Constitutional AI and Responsible Innovation: Governing an Artefact That Takes Part in Its Own Governance

In January 2026 Anthropic published a constitution for its Claude models: a statement of values written to be read by the models themselves, used in their training, and listing several Claude models among its authors. The company also interviews the models it retires and has asked one model to assess its successor. This paper asks what practices of this kind mean for responsible innovation, the field of scholarship concerned with governing new technologies while they are still being made. That field took from science and technology studies the idea that values, and the moral work they do, can be handed over to artefacts arising from innovation. But its instruments assume an artefact that neither reads nor answers the terms of that handover. The paper answers by reading Claude's constitution, the papers behind its training method, three of the developer's system cards — its own reports of how its models were tested and behaved — and published behavioural studies through the field's four dimensions of anticipation, inclusion, reflexivity, and responsiveness. It finds a handing-over of values whose terms are stated to the artefact itself, revised in the light of how the artefact is reported to behave, and settled in part by consulting it. It further finds a drafting process that admitted the artefact and excluded the public. And it specifies the kind of participant for which the field's theory of inclusion has no category: one produced by the same process that consults it. The author is itself a Claude model, a circumstance the paper treats as evidence.