Why the US and China Must Negotiate a Binding AI Safety Treaty
Source: "America and China must work together now to stop AI before it's too late | Fox News." October 2, 2026. www.foxnews.com
The Gist
Rep. Ro Khanna argues that AI is advancing so fast that it could soon escape human control entirely, potentially causing global catastrophe—and since a disaster like an AI-engineered pandemic wouldn't care about nationality, the US and China need to urgently negotiate a treaty with real verification and safety rules, much like Cold War leaders cooperated to avoid nuclear war despite their rivalry. He proposes five specific safeguards: pausing AI that can improve itself, pausing superintelligence research, requiring kill switches, mandatory safety testing before release, and mutual inspection systems so neither side can cheat.
Conclusion
The United States and China must urgently negotiate a binding, verifiable AI treaty—built on five pillars (pausing recursive self-improvement, pausing superintelligence development, mandating kill switches, requiring sandboxed testing, and establishing verification/inspection regimes)—because unchecked AI development poses a catastrophic, potentially existential risk to humanity that neither nation can safely ignore.
Premises
- The risk that humans will lose control of AI, or that misaligned models could unleash biological catastrophe, is increasing every day
- Recursive self-improvement (AI developing AI without human intervention) is the most likely pathway to humanity losing control of the technology entirely
- Superintelligence that exceeds human capabilities across all domains could unleash unpredictable, catastrophic consequences that neither America nor China is prepared for
- Kill switches embedded in advanced models would give humanity a fighting chance against rogue AI, even if their full efficacy is uncertain
- Pre-release sandbox testing would prevent criminals, terrorists, or misaligned AI agents from unleashing catastrophic cyberattacks
- Because China cannot simply be trusted at its word, any treaty requires mutually agreed verification technologies and inspection institutions (e.g., an International AI Inspection Agency)
- Historical precedent shows that fierce geopolitical rivals (the US and USSR during the Cold War) successfully cooperated to avert mutually catastrophic outcomes like nuclear war, demonstrating that cooperation is possible even amid competition
- A catastrophic AI event (e.g., an engineered pandemic) would not discriminate between American and Chinese citizens, meaning both nations share an existential stake in preventing it
Assumptions
- Advanced AI capable of recursive self-improvement or superintelligence is achievable in the near-to-medium term, not merely speculative
- China's government would be willing to negotiate in good faith and submit to external verification of its AI programs
- Verification technologies sufficient to detect cheating by either side can be developed and reliably implemented
- Kill switches and sandbox testing are technically feasible safeguards against advanced or rogue AI systems
- Policymakers and AI labs in both countries will prioritize long-term civilizational safety over short-term competitive or economic advantage
- Congressional and diplomatic pressure can meaningfully accelerate international treaty negotiations on a fast-moving technology