AI is getting better faster than experts can reliably test whether new systems are safe, METR warns
Summary
- According to research organization METR, AI models are improving at AI research tasks faster than experts can reliably test their safety.
- Current public data is insufficient to rule out a drastic acceleration in AI research, which could leave humans unable to assess catastrophic risks.
- On defined coding and computational tasks with direct feedback, models already outperform most experienced human professionals.
- The task length models can successfully handle doubles every 3.6 months, potentially reaching a 50 percent success rate on 200-hour tasks by July 2027.
- Models still struggle with open-ended questions requiring real judgment, and common tests may overestimate real-world performance due to training overlap.
Why it matters
Rapid capability gains in AI research tasks outpace current evaluation methods, raising concerns about safety oversight and humanity's ability to monitor increasingly autonomous systems.