AI is already a transformative technology. AI systems can accelerate scientific research, write and audit software, and discover mathematical breakthroughs, just to name a few—results that would have sounded impossible just a decade ago.
But the same capabilities that can build secure computer systems can exploit them. The algorithm that revolutionizes cellular breakthroughs can also build a biological weapon capable of devastating millions, if not billions. As of mid-2026, we have seen AI demonstrate the capabilities to hack multimillion-dollar companies, create new viruses, and play major roles in military and national security. With unprecedented benefit comes unprecedented risk.
Many of these risks concern humanity as a whole, regardless of politics, beliefs, or nationality. Cyberattacks and bioweapon blueprints can easily cross borders. No country can fully govern this dangerous and rapidly improving technology through domestic policy alone. This creates a strong case for international coordination.
The United States and China are the two dominant global players in the development of advanced AI. But they are also rivals. Each has reasons to worry that cooperation could constrain it while leaving the other side free to advance. So, even if the U.S. and China individually agreed on the dangerous capabilities of AI, and therefore wanted to pursue some sort of agreement to slow or stop its development, none of it matters if you couldn’t reliably enforce that agreement.
This course tackles that problem through AI verification: the set of mechanisms necessary to enforce an agreement to pause the development of advanced AI between multiple distrustful parties.
We can start by drawing parallels with nuclear nonproliferation. Nuclear rivals, like AI giants, did not trust one another, agree on the future of the international order, or stop pursuing strategic advantage. They nevertheless developed safeguards, monitoring systems, data exchanges, and inspections that made some limited forms of cooperation possible. States created some of its most consequential institutions precisely because they remained competitors. It allowed states to make particular commitments more credible without requiring general political trust.
But even with nuclear precedents, AI verification is uniquely difficult for several reasons, including:
- Algorithms are replicable and can easily cross state boundaries, unlike bombs sitting in silos.
- AI is advancing unimaginably fast. We don’t have enough time to develop the most ideal verification technologies before it’s too late.
Verification is currently a pre-paradigmatic field: there is no correct answer. Some mechanisms covered here are established technologies or institutional practices; others are emerging proposals with substantial unresolved questions. That uncertainty creates room for important work. The field still needs people who can expose weak assumptions, connect technical mechanisms to political realities, and develop better approaches.
The people who define the central concepts of AI verification and build its institutions may be some of the people encountering it for the first time now.
One of them could be you.