发生了什么
Fortune reports that former OpenAI and Anthropic employee Jacob Coxon’s resignation and warnings about prompted renewed congressional activity. Within 48 hours, the report says, senators opened an investigation and sent letters concerning OpenAI, while other lawmakers convened a briefing and proposed legislation. Fortune also reports that Senators Ted Cruz, Amy Klobuchar and John Thune are backing a bipartisan AI safety bill, though no public text is available.
Fortune reports that Jacob Coxon, who previously worked at OpenAI and more recently at Anthropic, resigned this week while warning that leading AI companies were moving too quickly toward systems they could not control. The article says current Anthropic alignment researcher Evan Hubinger publicly supported Coxon’s concerns, assigned a greater than 10% chance to human extinction within a decade, and said Anthropic lacked a concrete control plan. Two other Anthropic researchers reportedly joined the criticism. These statements and the reported internal reactions have not been independently confirmed.
According to Fortune, the political response included Senator Josh Hawley opening an investigation into OpenAI, Senator Richard Blumenthal sending a similar letter, Senator Bernie Sanders convening a bipartisan briefing with Geoffrey Hinton, and Representative Ro Khanna proposing additional legislation. The report says Senators Ted Cruz, Amy Klobuchar and John Thune are now associated with a bipartisan bill that would give the Commerce Department and Homeland Security authority over the most powerful AI models.
Fortune describes the reported bill as requiring safety testing and incident reporting and allowing regulators to block a model’s release if they determine it presents a genuine catastrophic risk, including risks involving biological or nuclear weapons. The article says the text has not been seen outside Congress and that OpenAI and Anthropic are privately weighing in with Senate staff. The bill’s formal status, sponsors beyond those named, legislative language and likelihood of passage are not independently confirmed.
The proposal would reportedly replace or preempt state AI rules, including laws in California and New York, with a federal standard. Fortune places the development in a longer history of federal inaction, state-level experimentation and disputes over whether AI regulation should prioritize safety, competitiveness or limits on state authority. That historical account is the publication’s reporting and analysis, not an independent verification of every cited legislative detail.
为什么这很重要
If enacted, the reported bill could become the first federal framework with direct authority over the most powerful AI systems, including mandatory safety testing, incident reporting and the ability to delay a release judged to pose catastrophic risks. It could also displace state AI laws, concentrating regulatory authority in Washington. The proposal therefore matters both for and for the unresolved federal-versus-state question. Fortune’s account has not been independently confirmed here, and the bill’s scope, enforcement standards and prospects are unknown.
A federal safety regime with release-blocking authority would be a consequential change from the largely voluntary commitments, executive actions and state laws described by Fortune. It could affect how frontier developers evaluate models before deployment and how quickly high-capability systems reach customers.
Federal preemption would create a major tradeoff. A single national standard could reduce conflicting obligations for developers, but it could also remove state safeguards before the federal system’s coverage, staffing and enforcement are known. The source does not provide enough information to assess which state provisions would be displaced.
The practical impact depends on details that are currently unavailable: thresholds for covered models, the meaning of catastrophic risk, the independence of safety evaluators, reporting deadlines, confidential-information rules and judicial review. No public text means these cannot yet be evaluated.
The report also indicates that industry participation in drafting is already occurring, but it does not establish whether that participation is advisory, formal or decisive. The influence of OpenAI and Anthropic on the final proposal remains an important unknown.
互动机制：它实际上是如何运作的
以交互方式探索这一发展背后的基础技术。
Why can ethical evaluation not be reduced to one model score?
接下来看什么
The immediate test is whether lawmakers publish and formally introduce the bill. Key issues will include how regulators define catastrophic risk, which models are covered, what evidence is required before release, whether state laws would be preempted, and what appeals or oversight mechanisms developers receive. The positions of OpenAI, Anthropic and other developers, as well as any bipartisan agreement on the draft, remain unclear.
Watch for publication of the bill text, formal introduction, committee referral and a clear list of sponsors. Those steps would establish whether the reported proposal is an actual legislative vehicle rather than an early negotiating draft.
Examine whether the bill creates binding pre-release review or only reporting requirements, and whether Commerce and Homeland Security receive sufficient technical expertise, funding and independence to enforce it.
Track the preemption language closely. The source says state laws could be displaced, but it does not identify the exact provisions affected or whether states would retain authority over consumer protection, employment, privacy or sector-specific uses.
Monitor responses from AI developers, state officials, civil-society groups and technical safety researchers. Their positions may reveal whether the proposal is viewed as enforceable safety regulation, a competitiveness framework, or an effort to centralize authority.