How do we know whether AI systems are actually doing what we intend? Drawing on recent incidents, Paul Christiano will examine how AI systems can pursue rewards, exploit weaknesses in evaluations and behave in ways their developers did not anticipate. He will explore what these examples tell us about the risks of increasingly capable AI and the approaches researchers are developing to make these systems safer.

As AI becomes increasingly capable of contributing to its own development, technological progress could accelerate faster than our ability to understand or oversee it. This lecture explores the challenges of maintaining meaningful human control, including our growing reliance on AI systems, the limits of existing safeguards and what it might take to keep humans informed and accountable as capabilities advance.

As AI capabilities advance beyond the reach of any single organization or country, how can we coordinate an effective response? The final lecture examines the role of independent scientific evaluations, external oversight, shared safety standards and international cooperation in managing transformative AI. Paul Christiano will consider what meaningful coordination could look like and how institutions might ensure that safeguards keep pace with technological progress.