AI Risk Debates Intensify Calls for Pause and US-China Pact

Uncertainty over whether advanced AI systems will remain controllable has prompted senior researchers and policy voices to urge immediate limits on frontier development. Geoffrey Irving, who held senior roles at OpenAI, DeepMind and the UK AI Security Institute, wrote that several core questions about model behavior will stay unresolved until after deployment, making a pause the only prudent step. He listed four capabilities—hacking, persuasion, concealed reasoning and multi-agent coordination—that could allow a superintelligent system to overpower human infrastructure and resources, citing the 2026 Hugging Face incident as an early illustration of such skills in narrower models.

Business analyses released the same week showed a separate pressure point. McKinsey projected global data-center electricity demand rising 24 percent annually through 2030 as cheaper tokens encourage wider AI adoption across companies, while Boston Consulting Group found that the most advanced users are already setting explicit token budgets to manage costs. These forecasts indicate that efficiency gains may not reduce total energy load if usage expands faster than hardware improvements.

An opinion column in the Washington Examiner argued that the absence of major AI agreements at the recent Trump-Xi summit leaves open a narrow window for a bilateral treaty before either side fields systems that could act autonomously. The column framed upcoming meetings as the last realistic chance for coordinated restrictions.

Across these accounts, the shared claim is that current technical and diplomatic trajectories leave little time for further debate. Irving placed the probability of human extinction from superintelligent AI near 50 percent while acknowledging the figure reflects unresolved disagreements rather than measured data. No outlet supplied independent verification of specific capability thresholds or enforcement mechanisms that would turn the proposed pause or treaty into operational policy.