When an AI system causes real harm — a bad medical recommendation, a trading loss, a safety failure — liability is murky. The developer, the deployer, the user, or no one? Without clear liability, incentives for safety are weak. What liability regime would actually make AI safer rather than just creating legal theater?