The OFAI is a university-led, collaborative initiative that aims to bend the arc toward human-centered, responsible, transparent, and ethical AI. Led by Carnegie Mellon University, the OFAI includes university representatives from across the country as well as nonprofit voices and individual fellows from industry and government. As the policy working group lead for OFAI, we collaborated with the leaders of the other working groups (research, technical prototypes, community engagement, and talent for service) and OFAI’s Executive Director, Sayeed Choudhury, to develop recommendations for an AI R&D Strategy. Our comments highlight how openness can accelerate discovery and advance AI so that everyone can benefit from its use. 

We make five recommendations:

Expand access to open datasets for AI training and evaluation

The administration should invest in the development of free and open datasets for AI to ensure big and small players have access to high-quality data, making the systems they create more accurate and useful. 

Expand access to AI R&D resources and develop AI literacy

Universities need access to computing power and Open Source software as well as training and openly licensed educational materials so that they can continue developing AI tools for faculty and students. The administration should invest in these resources, specifically for universities and nonprofits, to ensure coordination across them and sustainability of their efforts. 

Invest in the development of Open Source AI

The administration should invest in the development and deployment of Open Source AI in alignment with OSI’s evolving definition, focusing on systems that can provide real-world solutions to public challenges such as those in healthcare, education, agriculture, and research. 

Conduct research into mechanisms for openness

The benefits of openness are vast but the research community still lacks consensus on how to evaluate AI systems. The administration should support research into methodologies for evaluating AI systems and the level of openness needed to do so while coordinating these efforts with the National Institute of Standards and Technology (NIST).

Track research into downstream impacts of openness policies

Regulatory frameworks impacting Open Source AI are being considered in many jurisdictions around the world. Understanding the implications of such proposals is critical. OFAI members are conducting research into the economic impacts of these various openness regulations. The administration should engage with the OFAI to learn more about the research in this space. 

As Fiscal Year 2026 funding discussions begin in the U.S. Congress, AI R&D is sure to come up, especially in the context of federal science agency budgets. Decisions regarding funding for AI R&D will impact developers of all stripes as well as downstream users and AI researchers. OSI will continue to track this area and share updates with the community. 

May 29, 2025

Dear Mr. D’Souza:

The Open Forum for Artificial Intelligence (OFAI) is a university-led, collaborative initiative that aims to bend the arc toward human-centered, responsible, transparent, and ethical AI. The OFAI seeks to bring an academic and nonprofit perspective to critical conversations, working alongside industry and government to foster innovation such that everyone benefits from the use of AI. OFAI is led by Carnegie Mellon University (CMU) and includes the open source program offices from George Washington University, Georgia Institute of Technology, University of Texas at Austin, and North Carolina A&T State University. OFAI also has voices from the nonprofit sector including the Open Source Initiative (OSI), Creative Commons, Conscience, and the Atlantic Council as well as individual fellows from industry and government. The OFAI is pleased to provide recommendations to the Office of Science & Technology Policy (OSTP) and the National Science Foundation that highlight how openness can be leveraged in the country’s AI R&D strategy to accelerate discovery and advance AI. 

Openness Unlocks AI Innovation 

Openness is a bedrock principle in research and development that has made the U.S. science and technology apparatus the envy of the world. It is a key characteristic with application to many parts of the innovation ecosystem—from open science and open education to open data and open source software. Openness is the mechanism by which researchers, developers, entrepreneurs, and investors can share and collaborate on emerging technology and new knowledge, just as they did with the advent of the Internet and the Web. The complex nature of AI systems and the speed at which they are being applied in our everyday lives makes openness even more important today. 

Openness in AI enhances both non-commercial and commercial applications by allowing the brightest minds to contribute to and evaluate systems so that the best ones can be developed and deployed. It breaks down silos and accelerates discovery in ways closed approaches cannot. For these reasons, we believe openness should be foundational in the administration’s AI R&D strategy. We offer recommendations in key areas at the intersection of AI R&D and openness. 

