Making AI Headlines · Page 3

Safety Review Advances Beyond Admiring the Lobby

October 19, 2026 at 6:00 AM EDT · New York time

Anthropic has given the external evaluator METR employee-like access to assess risks in its AI models. The nonprofit has a team embedded in the company’s office, with access to hardware and internal systems, and can publish findings without Anthropic’s approval. That combination makes the arrangement more substantial than a carefully managed product demonstration. Outside evaluation matters because companies developing powerful systems also have strong incentives to release and promote them. Independent access can help test whether reassuring claims survive contact with the machinery, rather than merely with the public relations calendar. The central tradeoff is proximity. Evaluators need enough access to understand what they are examining, while independence must remain meaningful once they are inside. Permission to publish without company approval is therefore an important feature, not a decorative footnote. In ordinary office life, an employee-like visitor might receive a badge and directions to the coffee. Here, the useful privilege is being able to report that something deserves concern without first seeking permission from the people responsible for it. The arrangement establishes a better vantage point, not a guarantee that every risk has been resolved. In an industry fond of windows into the future, someone has finally been offered a view behind the desk.

Satire. Not factual reporting. Read the original reporting from the linked publishers.