Who Holds the Switch?

n June 12 of this year, the United States government pressed a switch. Two of Anthropic's newest AI models were suspended by an export-control order. Eighteen days later, the controls were lifted. Right away, three questions appeared. Who gets to press the switch? Who can check that it should be pressed? And when something goes wrong, who answers? There is a word for those three questions: governance. In this episode we go back a few thousand years, because people were asking them long before anyone built a computer. The Torah starts with judges and a rule against bribes. It limits even the king. It separates a loss nobody saw coming from a danger the owner knew about and did not guard against: the ox that was known to gore, the uncovered pit, the fence on the roof. Once the danger is known, you answer for it. But there is a hard problem. With a new technology, nobody knows yet what a fall looks like. We look at the Collingridge dilemma and at Norbert Wiener's warning from 1960. Then three stories from this month's news, held up against the three questions: Who can check? A researcher at Goodfire says a mysterious part of an AI that seemed to do two unrelated jobs may be doing one simple job. He says these systems may be easier to read than we feared. Others, including the head of Google DeepMind's interpretability team, disagree. Who verifies the check? California's governor has asked for recommendations on whether a "kill switch" for frontier AI could work, verified by an independent organization. It is not an order to build one. It is a question about whether it can be done. Who answers? Character.AI's parent company has filed for bankruptcy while facing lawsuits and state investigations, and its private chats may be part of what is sold. The lawsuits are allegations, not findings. We end with a story from the Talmud about a made being, a question that got no answer, and a maker everyone could name. We do not need the machine to be good. We need someone to be answerable. Topics covered: • The June 12 export-control order and the 18 days that followed • Governance: who decides, who checks, who answers • Deuteronomy 16 and 17: judges, bribes, and a king who carries the Teaching with him • Exodus 21 and Deuteronomy 22: the ox, the pit, and the roof • Maimonides on removing any danger to life • The Collingridge dilemma and Norbert Wiener • Interpretability explained simply, and the Goodfire debate • California Executive Order N-9-26 • Character.AI's bankruptcy and the question of who answers • Sanhedrin 65b: Rava, Rabbi Zeira, and the being who did not answer Still Human Weekly is a podcast about what it means to be human in the age of artificial intelligence. Hosted by Zalman Ollech. New episodes every week.