Hosted AI services keep firms and regulators within reach, while open-weight redistribution scatters model lineage across repositories, modifiers and deployers and creates enforcement gaps; policymakers should require portable provenance and assignable obligations to preserve accountability without squashing downstream innovation.

Main Finding

Branches of the same generative models (hosted/API vs open-weight redistributed variants) become differentially reachable by regulation as they diverge in distribution relationships and organisational control. Provider-controlled services retain clear, addressable control points; open-weight redistribution preserves model lineage across many actors and hosts, creating regulatory gaps and requiring governance that assigns differentiated responsibilities and portable records. The article formalises this as a BOAP-informed "branching regulatory biography" and the concept of "regulatory reachability" to show which instruments can target which actors through which control points at a given time.

Key Points

• Methodological innovation: applies Biographies of Artefacts and Practices (BOAP) to model lineages, producing "generative AI in context" as an analytic extension for technical continuity with diverging organisational distribution.
• Comparative case: matched documentary comparison of GPT-4 (hosted/API branches) and DeepSeek-R1 (hosted/API plus redistributed open-weight branches).
• Empirical scope: 132 documentary observations, reconstruction of five cross-site configuration moments spanning 2022–2025, plus a bounded mid‑2026 status check across the EU, US and China.
• Early governance modes: reliance on inherited controls — platform rules, algorithmic-service arrangements, data and compute constraints — that map well onto provider-driven services.
• Emergent governance modes: development of provider-, model- and output-oriented regulatory pathways as control points shift and regulatory instruments adapt.
• Divergence effects: provider-controlled services kept identifiable actors and enforceable control points; open-weight redistribution fractured control, allowing model lineage to persist across repositories, modifiers, hosts and deployers outside original developer control.
• Organisational dynamics: firms and other actors reconfigured later control points through contractual, technical and operational responses, reshaping regulatory reachability over time.
• Governance prescription: distributed-model governance needs differentiated responsibilities, portable records, and documentation treated as supporting evidence for accountable legal actors (not a substitute for legal obligations).

Data & Methods

• Approach: documentary biography using BOAP principles — tracing artefacts (models), practices (distribution and deployment), and organisational settings over time.
• Materials: 132 documentary observations (public documents, regulatory filings, platform policies, repository traces, technical release notes, press statements) collected across jurisdictions and sites.
• Comparative design: matched comparison of two model releases’ branches:
  • GPT-4: analysis concentrated on hosted and API-delivered branches under provider control.
  • DeepSeek-R1: included hosted/API branches plus open-weight repository releases, derivatives and public-serving routes enabling redistribution.
• Temporal reconstruction: identification of five cross-site configuration moments (2022–2025) that mark changes in configuration and control, with a bounded status check mid-2026 to capture subsequent shifts.
• Analytical concepts: "branching regulatory biography" to map divergence across branches; "regulatory reachability" to determine which instruments can plausibly target which actors via which control points at particular times.

Implications for AI Economics

• Market structure and competition
  • Redistribution of open weights lowers barriers for downstream entrants and forks, increasing competition but also fragmenting control and accountability.
  • Provider-controlled hosted/API models preserve concentrated control and capture (pricing, data capture, upgrade paths), affecting platform power dynamics.
• Regulation and compliance costs
  • Differentiated regulatory reachability implies heterogeneous compliance costs across branches: hosted services face direct provider obligations; redistributed branches push compliance burdens downstream to deployers, hosts and modifiers.
  • Regulators may need tiered or assignable obligations (e.g., provenance, documentation, liability rules) to avoid enforcement gaps or duplication.
• Liability, externalities and enforcement
  • Open-weight diffusion creates weakly attributable harms and higher enforcement/friction costs; traceability and portable records become economically valuable public goods.
  • Treating documentation as supportive (not substitutive) of legal accountability changes incentive structures for investment in auditability and provenance tracking.
• Supply chains and transaction costs
  • Persistent model lineage across repositories increases the complexity of contractual and monitoring relationships, raising transaction costs for downstream firms (due diligence, indemnities).
  • Portable, standardised records reduce informational frictions and can lower ex-post dispute costs, but impose upfront documentation costs.
• Innovation and diffusion dynamics
  • Open-weight redistribution accelerates innovation and specialization (modifiers, fine-tuners, niche deployers) but can create negative externalities (misuse, reputational harms) that depress social returns absent governance.
  • Provider strategies (locking capabilities behind APIs, contractual restrictions) are strategic responses that shape incentives for investment and diffusion.
• Policy design considerations for economic efficiency
  • Target instruments to reachable control points per branch (provider obligations for hosted; provenance, deployer liability, and platform-host rules for redistributed models).
  • Promote portable provenance standards to reduce verification costs and enable assignment of responsibilities across the value chain.
  • Balance incentives: avoid regulatory approaches that either ossify incumbent control (stifling downstream innovation) or create unbounded externalities via unfettered redistribution.
• Research priorities
  • Quantify welfare trade-offs: open redistribution vs centralized hosting (competition, innovation, risk externalities).
  • Measure costs of documentation and traceability infrastructure relative to enforcement benefits.
  • Model cross-jurisdictional regulatory arbitrage and its effects on investment and branching behaviour.

Overall, the article argues that economic analysis and policy must treat generative-models as branching artefacts whose regulatory reachability varies by branch and time — requiring differentiated, assignable obligations and interoperable, portable records to align incentives, reduce enforcement costs, and manage externalities without unduly constraining beneficial diffusion.