1. United States
2. Pa.
3. Letter

On September 9, Axios reported that three researchers at Anthropic stated publicly that they believe advanced AI could kill all humans within this decade. The company's alignment-science lead put his personal estimate above 10% and acknowledged that Anthropic has no plan yet for controlling systems more capable than humans. A colleague resigned in order to say so publicly. OpenAI's leadership made similar comments days earlier. I am asking you to test these claims. Critics argue the labs benefit from both the alarm and from rules that entrench incumbents. If that is what is happening, a public record should establish it. If it is not, Congress is leaving a civilization-scale decision to a handful of private firms whose own employees say they cannot control what they are building. I am asking for four things. Hearings. Call these researchers and the CEOs of the leading labs to testify under oath about their internal risk estimates and the evidence behind them. Independent pre-deployment testing. Models above a defined capability threshold should be evaluated by third parties before public release, with results going to a federal regulator rather than to the developer's own board. Mandatory incident reporting. Dangerous capabilities, evaluation failures, and loss-of-control incidents should be reported to the government within a fixed window, as we already require of aviation and nuclear operators. Whistleblower protection. No researcher should have to quit in order to warn the public. Employees at AI developers need statutory protection when raising safety concerns, including protection against non-disparagement and equity-forfeiture clauses. These companies say they want oversight. Take them up on it. Hearings cost almost nothing. If the people building this technology are right about what they are building, the cost of not holding them is everything else.