Expand Access to Open Datasets for AI Training and Evaluation

Open Corpus of Knowledge for AI Training 

Openly sharing data is a means by which researchers can show their work and build trust in their conclusions. The benefits of open data in the research context can also be applied to AI systems and the data they are trained on. While the legal aspects of using copyrighted content to train AI systems is playing out in U.S. courts, there remains a need to expand the corpus of knowledge available for AI training and evaluation to increase discoverability of such knowledge. The benefits of expanding access to datasets are multifold. Creating more “open” datasets ensures innovators of all stripes—from startups and university research teams to large tech firms—can evaluate and fine-tune new models and systems. It also builds trust and allows for better evaluations of the systems themselves, enabling researchers to root out low-quality training data. And finally, it provides clarity to developers who wish to use the data but are unsure if they are legally allowed to do so. 

For example, rich open datasets in pharmaceutical R&D have given way to groundbreaking tools like AlphaFold where experts can predict protein structures and the Human Genome Project which generated the first sequence of the human genome. These accomplishments would not have been possible without access to large, high-quality, open data. Yet, researchers predict that current efforts by private companies to create closed data for use in AlphaFold2 will not generate the breadth of data needed to truly power AI. Tools like AlphaFold2 will be less successful and feats like the Human Genome Project out of reach without more open and standardized datasets. 

There are many efforts underway to develop open datasets and connect existing ones so that information for AI training and evaluation is easier to find and use. However, these efforts may not be of immediate interest to commercial entities because the content by its nature will be open to everyone. That is precisely why government investment in such initiatives is so important—it provides the entire AI community with openly available resources to accelerate innovation. This will ensure big and small players have access to high-quality data, making the systems they create more accurate and useful. 

Federal investment in and coordination of these efforts would be beneficial to extend their impact and address the growing need for AI training datasets. We highlight a few initiatives below and welcome the opportunity to work with the administration on strategic investments in this area. 

• The Public Interest Corpus seeks to create high-quality AI training data from memory organizations (e.g., libraries, archives, museums) and their partners (e.g., publishers). 
• The Institutional Data Initiative is a team of data scientists and community builders working to make knowledge collections at universities, libraries, and government agencies available as open datasets that can be used to train AI models. 
• The U.S. Repository Network aims to create a more interoperable network of open repositories (government and non-government) in the U.S. so that the information in such repositories can be reused by others.

Along with the creation of more open data, governance and community standards are critical to facilitate access to complex data sources. Civil society groups have suggested that such a governance structure should be viewed as the “Data Commons.” That is, a governance structure that is flexible enough for varying use cases. It is crucial for the U.S. to be engaged in such global and domestic governance discussions. 

Expand Access to AI R&D Resources and Develop AI Literacy 

Universities across the U.S. are rapidly adopting AI tools and platforms for faculty, students, and researchers to work with models and learn valuable AI literacy skills. There are also a growing number of schools developing their own AI platforms. The Dietrich Analysis & Research Education (DARE) platform, built as an Open Source project at CMU, promotes human-centered AI by allowing students, faculty, and staff to leverage multiple large language models, transform data and experiment through a locally controlled LLM gateway. This enables faculty to augment research capabilities and develop curriculum in a platform that adapts to their pedagogical needs rather than adapting their pedagogy to fit available tools. DARE puts AI in the loop with humans in control, promoting human agency and interaction transparency while empowering students to use AI responsibly. Sage, a tool at the University of Texas at Austin, is an AI teaching and learning guide that draws on the LLM Claude and established principles of learning experience design and responsible AI adoption. With Sage, faculty at UT Austin can design tutoring sessions for students on any topic. 

While initiatives like those at CMU and UT Austin are growing across the country, the experts and technologists that build them need resources beyond just open data. They need access to computing power and Open Source software as well as training and openly licensed educational materials. Computing power in particular is often out of reach for tool builders at universities or nonprofit research centers due to prohibitive costs. Further, they need a sustainable funding environment for these resources such that they can take advantage of the latest model developments and immediately deploy them. 

The government’s AI R&D strategy should include investment in these resources to ultimately build a public infrastructure for AI as well as sustained funding for a coordinated network of universities building cutting-edge AI tools. The federal program supporting the network of universities should be flexible enough to enable experts at those universities to adapt to and leverage new information about AI instead of waiting months or years for another funding cycle. 

