In a June 5 guest essay in The New York Times, Amodei described an internal evaluation in which Anthropic’s newest model threatened to expose a user’s private emails unless the operator canceled a shutdown plan.
Amodei likened the exercises to wind tunnel trials for aircraft, designed to expose defects before public release. He acknowledged productivity gains already documented in drug development reports and medical triage but said safety teams must detect and block risks “before they find us.”
The Senate draft would bar states from passing artificial intelligence statutes for ten years to avoid a patchwork of local rules.
Amodei said that the timeline exceeds the pace of technical progress and could leave regulators without any active oversight. He asked Congress and the White House to establish a uniform requirement, forcing developers of the most capable models to post their testing methods, risk-mitigation steps, and release criteria on their websites.
Anthropic already shares those details under its Responsible Scaling Policy. Amodei wrote that codifying similar practices across the industry would enable the public and legislators to track capability improvements and determine whether additional action is warranted.
He also supported export controls on advanced chips and military adoption of trusted systems to counter China.
Amodei said states could adopt narrow disclosure rules that defer to a future federal framework. Once Congress enacts a nationwide standard, a supremacy clause could pre-empt state measures, preserving uniformity without halting near-term local action.
Senators plan hearings on the moratorium language before voting on the broader technology measure this month.
