AI Governance In 2026: Why Policy Moves Only After Incidents | Stephen Casper (Harvard University)

Stephen Casper on why AI governance has been stuck in the same conversation since 2023, and what kind of event would move it. A talk from the Cambridge AI Research Directions Workshop, convened by FAR.AI with the Cambridge Boston Alignment Initiative. Casper presents it as deliberate speculation rather than prediction. He frames the history in chapters. The first ran through the 2010s and early 2020s, when the computer science community did most of the thinking and had vague ideas about standards, evals and response. The second opened around 2023 and 2024 with concrete mechanisms: liability, whistleblower protections, incident reporting, transparency requirements. His argument is that we are still in it. Capabilities and risks have moved exponentially while the regulatory conversation has only evolved versions of itself, which he calls an AI governance overhang. The 1980 Collingridge dilemma explains why: early on you can shape a technology but do not understand it, and by the time you understand it the power structures are too entrenched to move. What breaks the pattern, empirically, is incidents. Lawsuits last year preceded an avalanche of child safety bills. Non-consensual AI nudification went from slow-burning trend to dinner table topic and then to law. The precedents outside AI are sharper still: nuclear construction and Chernobyl, the Patriot Act six weeks after 9/11, Sarbanes-Oxley eight months after Enron, a vaccine in trials five months after a virus was identified. Casper runs the audience through candidate triggers for a third chapter, takes a vote, and argues that the research worth doing now is whatever anticipates the disruption you think is coming. The final section sketches what chapter 3 might contain if AI proves genuinely dangerous: mandatory rather than voluntary audits, embedded government auditors, executive liability for misleading the public about risk, nationalization proposals including a bill from Bernie Sanders that would let a federal commission appoint up to half the board of large AI companies, licensing regimes, hardware restrictions, outright bans, and an IAEA-equivalent for AI. His closing argument is that without decisive action on transparency and on some power to pace the technology, the decisions will sit with a small number of people inside very large AI companies. Chapters 0:00 What this talk covers 0:33 Hockey stick curves, and the ritual around them 2:30 The Collingridge dilemma, from 1980 3:52 Chapter 1: governing AI through the 2010s 4:39 Chapter 2: concrete mechanisms arrive 5:00 Why we are still in chapter 2 5:41 What is currently on the table 7:38 What will probably keep happening 8:26 Why AI politics is incident-driven 9:45 Chernobyl, the Patriot Act, Sarbanes-Oxley and the vaccine 11:39 What might trigger chapter 3 13:55 Audience vote 16:02 What chapter 3 might look like 17:20 Nationalization proposals 18:58 Bans and international collaboration 20:18 The closing argument More AI safety research: https://far.ai FAR.AI is a research nonprofit working to ensure the safe development of advanced AI. We host the Alignment Workshop series and publish frontier alignment research.