Safety & ethics
AI safety is the field devoted to making AI systems behave reliably, avoid causing harm, and remain under meaningful human control as they grow more capable.
Any powerful technology needs safety work. For cars it meant seat belts and crash tests. For AI it means making sure systems do what we intend. They should refuse dangerous requests, not deceive people, and fail in predictable ways. The field spans today's practical problems and longer-term worries about very capable systems.
Near-term safety covers things like stopping chatbots from giving weapon instructions, preventing biased decisions, and keeping agents from taking harmful actions. Longer-term research asks a harder question: how do we keep control of systems that may become smarter than their designers? It also asks how to check what a model is really doing inside.
Safety work uses training methods like RLHF, testing by red teams who try to break the model, guardrails around deployed systems, and, increasingly, government rules. It is an active, unsettled field. Experts disagree about how big the risks are, but they broadly agree the work matters.
Before releasing a new model, a company has testers spend weeks trying to make it help with cyberattacks, then fixes every weakness they find.