Measuring the openness of AI foundation models: competition and policy implications
Citation: Thibault Schrepel, Jason Potts (2025/03/05) Measuring the openness of AI foundation models: competition and policy implications.
DOI (original publisher): 10.1080/13600834.2025.2461953
Semantic Scholar (metadata): 10.1080/13600834.2025.2461953
Sci-Hub (fulltext): 10.1080/13600834.2025.2461953
Internet Archive Scholar (search for fulltext): Measuring the openness of AI foundation models: competition and policy implications
Wikidata (metadata): Q135645204
Download: https://www.tandfonline.com/doi/full/10.1080/13600834.2025.2461953
Tagged:
Summary
Most “openness” yardsticks for foundation models focus on technical artifacts; this paper instead measures license-level openness as a driver of the AI innovation commons. It defines an 18-variable index, clustered around three economic problems a license must solve for commons to work—knowledge (e.g., accessibility, transparency, documentation), implicit contracting (e.g., contribution policies, exit rights, anti-opportunism), and collective-action governance (e.g., access & use rights, participatory governance, interoperability). The authors hand-score 11 prominent models (GPT-4, Gemini Ultra, Llama 3, Midjourney V6, Claude 3, xAI, Mistral 8×7B, BLOOM, Cohere Aya, Cohere Command R, Falcon 180B) across these variables (0/1/2 per variable; 198 scored items total) and release the evidence table. Headline findings: today’s models score highest on documentation/support, collaboration platforms, derivative-works and lowest on exit rights, costs of maintenance, participatory governance, credit/revenue sharing; openness rankings diverge from technical-only indices, with Cohere Aya and BLOOM-560M leading on “knowledge accessibility,” while GPT-4, Gemini, Midjourney V6, Command R, Claude 3 score 0/2 on knowledge accessibility and (except Command R) on interoperability. Aggregate scores are 64/132 for knowledge vs. 35/132 for each of implicit contracting and governance—pinpointing where “open” claims break down.
Theoretical and Practical Relevance
Turns “open vs. closed” into a license-design checklist that antitrust and policy actors can act on. For policymakers, it proposes a minimum audit (review how existing rules affect the 18 variables) and a maximal, three-pillar program: (1) legal exemptions for genuinely open systems (not just technical definitions), (2) economic measures (tax incentives, procurement preference, funding, institutional support) to tilt toward open models, and (3) technical support (transparency duties when public funds are used; open data repositories; interoperability standards). For enforcers, it argues the closed end of the spectrum deserves priority scrutiny (licenses that restrict forking, interoperability, or access), whereas more-open licenses reduce anticompetitive leverage by enabling scrutiny, forking, and entry. The authors publish the 198-item evidence sheet so others can reuse the rubric and recompute weights for procurement, exemptions, or oversight.