概述
They focus on risks such as cyberattacks, chemical and biological misuse, and loss of human control, often evaluating models before public release. They matter because they give governments independent technical expertise on frontier AI, although most depend on voluntary cooperation from companies rather than legal powers.
深入探討
The UK set up its AI Safety Institute in November 2023, around the Bletchley Park AI Safety Summit. It grew out of the Frontier AI Taskforce and became one of the best-funded government AI evaluation teams, hiring researchers from leading labs. In February 2025 it was renamed the AI Security Institute. The new name reflected a narrower focus on national-security harms such as cyber misuse, chemical and biological risks, and crime, rather than broader issues such as bias. The US AI Safety Institute was created within NIST in late 2023, after President Biden's executive order on AI. In June 2025 the Commerce Department reorganized it as the Center for AI Standards and Innovation (CAISI). The mandate shifted toward voluntary standards, national-security-relevant evaluations, and assessing foreign AI systems, while dropping the "safety" branding. Japan, Singapore, Canada, South Korea, France and others have set up similar bodies. The EU's AI Office has some parallel functions, but it is a regulator. In November 2024 the International Network of AI Safety Institutes held its first meeting in San Francisco. Its founding members were Australia, Canada, the EU, France, Japan, Kenya, South Korea, Singapore, the UK and the US. A common misconception is that these institutes can block model releases. The UK and US bodies have no general statutory power to compel access or stop deployment. Their access comes from voluntary agreements, and their findings shape policy and company decisions without being binding. The EU AI Office is different: under the AI Act it can request information from providers of general-purpose AI models with systemic risk and enforce obligations on them.
戰略影響
風險與安全
災難性和日常的人工智慧危害都取決於誰了解風險以及誰能夠採取行動。
更明確的決策
民眾和專業素養決定強而有力的安全政策在政治上是否可行。
突破炒作
清晰的解釋可以減少炒作、實驗室公關和模糊道德劇場的影響。
The Future of AI Safety Institutes Around the World
The institutes' role is still changing. Several have moved from broad safety research toward national security, standards and industrial competitiveness, and their priorities may keep shifting with governments. Key open questions are whether any of them will get statutory powers, how their work will connect to regulators such as the EU AI Office, and whether international collaboration will deepen or split along geopolitical lines. Their long-term value may depend on shared, reproducible evaluation methods and enough funding to keep up with frontier labs.
現實世界的實施
The UK and US institutes jointly tested an upgraded Claude 3.5 Sonnet and OpenAI's o1 before deployment and published summaries of the results in late 2024.
The UK institute released Inspect, an open-source framework that researchers and other governments use to build and run AI evaluations.
In 2024, the US institute signed agreements with OpenAI and Anthropic that gave it access to major new models before and after public release.
Institutes from several countries have run joint testing exercises to compare methods, for example evaluating agent behavior in different languages.
風險與防護欄
將存在風險視為科幻小說，同時能力複合。
混淆了表面產品安全與高度自治下的對準。
只給非英語和非專業觀眾留下低品質的資源。
實施路線圖
單獨的產品危害、誤用和失控/失調風險。
詢問哪些證據會改變您對時間表和嚴重性的看法。
比起行銷主張，更喜歡主要來源和具體評估。
確定一條行動路徑：職業、政策、資金或技能——而不僅僅是意識。
不斷探索
Free newsletter
Get the daily AI briefing
Three verified AI stories every weekday morning, written in plain English. Free forever, no ads.
One email each weekday. Unsubscribe in one click. We never sell or share your address.
Test yourself
Take the AI Safety Institutes Around the World quiz
Instant feedback on every answer, and a shareable certificate with a verifiable ID once you pass a course.
Support free AI education. AI Understanding is a 501(c)(3) nonprofit — no ads, no paywall, ever. Make a donation
常見問題
What is AI Safety Institutes Around the World?
AI safety institutes are government bodies that test and study advanced AI models. They focus on risks such as cyberattacks, chemical and biological misuse, and loss of human control, often evaluating models before public release. They matter because they give governments independent technical expertise on frontier AI, although most depend on voluntary cooperation from companies rather than legal powers.
What was the UK AI Safety Institute renamed in February 2025?
The new name, AI Security Institute, reflected a narrower focus on national-security risks such as cyber, chemical and biological misuse, and crime.
The US AI Safety Institute was reorganized into which body in 2025?
The Commerce Department turned it into CAISI, which focuses on standards, national-security evaluations and assessing foreign systems.
Where did the International Network of AI Safety Institutes hold its first meeting?
The first meeting was held in San Francisco in November 2024.
What legal power do the UK and US institutes generally have over model releases?
Their access comes from voluntary agreements, and their findings are advisory rather than binding.
What is Inspect?
Inspect organizes evaluations into datasets, solvers and scorers, and supports sandboxed agent testing.
繼續學習
相關指南
為此主題精選的更多指南