Hard Choices in Artificial Intelligence

Mathematical formalism alone cannot solve AI safety problems. The paper argues that technical solutions are insufficient for addressing safety and ethical concerns in AI systems—instead, decisions about AI safety require deliberation about politics and the specific contexts where these systems will be deployed.

Vagueness in AI safety debates reflects genuine design challenges, not just unclear thinking. Rather than treating disagreements about AI safety as problems to eliminate, the paper reframes "vagueness" as pointing to distinct, concrete design challenges at different stages of AI development that require explicit choices.

Stakeholder engagement should be central to AI safety, not peripheral. The paper emphasizes that creating feedback channels with diverse stakeholders is essential for exhaustively identifying and addressing safety issues, suggesting that AI developers need input from affected communities, not just technical experts.

AI safety decisions are inherently political and require democratic deliberation. The paper argues that determining what makes an AI system "safe" involves value judgments about how society should function, making deliberation itself—not just the procedures that follow from it—the actual goal of responsible AI development.

Hard Choices in AI require identifying overlaps between design decisions and broader social challenges. The framework helps developers recognize when technical choices about AI systems intersect with major sociotechnical issues, enabling more comprehensive consideration of consequences before deployment.