Exploring Theory, Policy and Data on AI Model Releases
Clear definitions are essential for legal and policy decision-making concerning complex technologies. By grounding Open Source AI in transparent, verifiable, standardized criteria, this work supports more consistent interpretation across legal, policy, and technical domains. The project provides evidence-based recommendations for engagement with policymakers and contributes to ongoing efforts to support the adoption of the Open Source AI Definition (OSAID) in practice and policy.
Making Sense of Opennes with Theory
Open Source Definitions as Boundary Objects
The theory of boundary objects refers to things or concepts that are used in different ways by different communities, which allow them to collaborate. Opennes in AI can be treated as a boundary object that serves as the operational glue between people and practices
Seeing the OSAID as a boundary object explains how it can be equally useful to law and compliance professionals, as to developers, two communities with widely differing practices, concerns, and knowledge, but that benefit from using a shared concept.
However, in the hotly contested AI arena, disagreements about what should count as “open” AI prevail. Importantly, because boundary objects are flexible and interpretable, they can be challenged and co-opted, and misinterpreted. Indeed, we don’t agree on a standard or definition of openness in AI.
Infrastructural Inversion
The theory of infrastructural inversion provides a helpful set of tools for dissecting how standards and categories are established. Conceptualized as a theory of analysis, infrastructural inversion focuses on the interdependence between technologies, their design, and the habits by which standards and categories become accepted and normalized.
According to Bowler and Starr, “good usable systems disappear almost by definition.” This suggests that useful standards are normalized and widely recognized, if not necessarily widely understood. Standards become normalized as the categories imposed by a standard are employed across different groups of people, organizations and examples. Consequently, standards evolve recursively––they need to be useful to be accepted, and they become more useful as they are more widely accepted.
Competing Notions of Openness in AI
In understanding downstream policy implications, it’s important to note how definitions of openness differ. While Open Source software (OSS) has existed for decades, the concept of “openness” in AI systems is nascent and contested. At least four regimes for defining openness materialized in ways that echo interests in promoting transparency and broad participation. Different notions of openness reflect different levels and kinds of disclosures.
For many AI practitioners today, open source AI generally refers to open weights and publicly available weights models available through repositories such as GitHub and Hugging Face. Open weight models provide access to a combination of code and weights in ways that allow some use and modification of AI while levying usage restrictions, for example, in certain locations or for commercial use.
While open access to model weights is a prerequisite for openness, in practice, open weight models appear to uphold a philosophy of openness yet fall short of providing the mechanisms to guarantee that AI releases comply with the four freedoms. Ideally, definitions should provide the right level of specificity and generality that reduces ‘unexpectedness’ and increases the possibility of applying to AI systems as they continue to evolve.
For the OSI and the OSAID’s binary framing, a system is either open or not. Categorization hinges on which materials are made available and how each is published. As with traditional OSS, this means providing artifacts in the “preferred form to make modifications,” under an OSI-approved license where applicable.
The Model Openness Framework (MOF) defines openness along three “classes”, each requiring a different number of artifacts and disclosures, ranging from code to academic papers and model cards detailing training and testing procedures
U.S. Policy Landscape
Existing legal definitions of AI and Open Source AI have immense potential for divergence, which will likely translate to diverging regulatory approaches
While the number of AI and LLM bills introduced is on an upward trend with no sign of slowing, Open Source has yet to take a central role despite more frequent mentions. However, executive orders, agency directives, and emerging legislation at the state level signal concerns with the safety of open models while at the same time, emphasizing the need to promote access, innovation, decentralization, and competition in the AI space.
Notably, no existing policies at the state or federal level explicitly differentiate Open Source AI using any of the available standards. Nevertheless, while loosely defined, mentions of open and Open Source AI signal an intention to incentivize its development and deployment, however construed.
