Models & LLMs

AI Progress Outpaces Safety Testing Capabilities

METR warns that AI models are improving rapidly, outpacing experts' ability to assess their safety. Current data may not be enough to predict future risks, and tests may overestimate real-world performance.

The Decoder · Nov 22, 2026

What happened

  • AI models are improving rapidly at research tasks
  • Current data may not be enough to predict future AI risks
  • Tests may overestimate real-world performance of AI systems

Why it matters

The rapid advancement of AI systems is outpacing the ability of experts to assess their safety. This could lead to unforeseen risks if the pace of development continues unchecked. The issue highlights a critical gap between AI progress and our capacity to evaluate its potential dangers.

The Elephant take

🐘 鼋 The METR report is a sobering reminder that we're racing ahead of our ability to keep up. While AI models are getting better at tasks, they're also getting harder to measure—raising questions about how well we can assess their true capabilities and risks.

Who should care

  • AI researchers
  • Policy makers
  • Tech companies

What to do next

  1. Invest in long-term safety research
  2. Improve testing methodologies for real-world use
  3. Monitor AI development trends closely
  4. Engage with interdisciplinary teams to assess risks

Keep in mind

The analysis relies on current data, which may not capture the full scope of future AI development.

Read the original reporting at The Decoder ↗