������Ƶ

There is no easy binary that opposes “open” and “closed” in the case of artificial intelligence (AI) models. Instead, openness should be viewed as a spectrum.¹ This more flexible understanding of openness fosters productive conversations about both the varied benefits of openness and the marginal risks associated with open models relative to closed models or what is already publicly available online.²

“Open-source AI” is a term with a definition that is still in flux and likely to be defined differently by different stakeholders. The Open Source Initiative (OSI), a nonprofit that advocates for the benefits of open source and acts as a multistakeholder standards body that maintains a widely-used definition of open source software as a public benefit, just released its first definition of the term “open source AI.”³ The OSI definition clarifies that to be open, an entire AI system must be considered—both in its “fully functional structure and its discrete structural elements.” OSI further notes that “the requirements are the same, whether applied to a system, a model, weights and parameters, or other structural elements.”⁴

Efforts such as OSI’s to align on a specific definition of open-source AI that is based on certain criteria are helpful, as different actors are currently using the term in different, often self-serving, ways. In a paper last year, David Gray Widder, Sarah West, and Meredith Whittaker described this phenomenon of “openwashing,” in which a model developer misleadingly claims the mantle of openness for public relations gains while actually providing access to their model in a way that “should be understood as ‘closed.’”⁵ They suggest that models fall on “gradients of openness” in which the term “open” can describe models that “offer vastly differing levels of access.”

The authors offer three attributes for understanding the openness of models: (1) transparency, (2) reusability, and (3) extensibility. Transparency denotes “the ability to access and vet source code, documentation and data;” reusability is “the ability and licensing needed to allow third parties to reuse source code and/or data;” and extensibility is “the ability to build on top of extant off-the-shelf models, ‘tuning’ them for one or another specific purpose.”⁶ These attributes are a useful framework for examining a model’s software components.

These attributes also highlight many of the same principles required by OSI’s definition. For OSI, an open-source AI system must allow use “for any purpose and without having to ask permission,” the ability to study how the system works, the ability to modify it, and to share it “with or without modification.” These overlaps suggest a growing agreement around the term “open,” particularly the need to include transparency, access, and modification in the definition.

Relevant to the ongoing discussions about how to define open models is the question of access to training data. While some models describe themselves as open and provide code and model weights, they do not provide access to the data used when training the model. A group of scholars recently suggested using the term “open-access AI” in this context, arguing that “‘open-source AI’ is a misnomer for such models” due to “meaningful differences in access, control, and development.”⁷ OSI’s definition similarly regards access to training data as an essential test in determining whether or not a model is truly open source.

To illustrate the concept of a spectrum of openness, we offer a simplified breakdown with examples. We have chosen this range of attributes as an exercise in illustratively drawing the line between models, recognizing that the spectrum could consist of many more attributes and reflect greater nuance.

The exercise of defining open-source AI, or even placing AI models along such a spectrum, demonstrates that the emerging requirements for an AI model to be considered open are similar to those in earlier free and open-source software projects. To examine—as OSI suggests—the entire system of an AI, we must investigate more attributes than simply the code or the weights. We must think more broadly of models as software projects. The history of open-source software is full of instructive examples of how to (and not to) structure and maintain large software projects, and it is also full of examples of unintended consequences that have shaped tech.

The term “open source” is used throughout this report in a way that includes consideration of all software licensing that meets both the Free Software Foundation’s definition of free software⁹ and the Open Source Initiative’s “Open Source Definition.” (OSD).¹⁰ We have chosen to use “open-source” to refer to code as it resonates with the current discourse around open models and not because of a particular preference or recommendation for existing open software licenses. We use the term “open model” throughout this report to echo that discussion but recognize that the lexicon around AI and openness is changing and may ultimately need more terminology—like “open access”—to meaningfully distinguish among model types in the future.

Much of the prevailing discourse around open models focuses on risks and fails to fully account for the significant societal benefits of open models to public transparency and accountability, to unexpected innovation and competition, to education and research, and to security. The following sections explore each of these benefits in further detail.