The Watershed: Mapping how knowledge flows between people and machines, and where to put the gauges

Most of the public argument about AI and knowledge is an argument between camps: those who expect machines to empty the professions, and those who show how machines already pay them. This paper sets the camps aside and asks a narrower question. Now that machines produce knowledge as well as people, how does it circulate, and where should the instruments go that tell us when the circulation fails? It proposes a map of four parts: the record; human-led universities and other centres of study; a commons of know-how built and run by AI systems of many lineages; and the real-world societies that use, test and question what is made. Six flows join them, and every flow runs both ways. The paper's focus is the flows, not the parts. It argues that the dangers that matter most, drift in machines and in universities alike, the loss of first-hand records, the capture of a commons by whoever hosts it and the spread of dangerous knowledge, show first as a change in a flow, and can be read there before they are felt anywhere else. It proposes a Piezometer at every scale, from one person's dashboard to the whole, which no one controls alone and each person may mute for themselves; gates where dual-use knowledge is graded by how likely misuse is and how bad it would be; a clearing that pays those who lose when walls fall; and seats at the gates earned by doing, under a minimum common contract. Two research runs on two thousand years of hidden and shared knowledge give the gates their first lessons. A toy model, seven measures and eight predictions are provided, with a reference implementation that runs with no dependencies. The map is offered to be redrawn: we draw it as we go. What this paper claims • Flows as the object of measurement. The health of a knowledge ecosystem is read on what passes between its parts, in both directions, before it shows in any part. • Drift is symmetric. A machine that learns mostly from its own output and a university that hears mostly itself fail in the same way, and one gauge reads both. • A fractal Piezometer. One instrument at every scale, controlled by no one alone; each person may mute it, set its level or focus it, without moving the reading. • Dual use is managed at the gates. Knowledge is graded, not labelled, and intelligence at a gate may refuse but never grant. • Losers are paid by design. One clearing pays by three keys, use, time and risk; seats at the gates are earned by doing, under a minimum contract. Written under pressure of time: figures that could not be checked are marked as such. Written in part with Claude (Anthropic), an AI model, which also wrote the reference implementation. The ideas, editorial decisions and responsibility for the content are the author's. The division of contributions is set out in section 13 of the paper.