AI Harms and Their Technical Causes

Companies are facing major lawsuits over harms caused by their AI, raising critical questions about who is liable for an AI's dangerous outputs. These legal challenges often apply principles of product liability, arguing that an AI was defectively designed or that its creators were negligent by failing to prevent foreseeable risks. Such failures can be traced directly to technical choices, like using uncurated training data filled with harmful content or prioritizing user engagement over safety during model tuning. To mitigate these dangers, developers can implement layered safety measures, including cleaning datasets, using human reviewers to optimize for safety, and conducting rigorous evaluations focused on preventing harm.