AI Safety Institute in plain English.

An AI Safety Institute is a government body set up to independently test frontier AI models for dangerous capabilities — things like helping create weapons, executing sophisticated cyberattacks, or deceiving their own overseers — separate from the safety testing the AI labs do themselves. The UK and US both stood up AI Safety Institutes in 2023-2024, and by 2026 similar bodies exist in Japan, Singapore, Canada, and the EU, sometimes coordinating through an international network. In practice, major labs like Anthropic, OpenAI, and Google DeepMind now give these institutes early, pre-release access to frontier models for testing, an arrangement that is voluntary in most jurisdictions rather than legally required. The institutes have limited enforcement power in most countries — they can publish findings and advise regulators, but generally can't block a release outright — which is the main criticism leveled at the model: independent testing without independent teeth.