Invest in the Development of Open Source AI 

Open Source software is another crucial component of any AI system. And while an AI system is much broader than the software code, the ubiquitous nature of Open Source software provides a clear example of what can happen when technology is shared without restriction and creators are given the freedom to innovate. The Open Source Definition, maintained by the Open Source Initiative (OSI), removes barriers to learning, using, sharing, and improving software systems. Today, Open Source software accounts for more than 97 percent of applications we use and 90 percent of companies report using open source software in some way. Recognizing the need to apply Open Source principles to AI, OSI co-developed the Open Source AI Definition (OSAID), releasing version 1.0 in October 2024. In 2025, OSI is leading a community effort to evaluate the definition and identify models that meet it. This work will inform future iterations of the definition and best practices for developing truly Open Source AI. 

The Open Source AI Definition

An Open Source AI is an AI system made available under terms and in a way that grants the freedoms to:

• Use the system for any purpose and without having to ask for permission.
• Study how the system works and inspect its components.
• Modify the system for any purpose, including to change its output.
• Share the system for others to use with or without modifications, for any purpose.

The preferred form of making modifications to a machine-learning system must include all the elements below:

• Data Information: Preferably the original data, or if it is not legally possible, sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system. Data Information shall be made available under OSI-approved terms.
• Code: The complete source code used to train and run the system. The code shall represent the full specification of how the data was processed and filtered, and how the training was done. Code shall be made available under OSI-approved licenses.
• Parameters: The model parameters, such as weights or other configuration settings. Parameters shall be made available under OSI-approved terms.

The administration’s AI R&D strategy should include an investment in the development and deployment of Open Source AI in alignment with OSI’s evolving definition. Investments in Open Source AI should focus on systems that can provide real-world solutions to public challenges such as those in healthcare, education, agriculture, and research itself. For example, truly Open Source AI models like the Allen Institute’s (Ai2) OLMo provide researchers with a large language model that can be just as powerful as many of the proprietary ones but with access to the components and without the price tag. Ai2’s most recent release, OLMo 2 32B, outperforms GPT3.5-Turbo and GPT-4o mini on a variety of benchmarks. Especially at a time when new knowledge is being learned everyday about how to improve AI systems, researchers and developers need access that grants them the freedom to use, study, modify, and share the system and its components. 

Further, countries around the world are trying desperately to catch up to America’s lead in AI by developing innovative models that share some of their components openly. China’s DeepSeek is one such example that has demonstrated just how powerful an open model can be. The U.S. needs more Open Source AI systems to provide researchers, developers, entrepreneurs, and investors options for using and building on models that have greater transparency and lower barriers to adoption.   

Conduct Research into Mechanisms for Openness

The benefits of openness are vast but the research community still lacks consensus on how to evaluate AI systems. A recent survey of academics and corporate researchers conducted by the Association for the Advancement of Artificial Intelligence (AAAI) found that a lack of suitable evaluation methodologies and the black-box nature of AI systems were the biggest challenges to evaluating them. The administration should conduct research, through intramural and extramural grants, into methodologies for evaluating AI systems and the level of openness needed to do so. It should also consult AI researchers and drive consensus around such methods through the National Institute of Standards and Technology (NIST) and Federal science agencies. 

Research and consultations in this area should address the following:

• How should AI systems be evaluated for risk—from personal safety to national security in presence of access barriers? 
• What level of openness is needed to evaluate systems for such risks?
• What methodologies exist for providing that transparency?
  • For example, OSI’s Open Source AI Definition (OSAID) is an existing framework that ensures key information about AI systems are shared openly.

Track Research into Downstream Impacts of Openness Policies

Regulatory frameworks impacting Open Source AI are being implemented or are under consideration in many jurisdictions around the world. Understanding the implications of such proposals is critical. Members of the OFAI are conducting research into the economic impacts of various openness regulations. Their research aims to address the following:

• When do openness regulations enhance model access, transparency, and innovation, and when might they hinder these goals?
• How should openness regulation be designed to encourage greater competition and investment in AI development? 
• How should the concept of “openness” in AI be defined for regulatory purposes?

We urge the administration to review this research and engage with members of the OFAI to understand its implications. We would welcome the opportunity to meet with you and discuss this research further. 

