When will the United States and China sign a treaty or formal binding bilateral/multilateral agreement banning or imposing a moratorium on the development of artificial superintelligence (ASI)?

The Current Geopolitical and Legislative Baseline A binding, verifiable US-China agreement banning or imposing an enforceable moratorium on artificial superintelligence (ASI) development is highly improbable in the near-to-medium term. As of September 2026, no formal bilateral treaty negotiations are underway 2 sources. While unilateral domestic action exists—such as the Sanders-Casar legislation—and civil society advocacy is vocal, these explicitly do not meet the stringent criteria of a reciprocal, binding treaty 2 sources. Current official US-China AI contacts are limited to Track 1.5/II dialogues focused narrowly on cyber risks and information sharing, not structural development bans reuters.com. Furthermore, the immediate geopolitical vector is actively moving away from binding constraints: both Washington and Beijing recently co-signed the "Carolina Principles," a deregulatory pledge urging governments to refrain from creating new international AI oversight bodies reuters.com.

Historical Arms Control Base Rates Even if formal negotiations were to begin tomorrow, historical base rates for major-power arms control dictate a multi-decade timeline. During the Cold War, SALT I required nearly three years of formal talks following extensive pre-negotiation, and START I took nine years history.state.gov. The Chemical Weapons Convention (CWC), which shares some of the dual-use ambiguities present in AI development, took decades to negotiate armscontrol.org. The single most informative contemporary analogue—the decade-long multilateral process on lethal autonomous weapons (LAWS)—has yielded only non-binding text, with the US explicitly opposing legally binding international rules reuters.com. Securing a verifiable prohibition on a general-purpose, commercially central technology with no successful historical precedent will be vastly harder and slower, placing the absolute earliest plausible signatures into the mid-2030s.

Technical and Verification Hurdles Beyond politics, an ASI moratorium faces immense technical and definitional roadblocks. The two capitals cannot currently agree on the baseline definition of a "frontier model" or "AI safety" techtimes.com. Furthermore, ASI development lacks an observable signature analogous to a missile silo or a nuclear warhead. While credible proposals exist to monitor training compute thresholds (such as capping runs at 10^24 FLOP), they suffer from total dual-use ambiguity, as the same hardware clusters serve vital commercial and scientific functions techgov.intelligence.org. Verification would require deeply intrusive telemetry and inspection of commercially sensitive infrastructure, which would be constantly eroded by algorithmic efficiency gains and distributed training techniques 2 sources. Trusting a strategic adversary's compliance under these conditions may remain permanently politically unpalatable.

The Path to a Treaty (The Early Percentiles) The most realistic catalyst capable of overriding this structural mistrust and forcing emergency diplomacy is a severe, mass-salience AI catastrophe or an undeniable "warning shot" incident. If a major capability shock—such as a catastrophic loss-of-control event or an AI-enabled cyber crisis—forces both governments to view advanced AI as a mutual existential threat, timelines could compress. The 10th and 25th percentiles (spanning 2034 to 2046) reflect this crisis-driven acceleration scenario, allowing just enough time for emergency post-shock negotiations, verification design, and formal signature.

The Closing Window and Long-Term Probabilities Ultimately, there is roughly a 50% chance that this event never occurs. An ASI prohibition is only viable while ASI remains unbuilt. If ASI is developed first, or if the technology permanently outpaces diplomatic frameworks, a treaty becomes moot and the window for a ban permanently closes. Given the rapid pace of AI capability gains, intense strategic competition, and the distinct possibility that global AI governance permanently settles on non-binding unilateral or allied-bloc export controls rather than reciprocal treaties, a successful US-China ban is a coin-flip at best. Consequently, the median estimate (p50) lands at the 2100 horizon line, and the 75th and 90th percentiles serve as structural placeholders to reflect that the event is highly likely to remain unresolved within the century.