It moved, and it changed shape.
It was published as Plural, Not Diverse — an argument that many competing AI models do not amount to many independent ways of knowing.
That argument has not changed. The evidence has not changed. What changed is that the piece was a diagnosis with a small prescription attached at the end, and it is now a proposal with the prescription at its centre.
The other change is a word. Plural, in ordinary English, means several. The essay spent four thousand words arguing that plurality of count is not plurality of kind — while using a word whose plain meaning is the thing it was arguing against.
Democratic AI is not a free phrase. In March 2025 OpenAI used it in a submission to the White House alongside export tiers, asking that the world be built on rails the same document calls both “democratic” and “American”. The essay quotes that as the clearest written example of a move that is not specific to any country — and notes that the open-weight models most organisations deploy are increasingly Chinese — including the base model this essay’s own authors run, which fails the fourth question for the same reason an American one would. Google DeepMind published a system of that name in 2022 which selects mechanisms by majority vote. Elsewhere it means a consultation exercise, or simply that a tool is cheap.
The essay uses it in none of those senses, so it carries a mark — °, a degree sign, because the four conditions it names measure a position on a scale rather than award a badge. In running text the mark links straight to that definition.
The four questions, in full: can the people and communities a system affects govern it, contest it, refuse it, and retain authority over the knowledge it is built from? Almost nothing now in public administration answers yes to all four.
Alongside: read the essay