AI safety isn't one field run by one group. It's a mix of the companies building AI, governments testing it, and independent researchers checking everyone's work. Each has different incentives, which is why you need all three.
The AI labs
The companies building the most capable models, including Anthropic, Google DeepMind, and OpenAI, all run internal safety teams. They test models for dangerous abilities before release and publish safety policies describing what would make them pause.
The catch: they're grading their own homework, under intense commercial pressure to ship faster than rivals.
Governments
UK AI Security Institute: a government team that tests frontier models for high-risk capabilities and checks whether safeguards hold up.
US Center for AI Standards and Innovation (CAISI): works with NIST and industry on testing and evaluation methods. In May 2026, Microsoft announced partnerships with both CAISI and the UK institute to test its frontier models.
The EU AI Office: enforces the EU AI Act, including rules for the most powerful general-purpose models.
The catch: governments are still building the expertise to keep up, and testing is largely voluntary outside the EU.
International efforts
The International AI Safety Report 2026, led by Yoshua Bengio with more than 100 experts from over 30 countries, is the closest thing to a shared scientific picture of AI risk.
Independent researchers
Nonprofits like METR, Apollo Research, and Redwood Research test models from the outside, study deceptive behavior, and develop new safety techniques. University groups around the world contribute research on alignment and interpretability.
The catch: they depend on labs granting access to unreleased models, and they're small compared to the companies they study.
Why this matters
When a new model launches, the most important question is often: who tested it, and what did they find? We'll tell you every time.
Sources
Get every issue in your inbox. It's free, and you can unsubscribe anytime.