Act actors into macro (EU Commission, AI Office), meso (AI Board, standardisation bodies, Advisory Forum, Competent Authorities, Notified Bodies), and micro (developers, providers, deployers, testing hubs). Meso actors perform translation (meso–micro) and aggregation (micro–meso–macro).
• Role of bounded uncertainty: Some ambiguity in legal texts is useful — it forces multi-stakeholder negotiation about socio-technical meaning and creates learning signals that can inform later specification. Total ex-ante certainty would foreclose necessary adaptation.
• Boundary Negotiating Artifacts (BNAs): Extends the boundary-object idea to artifacts (technical platforms, sandbox frameworks, shared datasets, benchmarks, protocols) that mediate negotiations between legal requirements and technical practice. BNAs enable stakeholder coordination by making trade-offs, metrics, and contextual assumptions visible and negotiable.
• Technical sandboxes (AI Regulatory Sandboxes / AIRSes) are highlighted as prime BNAs. The paper specifies properties these infrastructures must have to be effective mediators (see Data & Methods / Implications).
• Risk of premature closure: While legal certainty is a long-term aim, prematurely locking in technical specifications (standards, tests) undermines adaptive governance and may produce brittle compliance regimes.

Data & Methods

• Methodological approach: conceptual and policy analysis rather than empirical measurement. The paper synthesises legal texts (AI Act provisions), policy instruments (NLF mechanisms), and academic literatures on regulatory learning, standardisation, and socio-technical systems.
• Theoretical tools:
  • Coleman’s bathtub model to represent macro–meso–micro causal flows of regulatory pressure and learning.
  • Boundary objects and a novel extension, Boundary Negotiating Artifacts (BNAs), to explain translation processes across stakeholder communities.
• Actor & mechanism mapping: functional classification of actors (Enforcers, Meso–Micro advisors, Meso–Macro advisors, Standardisers) and tracing of learning flows in the AI Act governance architecture (Articles and institutional roles).
• Comparative and normative reasoning: the authors situate the AI Act relative to other jurisdictions (US, China) to highlight differences in top-down vs. sectoral or industrial-policy approaches, and to argue for the value of institutionalised learning instruments (codes, sandboxes, testing environments).
• No original quantitative data analysis; evidence comes from policy documents, existing literature, and reasoned argumentation about plausible dynamics of implementation and learning.

Implications for AI Economics

• Investment and innovation decisions:
  • Regulatory uncertainty increases option value of delay for firms and investors; bounded uncertainty supported by sandboxes can reduce paralysis by providing experimentally validated compliance paths.
  • Adaptive regulatory processes create dynamic complementarities: firms that engage with BNAs/sandboxes early may shape de facto standards, capturing first-mover advantages or creating switching costs.
• Compliance costs and market structure:
  • Translation of abstract obligations into technical tests will shape compliance costs heterogeneously across firms; SMEs may be disproportionately affected unless sandbox/SME support (explicit in the AI Act) reduces entry barriers.
  • Standardisation trajectories emerging from sandboxes can create network effects and lock-in; the meso actors (standardisers, advisory forums) play an outsized role in setting downstream economic equilibria.
• Uncertainty as a policy instrument:
  • Regulators should treat uncertainty as a managed input rather than an externality — calibrating how much ambiguity to keep and how quickly to crystallise technical rules based on aggregated micro-level evidence. This has implications for dynamic models of firm behavior under regulatory learning (real options, investment under ambiguity).
• Information externalities and coordination:
  • BNAs and sandboxes reduce information asymmetries (between firms, auditors, and regulators) and mitigate coordination failures on testing protocols and compliance technologies. This is welfare-improving relative to uncoordinated multiplicity of private solutions, but raises questions about who governs the BNA (public vs. private control).
• Standardisation, competition, and capture risks:
  • Because harmonised standards and sandbox outcomes can shape market norms, there is potential for regulatory capture or capture-like path dependence if participation is skewed toward incumbents. Economic policy must ensure broad, inclusive participation to avoid reinforcing market concentration.
• Externalities and systemic risk:
  • The meso→macro feedback loop is essential to detect systemic risks (e.g., failures of widely used models). Effective aggregation of micro evidence determines the ability to internalize externalities at the macro level (amend laws, issue implementing acts). Poorly designed aggregation raises the likelihood of mispriced systemic risk.
• Policy design recommendations for economic efficiency:
  • Invest in BNAs/sandboxes designed to be interoperable, transparent, and accessible to SMEs to lower compliance cost dispersion and promote competition.
  • Build explicit processes for time-bound reduction of uncertainty: safeguard the value of adaptive learning but commit to predictable rule crystallisation once sufficient evidence is aggregated (limits the duration of regulatory option-value).
  • Encourage data-sharing and standardized reporting from sandboxes to reduce measurement costs and produce actionable macro-level signals for timely policy updates.
  • Design governance of sandboxes to limit capture (representation rules, public oversight, open logs), since sandbox outputs will have disproportionate economic effects through shaping standards.

Bottom line: For AI economics, the paper reframes regulatory uncertainty not only as a cost (investment hesitation, compliance expense) but also as an instrument enabling socially valuable learning. The economic challenge is to design meso-level BNAs (especially sandboxes) and incentives so as to capture the upside of learning while limiting the downside of prolonged ambiguity, capture, and uneven compliance burdens.