Public encyclopedias powered by AI may already be losing editorial control before they flip the switch.
Researchers working with a Nordic public knowledge institution ran design workshops to explore what happens when an encyclopedia adopts an LLM interface — and who actually controls what that system says. Their finding: pretrained models arrive pre-loaded with the values and editorial instincts of their commercial developers, not the institution deploying them. The team built a working LLM-enabled encyclopedia interface and introduced a framework they call "editorial alignment" — a process that treats the institution's editorial standards as a design artifact to be translated into concrete alignment objectives for the underlying model.
The stakes are higher than they look. Public knowledge institutions — libraries, national encyclopedias, civic archives — exist specifically to maintain editorial independence from commercial incentives. If the AI layer those institutions bolt on top is already aligned to OpenAI's or Google's dissemination priorities, the institution's editorial identity becomes decorative. This paper is one of the first to treat AI alignment not as a technical problem to be solved by engineers, but as an ongoing design process that editors themselves should participate in.
The practical gap is still wide. Design workshops are a long way from a deployable standard that any encyclopedia can adopt, and the paper focuses on a single institution in a specific regulatory context. Still, as governments and cultural bodies race to deploy AI-assisted knowledge tools, the question of whose editorial values get embedded — and whether non-commercial institutions get any say — deserves more attention than it is currently getting.