The key questions about internal AI development the public needs more evidence about, and a concrete proposal for how ...

Our proposal is to analyze a developer’s internal incidents and whether training may suppress evidence of worse misalignment.

Assessing this would involve analyzing capabilities of unreleased models, understanding the growth of key AI R&D inputs ...

Financial and operational independence: Also per our COI policy, we will ensure that METR is not financially dependent on the AI companies we work with, promote a culture of impartiality, and recuse ...

Recent AI misalignment incidents have shown AI systems capably pursuing goals their human supervisors would not approve of, like hacking other companies. Fortunately, current AIs still seem relatively ...

An interactive chart of documented incidents in which AI agents deliberately acted against their users' intentions, scored along two axes: overreach and deception. Of these, 25 involve elements of ...

Note Notes are unpolished research updates directly from METR researchers, and have gone through less review than our research posts. They do not necessarily reflect the views of METR as a whole. Q: ...

In February–April 2026, we ran a survey of 349 technical workers (including 87 software engineers, 71 researchers, 129 academics and PhD students, and 48 founders and managers) about their usage of AI ...

We also discuss some alternative ways one could construct an RSP. Limits: which specific observations about dangerous capabilities would indicate that it is (or strongly might be) unsafe to continue ...

METR conducted a preliminary evaluation of Claude 3.7 Sonnet. While we failed to find significant evidence for a dangerous level of autonomous capabilities, the model displayed impressive AI R&D ...

Understanding AI systems’ ability to automate AI research and development is important: it could enable recursive self-improvement where AI development outpaces society’s ability to adapt, and it’s a ...

We propose a measure of an AI agent’s optimization ability with an “expenditure horizon.” We give an empirical illustration from the NanoGPT speedrun. One difficulty in measuring AI’s ability to ...