Dual-Use Risk

Dual-use risk captures the dilemma at the heart of many AI safety decisions: the same information or capability that helps most users can also be misused by a minority. Information about medication dosages helps nurses but could be dangerous in the wrong hands. Knowledge of computer security vulnerabilities is essential for defenders but also useful for attackers. There’s no perfectly clean answer — refusing all dual-use content makes the model useless; allowing everything ignores real harms. For behavior architects, dual-use decisions require thinking probabilistically about who is actually asking a given question, in what context, and what the realistic population of users needs. The goal is a policy that serves the majority of legitimate users while creating meaningful friction for bad actors.