AI safety institutes
AI safety institutes are government-backed technical organizations that research, measure, and evaluate advanced AI capabilities and risks to support public policy. They may develop testing methods, guidance, standards work, or risk research. The label does not imply that every institute is an independent regulator, has the same statutory powers, or can certify a model as safe.
Origin and context
The United States announced its AI Safety Institute on 1 November 2023, and the UK launched an operating institute the following day by putting a frontier-model testing function on a permanent footing. The institutional pattern then spread, but two prominent names changed in 2025: the UK body became the AI Security Institute, with an explicit security and criminal-misuse emphasis, and the former US AI Safety Institute became NIST's Center for AI Standards and Innovation, or CAISI. Current official pages confirm those successor names and mandates, so AISI is now a historical and generic category as well as an acronym still used by some national bodies.
Why it matters
These institutes give governments in-house technical capacity to examine advanced systems instead of relying only on vendor claims or general-purpose regulators. Their work can inform evaluation practice, voluntary standards, security research, and policy decisions. The 2025 US and UK changes also show why readers must check the current mandate rather than infer it from the safety label: priorities can shift toward standards, innovation, national security, or specific demonstrable risks without the underlying organization disappearing.
Example
A ministry may ask its technical institute to design evaluations for cyber or biological capabilities, run research with model developers, and translate findings into measurement guidance. The institute supplies evidence and technical expertise. Whether it can compel access, impose conditions, or enforce a rule depends on separate law and its national mandate, not on being called an AISI.
How it differs
International Network for Advanced AI Measurement, Evaluation and Science
An AI safety institute is a national or jurisdictional organization. The International Network for Advanced AI Measurement, Evaluation, and Science is a coordination forum connecting institutes and equivalent offices. The network supports shared measurement and evaluation practices, but it is not a supranational institute or regulator.
Maturity and evidence
Multiple governments have created durable technical organizations in this family, supporting maturity 4. At the same time, the UK and US rebrandings demonstrate that the category is not institutionally uniform or terminologically fixed. The mature concept is government technical capacity for advanced-AI evaluation, not one standardized AISI charter.
Limits and open questions
Public information does not justify assuming pre-deployment access, legal independence, enforcement power, or identical methods across institutes. Published evaluations may cover only selected models and risks, and institutional priorities can change with governments. This entry is descriptive policy context, not assurance that a model, developer, or deployment meets a safety or legal threshold.
Related terms
References
- Prime Minister launches new AI Safety InstituteUK Government · 2023-11-02 · class A
- Tackling AI security risks to unleash growth and deliver Plan for ChangeUK Government · 2025-02-14 · class A
- Statement on Transforming the U.S. AI Safety Institute into the Center for AI Standards and InnovationU.S. Department of Commerce · 2025-06-03 · class A
- Renaming the US AI Safety Institute Is About Priorities, Not SemanticsTech Policy Press · 2025-07-03 · class B
- About usAI Security Institute / UK Government · 2025-02-14 · class A
- Center for AI Standards and Innovation (CAISI)National Institute of Standards and Technology · 2023-10-26 · class A
- International Network for Advanced AI Measurement, Evaluation, and Science Publishes Consensus Areas on Practices for Automated EvaluationsNational Institute of Standards and Technology · 2026-02-13 · class A
- At the Direction of President Biden, Department of Commerce to Establish U.S. Artificial Intelligence Safety Institute to Lead Efforts on AI SafetyU.S. Department of Commerce · 2023-11-01 · class A
Last updated: 2026-09-07
This term is also covered in the Skills Atlas as ai risk management skill.
This term is also covered in the Skills Atlas as model evaluation skill.