Charles Foster outlines METR's proposed assessments of alignmentAlignment is the effort to make an AI system pursue intended goals and follow relevant human values, instructions, and constraints., control and AI research acceleration. The questions concern evidence of misalignment in internal use and training, whether monitoring covers agents robustly, and how much AI is accelerating further AI development. These are assessment proposals, not a claim that current methods can prove a model safe.
Charles Foster discusses testing explicit safety claims by searching for counterexamples and examining logsAI agent observability makes an agent's state, actions, tool use, failures, resource use, and outcomes visible enough to understand and operate it.. He distinguishes a lack of observed misalignment from positive proof of alignment, and describes how monitoring failures and unwanted agent behavior can combine. Reported security incidents motivate the discussion rather than serving as independently reconstructed evidence in this interview.
Charles Foster emphasizes publishing the underlying evidence and explaining the assessment process. Access agreements, restrictions, redactions and conflicts of interest should be visible so readers understand both the findings and what remains unavailable. He describes independence and institutional neutrality as goals requiring disclosure, not as guarantees that evaluators have no limitations.
Charles Foster closes by discussing the skills needed across an expanding evaluation ecosystem: truth-seeking, adaptable methods, technical familiarity with agent systems and organizational analysis. Theo Jaffee asks how outside assessors can obtain meaningful access while remaining independentIndependent AI evaluation tests an AI system through reviewers who are meaningfully separate from the team that built it. of the companies they study.
Watch on YouTube
