The US AI Safety Institute (US AISI) was a federal entity for developing the science of AI safety, housed within the National Institute of Standards and Technology (NIST) under the Department of Commerce. It was established following commitments under the 2023 Biden AI executive order (later revoked via Executive Order 14365 — Ensuring a National Policy Framework for AI), and operated alongside a companion AISI Consortium, a multi-stakeholder convening body. This page preserves the pre-2025 framing and historical mission as a record.
The US AISI became CAISI, the Center for AI Standards and Innovation, in June 2025, when Commerce Secretary Howard Lutnick announced the institute's transformation (Source: nist.gov; fedscoop.com). Commentators read the renaming as a shift in priorities from safety toward security, standards, and innovation rather than a purely cosmetic change (Source: techpolicy.press). For the current mission, structure, and recent activity, including the AI Agent Standards Initiative, the May 5, 2026 pre-deployment evaluation agreements with Microsoft, Google, and xAI and their May 11 deletion, and the DeepSeek R1-0528 jailbreak benchmark, see NIST CAISI (Center for AI Standards and Innovation). (Correction, 2026-07-08: an earlier revision of this page dated the renaming to February 14, 2025; the announcement was June 2025 per the sources above.)
Mission and principles
The institute's strategic vision rested on two declared principles (see US AI Safety Institute — Vision, Mission, and Strategic Goals):
- Beneficial AI depends on AI safety.
- AI safety depends on science.
It framed safety as enabling deployment rather than constraining it, describing a progression from safety to trust to adoption to innovation.
Strategic goals
The strategic vision set out three goals:
- Advance AI safety science through empirical testing, evaluation, validation, and verification (TEVV) of frontier models, systems, and agents.
- Develop and disseminate safety practices, publishing metrics, benchmarks, protocols, and mitigation guidelines.
- Support institutions and coordination, including leading an international AI safety science network and promoting voluntary adoption.
Operating role
The institute described itself as the primary US government point of contact with advanced model developers for pre- and post-deployment voluntary safety testing. It collaborated across NIST labs, US agencies, and international partners including the OECD, the G7, and sibling AI safety institutes. Its portfolio was organized as agile, project-based work, with a stated commitment to "speed of relevance."
In 2024 the institute appointed alignment researcher Paul Christiano as head of AI safety, a role NIST describes as designing and conducting tests of frontier AI models (Source: nist.gov); he retained the safety post through the CAISI transition.
Policy context and status
The US AISI's strategic vision, published in May 2024, predates Executive Order 14365 — Ensuring a National Policy Framework for AI, which revoked the Biden AI executive order, and America's AI Action Plan, which reoriented federal AI policy toward innovation and competitiveness. Some specific commitments may have been superseded or reframed, and the mandate's scope under the current administration remained in flux. The institute's voluntary-testing posture contrasted with EU AI Act (Regulation 2024/1689)'s regulatory obligations and with China's mandatory filing regime (see China — Internet Information Service Algorithmic Recommendation Management Provisions).
Relationships
- related: NIST CAISI (Center for AI Standards and Innovation) (successor body, June 2025)
- related: Paul Christiano (head of AI safety, 2024–; retained through the CAISI transition)
- related: UK AI Safety Institute (AI Security Institute) (sibling institution; more empirical posture)
- related: NIST AI Risk Management Framework (AI RMF 1.0) (companion NIST framework)
- related: Executive Order 14365 — Ensuring a National Policy Framework for AI (later EO reshaping federal AI stance)
- related: America's AI Action Plan (subsequent federal AI strategy)
- related: AI Safety Cases and Frameworks
- related: AI Benchmarks and Evaluation