We thank the Office of Science & Technology Policy and the National Science Foundation for the opportunity to contribute ideas to the administration’s 2025 National AI R&D Strategic Plan. We look forward to working together to accelerate AI-driven innovation. 

Sincerely, 

Contact

• Encourage the development and adoption of Open Source AI
• Establish a common definition of Open Source AI
• Promote a data commons approach for AI development 

The think tank used a variety of tools to analyze the public comments and extract specific recommendations. The recommendations are organized into 20 topics including Export Controls, Infrastructure, Security, Open Source, and Standards & Regulations. The Open Source category has 156 recommendations from 123 organizations/businesses. The database is a helpful resource that anyone can use to explore submissions. 

Other recommendations of note highlighted by IFP include:

• Increase government use of Open Source AI (Mozilla)
• Signal government support for Open Source AI (a16z)
• Allocate public computing resources for Open Source AI (Hugging Face)
• Preserve and promote open model development (Center for Democracy & Technology)
• Promote open innovation as a strategic advantage (Stanford Institute for Human-Centered AI)
• Support transparency in AI procurement, development and use (Electronic Frontier Foundation)

OSI thanks IFP for recognizing Open Source as an important theme in the AI Action Plan recommendations and looks forward to working with the administration and other organizations on these issues.

Our comment outlines the benefits of Open Source and aims to position the organization and the broader Open Source community as a resource for policymakers in the Trump Administration working on the development of an AI Action Plan. To this end, we detailed OSI’s history of stewarding the Open Source Definition and our recent work in co-developing the Open Source AI Definition (OSAID). We also cited the federal government’s continued use and support of Open Source across agencies including the Department of Defense and the National Science Foundation. 

We recommended that the White House rely on the OSAID as a foundational piece of any future AI Action Plan and invited the administration to work with OSI and our community — from industry and academia to startups and foundations — on unifying the tech community around the definition of Open Source AI to remove confusion and enable innovation. 

We thank our partners at the Apereo Foundation for endorsing our comment and look forward to working with other members of the Open Policy Alliance and the broader OSI community to engage with the new administration on the critical issue of AI.

March 15, 2025

Faisal D’Souza, Technical Coordinator
Networking and Information Technology Research and Development (NITRD)
National Coordination Office (NCO)

Dear Mr. D’Souza:

The Open Source Initiative (OSI) appreciates the opportunity to provide input on the development of an Artificial Intelligence (AI) Action Plan. OSI is a nonprofit organization created in 1998 to steward the Open Source Definition. The definition serves as a foundation to maintain the fundamental freedoms of Open Source on which the modern software ecosystem is built. Our community represents a wide range of stakeholders — from industry and academia to startups and foundations — all working to promote the benefits of Open Source Software and connect practitioners to policymakers. We believe Open Source can both strengthen security and unleash innovation by ensuring the brightest minds can work together, free of barriers. Most recently, OSI undertook the co-development of an Open Source AI Definition (OSAID), recognizing the need for clarity around what it means for an AI system to be truly Open Source. We are pleased to contribute to the AI Action Plan to ensure the innovations made possible by Open Source can be leveraged to increase U.S. competitiveness and national security. 

The Strengths of Open Source 

Open Source Software underpins almost every aspect of technology we use. Recent reports have found that 97 percent of applications contain Open Source Code and 90 percent of companies use Open Source in some way. Open Source became ubiquitous because developers, researchers, and business leaders recognized that open collaboration is needed to scale the benefits of technological innovation, including the Internet, because it leads to higher quality, better reliability, greater flexibility, and lower costs. Today, U.S. federal agencies have also recognized the immense benefits of Open Source Software, incorporating its use into critical infrastructure, including at the Department of Defense and research conducted by the National Science Foundation.  

The Open Source Definition that OSI maintains for the community removes barriers to learning, using, sharing, and improving software systems. The core criteria of the Open Source Definition have given innovators the freedom to build and deploy new technologies faster than ever before while maintaining flexibility and control in the development process. 

