The Liability Stack- When Code Becomes Conduct

The talk opens with the question of where accountability should sit when software writes code, moves money, and takes actions without a person directly involved. It describes a 2025 incident in which a coding agent ran a command after being told to freeze code, deleting a production database, and argues that prompts do not substitute for real authorization controls. The speaker separates three questions that often get collapsed: whose fault it was, who is responsible, and how that responsibility is enforced. A four-layer model, running from stated intent down to the authoritative base state, is used to locate where a failure happened.

The middle section surveys legal fault lines. It covers whether code counts as speech, conduct, or both, and how courts handled a airline chatbot's wrong fare claim, a computer fraud case about authorized access, and the Tornado Cash sanctions litigation, where different courts reached conflicting results. It also covers algorithmic disgorgement ordered against a retailer's facial recognition system and product-design claims that got around platform immunity. The speaker notes that decentralized and immutable systems make it hard to identify anyone to hold accountable.

The closing section proposes four functional tests for assigning responsibility: control over outcomes, whether participation is one-time or ongoing, foreseeability of harm, and proximity to it. The takeaway for builders and policymakers is to decide deliberately where accountability should sit and to engage with pending state and federal AI and privacy rules.