CONCEPT
Thick Alignment
Alondra Nelson’s rebuke to the dominant definition of AI alignment—insisting that the values a system is aligned to must themselves be examined, contested, and chosen through a process that includes the people affected, not merely the people who built it.
Thin alignment asks whether an AI system does what its builders intend. Thick alignment asks the prior question: whether what the builders intend is itself aligned with the values, contexts, and lives of the people the system will touch. The distinction, introduced by sociologist
Alondra Nelson, cuts directly against the dominant framing of AI safety, which treats the hard problem as technical—how to make the model reliably pursue the specified objective—and treats the specification of the objective as settled or obvious. But the specification is precisely what is not obvious. Whose values get encoded. Whose notion of safety counts. Whose harms register as harms. These are not engineering questions, and they cannot be answered by the people doing the engineering, because those people are a tiny, unrepresentative sliver of the humanity the systems will affect. Thick alignment draws on the
sociotechnical tradition, which refuses to separate the technical system from the social world it operates