Global Treaty and Slowdown

The leaders of OpenAI, Google DeepMind, and Anthropic have signed a statement ranking extinction risk from AI with pandemics and nuclear war.¹ Yet the AI agreements signed so far rely on promises and self-reporting, not inspectors. Congress should fund the science of verification, open crisis channels with rivals, negotiate limits that can be checked, and set strict conditions for a time-limited domestic pause.

The Problem

The lab leaders' statement calls reducing AI extinction risk "a global priority alongside other societal-scale risks such as pandemics and nuclear war."¹ In 2023, 28 nations, the United States and China among them, warned of "serious, even catastrophic, harm" from frontier AI.²

Each side fears that if it slows, a rival will not. China's 2017 plan aims to make it "the world's primary AI innovation center" by 2030;³ Washington's 2025 AI Action Plan says the United States "is in a race to achieve global dominance."⁴ Only limits each side can check let both slow down safely.

Arms control shows what makes a promise hold. The 1987 INF Treaty eliminated 2,692 missiles under what President Reagan called "the most stringent verification regime in history."⁵ The Biological Weapons Convention has no inspectors; a suspicious state can only complain to the UN Security Council,⁶ and a 1979 Soviet anthrax outbreak blamed on tainted meat was later traced to a military facility.⁷ AI diplomacy so far looks like the bioweapons treaty, not the missile treaty. Three gaps stand out:

1. Pledges lack inspectors. The only binding AI treaty, the Council of Europe convention the United States signed, excludes national defense and most research and relies on a conference of its parties for oversight.⁸ The UN's new AI science panel issues "non-prescriptive" reports and leaves military AI out.⁹
2. Verification must be built. In 2001, the United States rejected a draft bioweapons inspection protocol that would "not improve our ability to verify" compliance and would put business secrets at risk.¹⁰ New START, whose 18 yearly inspections checked U.S. and Russian warheads, expired in February 2026.¹¹
3. Diplomacy is drifting. Summits that began at Bletchley with warnings of catastrophe reached Paris in February 2025 with a statement on "inclusive and sustainable" AI, signed by 63 countries including China, but not the United States or Britain.¹²

Why legislation: The President negotiates, but a treaty needs two-thirds of the Senate,¹³ and Congress controls the money, the verification agencies, and any domestic pause. Congress built this machinery once: the Arms Control and Disarmament Act of 1961 created an agency for arms-control negotiation and research, whose duties the State Department now carries.¹⁴ Washington and Beijing already agreed that humans must control decisions to use nuclear weapons.¹⁵ No federal law directs the government to develop ways to verify limits on AI. America negotiated arms control from strength; it should lead on AI the same way.

The Solution

A four-step staircase: each step stands alone, and each step up adds commitment. Scope: the most dangerous AI activities, such as the largest training runs or systems that could help make weapons of mass destruction. Export controls and incident records are addressed separately.

Step 1 — Build the tools to verify. Fund research on checking AI activity without exposing secrets: accounting for large training runs, chip features that confirm location or use, and confidential model audits. Require a yearly report to Congress on what can and cannot yet be verified. The rule Reagan quoted at the INF signing still applies: "trust, but verify."⁵

Step 2 — Open crisis channels. Seek shared definitions of serious AI incidents, protected government-to-government channels to report them, and common testing methods, starting with the United States, China, and the chipmaking nations. The 1963 Washington–Moscow hotline followed the Cuban missile crisis;¹⁶ an AI channel should come before one. Lock in human control over nuclear launch decisions first, already U.S. policy under the December 2024 defense law.¹⁷

Step 3 — Negotiate limits that can be checked. Pursue agreements on specified dangerous activities, with declarations, agreed access, confidential inspection, dispute procedures, and consequences for cheating. Protect unrelated research and trade secrets, and do not assume access to every foreign facility. Expand obligations only as verification proves itself.

Step 4 — Set the terms for a pause. Allow a time-limited domestic pause on a specified dangerous activity only if independent experts find a severe risk that less restrictive measures cannot control. The order names the activity, its length, and the evidence needed to lift it; courts review it promptly, and a broad pause lapses unless Congress renews it. A bipartisan House bill would let DHS order proportionate action, up to a shutdown, after a serious AI incident.¹⁸

Where to start: Step 1 is the floor; it binds no one. Step 3 is the heart.

Administration and enforcement: State leads negotiations, with Commerce, Energy, Defense, and the intelligence community supplying expertise. The executive delivers a negotiating and verification plan within 180 days. Binding commitments follow the treaty process, and any pause authority carries narrow triggers, short duration, prompt judicial review, and explicit appropriations.

Risks and Mitigations

• Cheating: Distributed training and hidden data centers may be hard to detect. Start with observable commitments and pilot verification; a comprehensive global slowdown may stay unenforceable.
• Handing rivals the lead: A pause that binds only America could reward less careful rivals and entrench today's leaders. Seek reciprocal limits first, and tie any unilateral pause to evidence, time limits, and independent review; weigh the costs of delay openly.
• Emergency power creep: Pause authority could outlive its purpose. Require specific findings, automatic expiration, congressional renewal, and court review; no agreement can override Americans' constitutional rights.

Similar Bills

Fit measures similarity to this proposal's mechanisms: High = direct precedent; Partial = useful component with material differences; Related = adjacent approach.

Federal

State

None. The Constitution forbids states to enter treaties,¹⁹ and we found no state bill seeking international AI limits.

What this adds: Existing law offers pieces: arms-control machinery, a nuclear human-control policy, and a proposed incident-triggered shutdown power. This proposal links them into a path from verification research to crisis channels, checkable limits, and a lawful, time-limited pause, so America can lead on restraint without betting on trust alone.