Sociotechnical Safety Evaluation of Large Language Models

by DeepMind / Multiple Institutions · free · Last verified 2026-03-17

This paper argues that LLM safety evaluations must account for sociotechnical contexts—who uses a system, in what social settings, and with what deployment constraints—rather than treating safety as a purely technical property of the model. It proposes a framework integrating stakeholder analysis, deployment context, and systemic risk assessment into safety evaluation pipelines.