Research / AI Governance and Public AI
AI Governance and Public AI
Elkins’ governance work examines the institutional conditions under which AI can serve public rather than merely commercial or geopolitical ends. It includes comparative AI regulation, open-source and open-weight model risk, public AI infrastructure, cultural data governance, standards participation, and the policy challenges of transparency, access, oversight, and accountability.
The work moves between standards bodies, technical-policy venues, public-interest AI organizations, and comparative governance scholarship. It treats AI governance as an institutional problem: who can inspect, use, adapt, regulate, localize, and benefit from AI systems once they become public infrastructure.
Selected work and projects
NIST AI Safety Institute Consortium / CAISI
Co-leading the MLA-sponsored team, 2024–present
Elkins co-leads the team representing the Modern Language Association at CAISI, the consortium launched by the National Institute of Standards and Technology to develop standards and guidelines for trustworthy and safe AI systems. The five-member MLA-sponsored team brings expertise in language, writing, interpretation, ethics, and humanistic inquiry to evaluations of model behavior in complex linguistic and ethical scenarios.
The work places humanities-based model evaluation within a national AI standards process, connecting questions of ambiguity, framing, persuasion, value, and linguistic edge cases to the practical assessment of trustworthy AI systems.
AI, IP & Culture Repository Co-Design Process
Elkins and Chun, on behalf of the Human-Centered AI Lab, 2026
Elkins contributed with Jon Chun to the AI, IP & Culture Repository co-design process, a civil society initiative developed in support of UNESCO’s work on AI ethics, intellectual property, cultural rights, and digital cultural sovereignty. The process was coordinated through the Intellectual Property and Culture working group within UNESCO’s CSO Network in AI Ethics.
Their contribution, “Provenance Infrastructure as a Safeguard for Cultural Commons in the Age of Generative AI,” was included among the report’s operational proposals offering practical solutions to AI-related challenges for cultural rights. The proposal advances technical mechanisms for tracking cultural-data provenance and supporting rights-aware AI training practices, with the goal of preserving consent, attribution, cultural context, and community authority as cultural materials move through AI pipelines.
The work extends Elkins’ governance research from model evaluation and public AI infrastructure into cultural data governance: how archives, public-domain materials, and cultural commons can remain accountable to the communities and histories they come from rather than being absorbed into opaque proprietary AI systems.
Comparative Global AI Regulation
Chun, Schroeder de Witt, and Elkins, arXiv 2410.21279, 2024
This paper develops a structured comparative framework for understanding three major approaches to AI governance: the European Union’s risk-based regulatory model, the United States’ decentralized and sectoral approach, and China’s state-directed but flexible system of oversight. It examines how different institutional arrangements produce different tradeoffs among safety, innovation, enforcement, rights, and state power.
The paper was coauthored with Christian Schroeder de Witt, Principal Investigator of the Oxford Witt Lab for Trust in AI at the University of Oxford, whose work focuses on multi-agent security and trustworthy multi-agent AI systems. The collaboration situates the governance paper within a technical-policy network spanning humanities, AI safety, multi-agent systems, and institutional regulation.
The framework has been used in later governance scholarship, including Olugbade’s “In search of a global governance mechanism for Artificial Intelligence (AI): a collective action perspective,” which cites the paper across its US, China, EU, and synthesis sections. It is also engaged by Yew, Marino, and Venkatasubramanian’s FAccT 2025 work on AI policy red-teaming and by Ilcic, Fuentes, and Lawler in Frontiers in AI as a reference framework for AI governance analysis.
Near- to Mid-Term Risks and Opportunities of Open-Source Generative AI
Eiras, Petrov, Vidgen, Schroeder de Witt, Pizzati, Elkins, et al., ICML 2024 oral
This ICML oral paper intervenes in the debate over open-source and open-weight generative AI by distinguishing among different forms of openness and their associated risks, benefits, and governance needs. It applies and operationalizes an AI openness taxonomy across forty large language models, extending Irene Solaiman’s gradient framework with risk-mitigation strategies and a temporal view of model development.
The author group brought together researchers from Oxford’s Foerster Lab for AI Research and Torr Vision Group, UC Berkeley, Notre Dame, Kenyon, Berklee, the Luxembourg Institute of Science and Technology, the University of Luxembourg, the University of Bath, ITESM, the National University Philippines, and other international institutions.
The paper has become part of subsequent work on AI openness and open-weight risk. Paris, Moon, and Guo name it at FAccT 2025 as one of three major openness frameworks, alongside the Model Openness Framework and Liesenfeld and Dingemanse’s Community-Driven Assessment Framework. Casper, O’Brien, Longpre, and coauthors cite it in “Open Technical Problems in Open-Weight AI Model Risk Management,” a 2026 TMLR consensus paper authored by researchers from MIT CSAIL, Stanford, Mila, Carnegie Mellon, Princeton, the UK AI Security Institute, the Center for AI Safety, and Hugging Face.
If Open Source Is to Win, It Must Go Public
Tan, Vincent, Elkins, and Sahlgren, Public AI / ICML 2025 CODEML Spotlight
This work argues that open weights alone do not democratize AI. Unlike ordinary software, AI systems require substantial activation resources: compute, post-training, deployment, localization, and oversight. Without public institutions and public infrastructure, open models can remain formally available but practically unequal.
The authors bring together public-interest computing, platform governance, philosophy, and AI science, including Joshua Tan of the Public AI Network and Metagov, Nicholas Vincent of Simon Fraser University, and Magnus Sahlgren of AI Sweden. The project shifts the open-source AI debate from model release to public capacity, calling for infrastructure that allows open models to function as public goods: usable, adaptable, locally governable, and accountable beyond the small set of actors with sufficient compute, technical staff, and deployment resources.