On November 5, 2023, a series of policy developments signalled a shift in how governments plan to address risks from advanced artificial intelligence. Over the same week three related initiatives were announced: a comprehensive United States Executive Order on AI, a G7 International Code of Conduct for organisations developing advanced AI systems, and an AI Safety Summit hosted by the United Kingdom at Bletchley Park that produced the Bletchley Declaration.
The US Executive Order
The US Executive Order runs to more than 100 pages and addresses a broad range of AI risks — from privacy and fairness concerns to catastrophic-risk considerations. It acknowledges the benefits AI can bring and instructs government agencies to identify opportunities to apply AI in government services, including appointing Chief AI Officers to lead those efforts.
A major strand of the Order directs work at the National Institute of Standards and Technology (NIST). NIST is tasked with developing evaluations of model capabilities and safety characteristics, expanding the Secure Software Development Framework to include secure development practices for frontier models, and producing a companion to the existing Risk Management Framework aimed at generative AI. The Order also launches a pilot for the National AI Research Resource (NAIRR), intended to improve academic researchers’ access to data and compute needed for AI safety research and beneficial applications.
The Order and accompanying materials reference related legislative and policy efforts such as the CREATE AI Act; Anthropic has publicly supported NAIRR and the CREATE AI Act since its founding.
The G7 International Code of Conduct
Developed through the G7 Hiroshima Process, the Code of Conduct for Organizations Developing Advanced AI Systems builds on voluntary company commitments announced by the White House. The Code outlines responsible practices to identify and mitigate risks across the AI development and deployment lifecycle, including mechanisms for evaluations, information sharing, governance, security procedures, and transparency measures.
If widely adopted by governments and companies, the Code could serve as an international baseline for best practice and inform domestic regulation. Anthropic has indicated its support for the G7 Code of Conduct and plans to incorporate its principles alongside the White House commitments into its development and deployment practices.
The Bletchley Park AI Safety Summit and the Bletchley Declaration
The UK-hosted AI Safety Summit convened government officials and experts from academia, industry and civil society to articulate concerns around so-called frontier AI and to discuss possible responses. Dario Amodei, CEO of Anthropic, delivered remarks describing Anthropic’s Responsible Scaling Policy, which the company views as a potential guide for other frontier developers and a prototype for regulatory approaches.
Twenty-eight countries participating at the summit issued the Bletchley Declaration, a collective statement calling for multi-stakeholder action to capture AI’s benefits and address its risks; the declaration’s signatories notably include China and a number of developing countries. The participants also agreed to support an international assessment of existing research on frontier AI capabilities and risks, intended to inform governments about the state of the science and to identify priority research areas for AI safety.
The Declaration and the proposed scientific review panel were welcomed by stakeholders who hope they will catalyse concrete, sustained international cooperation.
Government AI safety institutes and evaluation capacity
At the summit the UK announced that its Frontier AI Task Force will reconstitute as the AI Safety Institute, the first significant government initiative explicitly focused on evaluating the risks of frontier AI models. The AI Safety Institute has technical staff who have already been privately testing frontier models from labs including Anthropic — an engagement Anthropic describes as the most technically sophisticated government cooperation it has participated in so far.
The US announced plans for a US companion to the UK institute via a NIST consortium that will support development of methods to evaluate AI systems with emphasis on safety and trustworthiness. NIST will also develop guidelines for frontier threats red teaming — exercises to identify potentially dangerous dual-use capabilities of models. The communications around these initiatives note that increasing funding for NIST would strengthen its ability to deliver on this agenda.
More broadly, the write-up highlights longstanding challenges in evaluating AI systems and argues that governments should invest in their own measurement and monitoring capabilities. It cites encouraging steps by the US, UK and Singapore and refers to proposals emerging in Europe. Advancing evaluation science and establishing reliable, independent testing protocols are presented as foundational prerequisites for sensible regulation; such capabilities can also level the playing field by allowing less-resourced organisations to compete on safety evaluations and by providing objective information to governments, customers and the public.
Why this matters
The recent announcements mark the beginning of a new phase in AI safety and policy work. Major governments are showing unprecedented interest in and commitment to evaluating and monitoring AI systems, and much of their strategy depends on the ability to test and evaluate systems for capabilities, misuse potential and intrinsic safety properties. Stakeholders such as Anthropic have signalled willingness to participate and to help advance shared standards, with the stated aim of fostering a ‘‘safety race to the top.’’
Next steps
Concrete regulatory measures, funding decisions and the operational details of the new institutes will take shape in the coming months and years. Key milestones to watch include the initial workstreams and outputs of the UK AI Safety Institute and the proposed NIST consortium, the degree of adoption of the G7 Code of Conduct within industry, and progress on the Bletchley Declaration’s international scientific assessment.
