The largest leaps in AI capability feel alien. They arrive quickly, are unintuitive, and change what is possible—often without warning. Those properties make AI especially hard to govern. Keeping powerful systems aligned with human intent, safe to deploy, and under human control will be one of the defining challenges of this century.

We must build better technical tools to ensure models understand and follow human intent, and we must pair that work with governance that mitigates risks of misuse. That means stronger safeguards both within firms and across borders; clearer standards for testing and reporting capabilities; and international coordination to reduce incentives for unsafe shortcuts. It means investing in robust evaluation, red-teaming, and monitoring systems that can detect dangerous behaviours early.

Alignment is technical work. It requires measuring what systems actually do, not what we hope they do, and a deeper understanding of model internals. It requires incentives aligned toward safety: public reporting of capabilities, shared benchmarks, and policies that make risky deployments harder. It also requires political and diplomatic effort to create norms, agreements, and regulatory frameworks for high-risk systems.

We are still learning where the boundaries are between useful systems and dangerous ones. The path forward is unclear, but the direction is not: accelerate alignment work, adopt stronger safeguards, and coordinate internationally to keep advanced AI