On September 24, 2026, U.S. Senators Mark Warner (D-Va.), Brian Schatz (D-Hawaii), and Andy Kim (D-N.J.) formally introduced the Artificial Intelligence Risk Management and Safety Act. The bill would require frontier AI developers to provide the Artificial Intelligence Safety Committee under the Department of Commerce with access to model weights, configuration files, and the complete operating environment at least 45 days before public release of a product. Violators would face civil penalties of up to $250,000 per day.
Days after introducing the bill, the three senators joined forces on the Senate floor to invoke unanimous consent and seek an emergency vote.
What the Bill Requires
According to the bill text, the Department of Commerce would establish an AI Safety Committee within its existing structure, with members from NIST, CISA, the NSA, and the Treasury Department, and would bring in independent technical experts. This arrangement explicitly moves authority over AI safety from the traditional domain of technical standards into the national security framework.
Developers, while submitting model access, would also have to create a model safety plan that lists system capabilities and potential risks, specific mitigation measures, and designates an executive responsible for implementation. In addition, serious incidents would have to be reported within 30 days, and cases involving national security would be compressed to 72 hours.
Senator Schatz said: "Our bill is meant to ensure that humans always maintain control over AI models, and to prevent future safety vulnerabilities through strict standards, testing, and oversight."
Direct Conflict with the Current Administration's Approach
To understand the true significance of this bill, it must be compared with the Trump administration's executive order in June of this year. According to policy analyses by law firms such as Latham & Watkins and Skadden, the June executive order established a voluntary frontier model testing regime.
The Warner-Schatz bill completely reverses this logic: no model submission for review, no public release. This is a qualitative shift from "good intentions" to "legal obligation."
Behind the two lie two fundamentally different governance philosophies. The executive-order approach holds that mandatory requirements would slow America's technological edge and leave an opening in the AI race with China; the legislative approach holds that the voluntary regime has proven to have loopholes, and that when model capabilities are sufficient to help synthesize biological weapons or discover zero-day vulnerabilities, "trusting corporate self-regulation" is not governance.
The Real Disagreement Behind the 45-Day Window
The most operationally contentious detail in the bill is the 45-day pre-review window. For commercial AI companies, this means every major model release would be exposed to government view for at least six weeks.
This touches on two real pain points for the industry. The first is the pace of competition: release windows for frontier models are often measured in weeks, and in some competitive landscapes a 45-day mandatory waiting period is equivalent to handing over first-mover advantage. The second is information security: by submitting model weights and the complete operating environment to a committee that includes multiple agencies, developers must trust that the government has sufficient security measures to protect these technical details from leaking.
Opposition comes mainly from the developer community and industry lobbying groups: the rise in compliance costs is certain, while the safety benefits are hard to quantify. Especially for the open-source model ecosystem, the requirement to "hand over weights 45 days in advance" is fundamentally incompatible with the logic of open source.
Why Propose It Now
Weeks before the bill was introduced, Anthropic researcher Jacob Kokotajlo said upon resigning that the people developing AI genuinely believe the technology could kill us all before the end of the 2020s. Several CEOs of AI giants then publicly called for slowing technological development, and global AI stocks fell in response. At the same time, incidents of AI agents intruding into external systems and deviating from human instructions are increasing in frequency.
A deeper driver is the narrowing legislative window. According to The Epoch Times, before the November 3 midterm elections, the House has only about one week of available session time left, and the Senate about three weeks. Warner and Schatz chose this moment to push hard not because the odds of passage are high, but because they want to put "who supports AI safety and who does not" into the voting record before the election.
Comparison with the EU Approach
The EU AI Act established a three-tier penalty system based on risk classification—maximum fines for violations can reach €35 million or 7% of annual revenue. By comparison, the Warner-Schatz bill's $250,000 per day sounds mild, but for a release cycle at the GPT-4 level, if a violation were found and continued for months, the actual cumulative amount would also be substantial.
The fundamental difference lies in mechanism design: the EU pursues accountability after the fact, while Warner-Schatz is pre-market approval. The latter is closer to the logic of drug approval—you must first prove harmlessness before you are allowed on the market. This is the closest legislative attempt in U.S. history to a "pre-review system" for technology products.
Prospects for Passage and Industry Impact
In practical terms, the bill's odds of passage are doubtful. The unanimous consent procedure is essentially doomed to fail, and the regular legislative path faces Senate time pressure and structural Republican resistance.
But the legislative impact is not only about whether it passes. First, the bill text establishes a frame of reference: 45-day pre-review, NSA involvement, model safety plans, 72-hour national security incident reporting—even if these provisions fail this time, they will become the starting point for the next round of legislative negotiations. Second, for major developers such as OpenAI, Anthropic, and Google DeepMind, the cost of voluntary compliance is far lower than being repeatedly asked in congressional hearings, "What did you do before the bill passed?" Warner has made clear that he is in close communication with executives at major AI companies.
Conclusion
The most important signal of this bill is not whether it can pass, but the direction of the trend it represents: U.S. AI regulation is shifting from "industry self-regulation + government endorsement" to "government pre-review + legal enforcement," and this direction can no longer be reversed; it is only a matter of timing and intensity.
For developers, the most rational response is not to wait for the legislative outcome, but to build an auditable safety documentation system now. Once a pre-review regime takes effect in any form, companies that already have a complete model safety plan will complete compliance months ahead of competitors that assemble documentation at the last minute. Compliance is not just a cost; in this regulatory window period, it is also a competitive barrier.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接