Global pause

What it does

Every country would stop training new AI models more powerful than today's best, with no end date, until the safety problems are solved. Inspectors would watch chip supplies and data centers to catch anyone who cheats.

Applies to: Frontier training runs in every signatory country.

• Every country stops frontier AI training indefinitely, with verification.

Where things stand

Frontier training compute keeps growing about 4 to 5 times per year. Only voluntary company safety policies and U.S. export controls constrain it.

Why it costs ~5 months

America stops, and so does China, except for what it can hide. We assume it keeps about 30% of its progress. Net: about a month and a half a year in.

Biggest unknown: Whether China would truly stop and whether a pause could be verified while algorithmic research continues without big training runs.

Why it lowers p(doom) by ~1.3%

Stops frontier AI development everywhere, with no end date, until the world agrees it is safe to continue.

Stopping development everywhere is the most direct way to stop catastrophic AI. The catch is enforcement, plus the risk of a rush if the pause breaks down.

The strongest case that it costs more

A pause that breaks down could leave a hardware overhang and a sudden sprint. Cheaters would gain the most.

The debate

For

• Eliezer Yudkowsky TIME op-ed 'Pausing AI Developments Isn't Enough. We Need to Shut it All Down' (Mar 29, 2023): indefinite worldwide moratorium on large training runs, track all GPUs, enforce by force if needed.
• PauseAI (founded by Joep Meindertsma, 2023) Calls for a global treaty pausing training of the most powerful AI; held protests in London, Brussels, San Francisco and Washington in 2026.
• Sen. Chris Van Hollen, Sen. Alex Padilla and 15 other Democratic senators Letter (Sept 2026) urging Trump to negotiate AI guardrails with Xi and consider 'a mutual pause in frontier AI development', with technical verification.
• Future of Life Institute open letter (signers incl. Yoshua Bengio, Stuart Russell, Elon Musk, Steve Wozniak) March 2023 letter calling for at least a 6-month pause on training systems more powerful than GPT-4; 30,000+ signatures.

Against

• Yann LeCun and Andrew Ng 'Why the 6-month AI Pause is a Bad Idea' (April 2023): a pause is unrealistic and counterproductive.
• China Foreign Ministry spokesperson Guo Jiakun Sept 2026: called U.S. AI CEOs' slowdown calls 'fear-mongering' that would undermine global AI governance.
• Dario Amodei (Anthropic) Sept 2026 essay backs pacing but says it 'does not mean halting', warning that if the U.S. restrains itself and China defects, China could gain dominance.

Sources

• Dario Amodei: We Must Pace the Frontier (Sept 2026): Leading lab CEO arguing for slowdown, not pause; argues 1-2 extra years for alignment could greatly cut risk.
• Rest of World: China rejects Amodei call to slow AI: China's official rejection, Sept 2026.
• Padilla/Van Hollen letter press release: First congressional call for a mutual U.S.-China pause.
• Al Jazeera: As AI leaders warn of catastrophe, US and China shun slowdown calls (Sept 23, 2026): Both governments reject slowing; David Krueger calls for an indefinite international moratorium.