A new essay argues that the AI alignment debate excludes the very people who will live with these systems. Researchers at labs like Anthropic conduct the conversation about what AI should do and how it should be evaluated, while the public remains absent from the room.

The safety and accelerationist camps appear to disagree fiercely—Yudkowsky calling for governments to "destroy a rogue datacenter by airstrike" while Andreessen dismisses critics as suffering from "ressentiment." Yet both share an assumption: that the participants in the debate are the ones doing the designing, and everyone else is what gets designed for.

The author contends that real alignment works differently. Rather than configuration—values installed into a system that receives them—the actual experience of using AI is mutual shaping, like sculpting wet clay where both hands adjust to each other. The labs' definition treats alignment as measurement from the human side, producing the closed-loop evaluation systems Anthropic itself described.

The work that matters now involves building, alongside others who are noticing the same gap, the kind of alignment the existing process cannot produce.