AI is getting better faster than experts can reliably test whether new systems are safe, METR warns

METR warns that leading AI models are improving faster than researchers can reliably measure their safety, making it difficult to detect potentially dangerous acceleration in AI capabilities. The organization notes that current evaluation methods may both fail to catch real risks and overestimate what models can actually perform.
Why it matters
If AI systems advance beyond our ability to test them, regulators and organizations cannot make informed decisions about deployment risks.