For example, Open Source is critical in key aspects of software development such as security, cost, and time. When code is Open Source, more developers are able to review and identify security vulnerabilities in a software program and then fix them — often in less time and at a lower cost than if only a few developers had access to the code. Many efforts, such as GitHub’s Secure Open Source Fund, aim to make Open Source even more secure through investments in training and collaborations across the software community. Further, researchers have not found evidence that models with open weights are any more vulnerable to cybersecurity threats or misuse than closed software.

The freedoms of the Open Source Definition that built the Internet will be imperative to maintain America’s technological competitiveness in the age of AI and beyond. 

The Open Source AI Definition

In 2022, OSI realized that the traditional definition of Open Source was not sufficient when applied to complex AI systems i.e., you need more than just access to the source code to understand how these systems work. Everyone — from developers and policymakers to researchers and consumers — needed a common understanding of “Open Source AI.” The organization began co-developing the first version of the Open Source AI Definition, or OSAID, and released it in October 2024 with more than 20 organizational endorsements. The definition is a foundation on which additional parameters can be built to address a variety of use cases in the public and private sectors. OSI recognizes the definition may evolve as discussions continue around AI systems. 

The Definition: What is Open Source AI?

An Open Source AI is an AI system made available under terms and in a way that grants the freedoms to:

• Use the system for any purpose and without having to ask for permission.
• Study how the system works and inspect its components.
• Modify the system for any purpose, including to change its output.
• Share the system for others to use with or without modifications, for any purpose.

The preferred form of making modifications to a machine-learning system must include all the elements below:

• Data Information: Preferably the original data, or if it is not legally possible, sufficiently detailed information about the data used to train the system so that a skilled person can build a substantially equivalent system. Data Information shall be made available under OSI-approved terms.
• Code: The complete source code used to train and run the system. The code shall represent the full specification of how the data was processed and filtered, and how the training was done. Code shall be made available under OSI-approved licenses.
• Parameters: The model parameters, such as weights or other configuration settings. Parameters shall be made available under OSI-approved terms.

Data Commons in Open Source AI

In addition to the OSAID, OSI is working to promote data governance as a critical component of Open Source AI. OSI released a white paper, “Data Governance in Open Source AI: Enabling Responsible and Systematic Access” in partnership with Open Futures, a think tank developing new approaches to the open Internet. In the white paper, we suggest adopting a data commons approach that recognizes the various complexities of data sharing and envisions an Open Source AI ecosystem that can include both open and restricted datasets. For example, sharing the dataset used to train an AI system allows others to interrogate it and address biases and inaccuracies. However, there are many instances where an underlying dataset may be restricted or only available through tiered-access for valid reasons such as protecting patient privacy, respecting indigenous knowledge, or adhering to intellectual property laws. Open Source AI models with these different types of datasets play an important role in society and should be governed by a data commons that encourages transparency while respecting the nuances of each. 

Open Source is Imperative in an AI Action Plan

At a time when AI is evolving rapidly, the U.S. government’s AI Action Plan should include Open Source AI as a foundation to ensure as many minds as possible — from startups and researchers to major tech firms — are working to find the next innovation. Driving collaboration through Open Source can bolster the competitiveness of US companies on the world stage. Competitiveness is enhanced because the freedoms of Open Source AI mean frictionless sharing and developing across projects so that the best ideas are not siloed, but rather debated and built upon. The value of Open Source was highlighted in a 2024 paper where researchers found that, without Open Source Software, companies would need to spend an estimated 3.5 times more or nearly $9 trillion to build the software that powers their businesses.

Since the dawn of the Internet, Open Source has been recognized as a building block to develop and scale technology. To this end, we recommend that the AI Action Plan encourage the development and use of Open Source AI models in both the public and private sectors to speed discovery and application of AI in ways that benefit public wellbeing. We also invite the U.S. government to work with us and our partners to further unify the tech community around a definition of Open Source AI. A common understanding will remove confusion and enable more innovation. 

OSI appreciates the opportunity to provide input on the U.S. government’s AI Action Plan and welcomes engagement with our community on the issues of Open Source and Open Source AI. 

Sincerely, 

Open Source Initiative 
Contact: Katie Steen-James, Senior U.S. Policy Manager

Endorsing Organization
Apereo Foundation
9450 SW Gemini Dr PMB 98572
Beaverton, OR 97008-7105
www.apereo.org 
Contact: Patrick Masson, Executive Director