The strategic vision is the inaugural public strategy document of the US AI Safety Institute (NIST AISI), released May 21, 2024 by Commerce Secretary Gina Raimondo. It sets out the institute's mission principles, the challenges it intends to address, and three strategic goals. The institute is housed within NIST and operates alongside the AISI Consortium; its mandate as described in the document is scientific and voluntary rather than regulatory.
Provenance
The document was authored by the US AI Safety Institute (NIST AISI) and published as a government strategy document on May 21, 2024 (URL: https://www.nist.gov/system/files/documents/2024/05/21/AISI-vision-21May2024.pdf). The institute was established under commitments from the Executive Order 14110 — Safe, Secure, and Trustworthy AI era, housed within NIST, and paired with the AISI Consortium as a companion body.
Mission principles
The document states two principles for the institute: that beneficial AI depends on AI safety, and that AI safety depends on science. It frames safety as enabling rather than opposing deployment, following a chain from safety to trust to adoption to innovation, and draws analogies to the historical development of aviation, electricity, automobiles, and pharmaceuticals.
Challenges identified
The document lists challenges it states the institute is intended to address:
- No commonly accepted definitions of AI safety capabilities and their measurement.
- Underdeveloped testing, evaluation, validation, and verification (TEVV) methods.
- Absence of scientifically established lifecycle risk mitigations.
- Insufficient understanding of the relationship between architecture and behavior.
- Ad hoc coordination across industry, civil society, and government.
Three strategic goals
Goal 1 — Advance AI safety science. The institute is to perform and coordinate technical research, including synthetic content detection, model security, and technical safeguards; conduct pre-deployment TEVV of advanced models for chemical, biological, cyber, and oversight risks; and conduct TEVV on present-day risks. The document states that the institute intends to be "the primary U.S. government point of contact with advanced model developers for potential pre-deployment and post-deployment AI safety testing."
Goal 2 — Develop and disseminate safety practices. The institute is to publish metrics, evaluation tools, methodological guidelines, protocols, and benchmarks, and to develop risk-based mitigation guidelines and safety mechanisms.
Goal 3 — Support institutions, communities, and coordination. The institute is to promote adoption of its guidelines and to lead an international network on AI safety science, partnering with sibling AI safety institutes, the OECD, and the G7.
Operating model
The document describes an agile, project-based portfolio operating at what it calls "speed of relevance." It states that the institute leverages the AISI Consortium for community consensus and adopts an explicit international orientation.
The document presents the institute's role through several claims. It states that the institute is housed within NIST. It positions the institute as the US government's primary contact for pre-deployment and post-deployment safety testing with frontier labs, a role the document declares for itself. The mandate is described as voluntary and scientific rather than regulatory: the document does not claim enforcement authority and frames its outputs as guidelines, benchmarks, and recommendations. Safety is framed instrumentally, as an enabler of innovation.
Comparison with other approaches
Relative to the UK AI Safety Institute (AI Security Institute), the US document is oriented toward science-building and coordination rather than immediate testing results. The UK institute published empirical capability evaluations (UK AISI — Advanced AI Evaluations May Update) within months of founding, and later consolidated two years and 30 or more frontier models of evaluation data into the Frontier AI Trends Report 2025. The US May 2024 vision describes what the institute will do; the UK Trends Report describes what its counterpart found after two years of doing it, making the two documents the strategic (US) and empirical (UK) counterparts of the AI safety institute model.
Relative to the EU AI Act (Regulation 2024/1689), the US institute is explicitly voluntary and guidance-based, whereas the EU approach is regulatory and risk-categorical. Relative to the Chinese regulatory framework (China — Interim Measures for the Management of Generative AI Services, China — Internet Information Service Algorithmic Recommendation Management Provisions, China — Provisions on the Administration of Deep Synthesis Internet Information Services), the US document emphasizes science and voluntary standards, while the Chinese measures emphasize content control, filing regimes, and real-name verification.
Relationships
- supports: AI Benchmarks and Evaluation
- supports: AI Safety Cases and Frameworks
- related: UK AISI — Advanced AI Evaluations May Update (sibling institution, different operating posture)
- related: UK AISI Frontier AI Trends Report 2025 (UK empirical counterpart to this US strategic vision; the two-year evaluation retrospective)
- related: NIST AI Risk Management Framework (AI RMF 1.0) (same host institution; complementary framework)
- related: Executive Order 14365 — Ensuring a National Policy Framework for AI (subsequent EO that reshaped federal AI policy posture)
- related: America's AI Action Plan (later strategy that updates US federal role)
- instance-of: AI Safety Cases and Frameworks