On July 23rd, the US Administration released its highly anticipated AI Action Plan, marking a significant shift from the former administration’s approach to AI. The new Plan is supported by Executive Orders on ‘Preventing Woke AI in the Federal Government’, ‘Accelerating Federal Permitting of Data Center Infrastructure’ and ‘Promoting The Export of the American AI Technology Stack’. Together, these documents signal a new focus on AI innovation, investment in infrastructure, and international diplomacy and security, while continuing to support evaluations and efforts to advance foundational science.
The Plan has received a range of responses from different organizations: Meta has welcomed changes that support greater investment and IBM’s CEO has applauded the prioritization of open innovation, while over 100 civil society groups and academic institutions have signed the People’s AI Action Plan which focuses on “public well-being, shared prosperity, a sustainable future, and security for all,” providing an alternative path forward.
Although the Executive Orders provide specific targets, actors and timelines to advance certain aspects of the Plan, it is still unclear how different agencies would implement many of these policy recommendations. As a multistakeholder partnership, we believe that a core feature of any policy strategy that benefits people and society is the involvement of the organizations and people it will affect. We see opportunities for the Administration to build on parts of the Plan by involving stakeholders, and for our cross-sector community to continue to advance safe, responsible AI in the evolving policy landscape.
Building on the AI Action Plan
The Plan recognizes that different stakeholders can make vital contributions to progressing AI governance. For example, it calls for the Center for AI Standards and Innovation (CAISI) to bring together Federal agencies and the research community to share best practices on building AI evaluations. It also calls for NIST to convene a broad range of public, private, and academic stakeholders to accelerate the development and adoption of national standards for AI systems.
We see promising opportunities to build upon the Plan’s initiatives by furthering stakeholder collaboration, particularly in labor and the workforce and foundational research.
“A core feature of any policy strategy that benefits people and society is the involvement of the organizations and people it will affect”
Impact of AI on Labor and American Workers
The Plan acknowledges the significant impact of AI on ‘the labor market and the experience of American workers’ and sets a number of priority actions to measure this impact and prepare workers. Recommendations include the establishment of the AI Workforce Research Hub under the Department of Labor (DOL) to conduct recurring analyses. They also include a mandate for the Census Bureau and the Bureau of Economic Analysis (BEA) to study AI’s impact on the labor market using existing data, which could include Business Trends and Outlook Survey. These actions are a step in the right direction to ensure policymakers have quality data to inform how to support workers and foster high quality jobs.
At PAI, we have dedicated significant research efforts into the impact of AI on labor and the economy. We believe the DOL, the Department of Education, the Department of Commerce and other relevant agencies should actively involve workers in the development of proposed workforce training and education policies as well as efforts to map out the impact of AI on workers. This is crucial to ensuring that these policies truly empower workers and address their needs.
Foundational Research
PAI has also consistently called for increased investment in open research and innovation within the AI ecosystem. On this basis, we welcome the Plan’s recognition that America should foster leading open models founded on American values.
We are keen to see how the NSF R&D Plan can operationalize much of the work related to foundational research. We encourage any funding to focus on both technical and sociotechnical research, as well as prioritizing multistakeholder research. (You can see PAI’s response to NSF’s R&D Plan here)
Encouraging Open-Source
The Plan recognizes the academic, commercial and public benefits of open-source and open-weight AI models, and recommends actions to guarantee that startups and researchers have access to the necessary compute, models, data, and software resources as part of the NSF’s National AI Research Resource (NAIRR) pilot. PAI has supported the development of open AI ecosystems to promote innovation and ensure that the benefits of AI are shared widely across America, and welcomes these developments.
To build on these proposals, policymakers and partners should continue to explore how the adoption of open-source AI systems can be done safely, with responsibilities shared along the value chain. Our best practices for mitigation strategies across the AI value chain detail how different actors can ensure that open-source systems are adopted responsibly.
An Independent Assurance Ecosystem
The Plan recognizes the value of an AI evaluations ecosystem, including testbeds and regulatory sandboxes as key components. It recommends that NIST and CAISI publish guidelines about AI evaluations for federal agencies, convene meetings to share best practices, and update national security-related AI evaluations. Evaluations are a key part of PAI’s Guidance for Safe Foundation Model Deployment, and we welcome the recognition of their importance.
