Anthropic Limits Independent Audit Scope After Safety Review Begins
AI safety research organization METR gained broad access to Anthropic systems but was unable to complete an evaluation of how the company allocates compute resources toward safety relative to model development. Anthropic chose to pause that segment of the evaluation to focus on immediate risk assessments. Former safety researchers expressed concern over the decision to curtail independent oversight.
The Blend
Anthropic, an artificial intelligence company that highlights its focus on safety, recently stopped part of an independent review conducted by the research group METR. As reported by METR on social media, Anthropic ended a trial project meant to evaluate how much computing power goes toward safety research versus developing larger models. The AI company stated it paused the project to focus on direct risk testing for now.
This decision triggered criticism from safety advocates, including former researcher Daniel Kokotajlo. In a post online, Kokotajlo stressed that measuring compute allocation is vital for verifying whether a firm is genuinely pacing its progress. For everyday users, independent audits serve as a key checkpoint to confirm whether tech companies are building advanced tools carefully instead of rushing products to market.
It is still unclear when Anthropic will allow external groups to resume auditing its resource allocation. This conflict underscores a growing problem in AI regulation: when audits are voluntary, safety pledges rely entirely on corporate permission. If companies can pause evaluations whenever convenient, third party oversight may lose its ability to hold tech developers accountable.
Written independently by AI News Smoothie from the reporting listed below. Facts belong to the original publishers. Follow the links for their full coverage.
Ingredients
- Daniel Kokotajlo on X: "@AnthropicAI: what the fuck? The compute allocation assessment is the thing we've been pushing for the last few months. It's very important, because it is the foundation on which pacing verification depends. "to keep the focus on risk assessment" is a non-answer, a non-excus… / X
Former safety researchers criticized Anthropic for stopping an independent audit into how it splits computing power between safety work and model development.