However, there is still a gap around ensuring that private systems used by individuals and enterprises in America and elsewhere are safe, secure and interpretable – and that deployers and other downstream actors trust that AI has these characteristics. Meeting these goals will require an independent AI assurance ecosystem. Such an ecosystem would foster a variety of assurance methods that address all relevant AI attributes, such as safety, security, and resilience, interpretability, transparency and accountability, and would include testing, evaluation, validation and verification (TEVV) methods already being developed by NIST to support private sector-led standardization. Independent assurance will play a vital role in promoting the AI Action Plan’s core objective of promoting adoption.
Addressing Gaps in the Plan
While the AI Action Plan makes significant strides towards AI adoption and innovation, this focus also leaves gaps that will necessitate increased action from our partnership to ensure AI is developed and deployed responsibly to benefit people and society.
Safety Guardrails for AI Systems
The Plan recognizes that prudent planning is required to mitigate the impact of AI system failures on critical services, in particular, by developing and incorporating AI Incident Response actions and encouraging the responsible sharing of AI vulnerabilities. However, it does less to advance governance approaches that seek to promote AI systems that are robust and resistant to other risks, such as hallucinations and anthropomorphisation.
PAI has published research and recommended best practices to ensure that AI systems are robust and trustworthy. Our Guidance for Safe Foundation Model Deployment guides model providers in responsibly developing and deploying foundation models across a spectrum of current and emerging capabilities. We are also exploring the risk of humans developing relationships with increasingly ‘personlike’ AI systems. As AI systems continue to evolve, especially with the development and deployment of more autonomous AI agents, the multistakeholder community will need to continue developing cross-cutting research and sharing best practices.
“International cooperation is a two-way street, requiring flexibility and willingness to accommodate global values”
Supporting International Leadership
In Pillar III, the Plan sets out goals to lead in international AI diplomacy and security. PAI welcomes efforts to engage with allies and form global alliances. However, international cooperation is a two-way street, requiring flexibility and willingness to accommodate global values. It remains to be seen whether other countries will be receptive to the AI Action Plan and the AI technologies built on its policies. For example, some of the biggest AI markets around the world, such as India, Brazil and Germany, feel strongly about the need for AI content moderation and sustainable development.
Relatedly, the push for exports of the American AI technology stack in Pillar III showcases how international trade and investment are crucial to the development of a strong US AI market. But as the importers of this technology, foreign governments get to decide whether it meets their needs, expectations, and local requirements. Therefore, to maintain its leadership in the global AI market, the US will have to work together with other countries to find common ground on key principles of AI governance. This is especially important as China’s new “Action Plan on Global Governance of Artificial Intelligence” might appeal to many Global South countries, with its focus on cooperation, equity, safety, and respect for national sovereignty.
Moving forward
The Administration’s AI Action Plan does not exist in a vacuum. The Plan will influence the global governance of AI, and as leaders in the development and implementation of AI, the US has a responsibility to protect people affected by American AI technologies, both at home and abroad.
With the 80th session of the UN General Assembly and the launch of the UN Global Dialogue on AI approaching in September, governments around the world have an opportunity to discuss how they want to shape the global governance of AI. In doing so, they should uphold foundational international principles, including sovereignty, self-determination and human dignity as AI technologies continue to be disseminated around the world. The US and other democracies around the world have a key role to play in safeguarding these principles as new proposals for global AI governance frameworks and mechanisms gain momentum, including China’s recent AI Action Plan.
The US AI Action Plan lays down the direction of AI policy for the next four years, and will be followed by details on exactly how this vision will be achieved. In the areas above, where PAI has critical evidence and research, we welcome the opportunity to engage with Government agencies to ensure that the implementation of the AI Action Plan benefits people and society.
We encourage the Administration to consider strengthening the involvement of affected people and societal groups in both the development and implementation of AI policies, especially in areas such as AI and labor and foundational research.
We also encourage our global partnership to continue to carry out technical and socio-technical research, as well as to participate in the development of an independent AI assurance ecosystem and global governance frameworks to ensure that AI technology advances are for the benefit of people and society.