AI exposition

Guiding an AI to write for a human reader

Index of the ai-exposition project: what is known about where AI-drafted explanatory writing fails a human reader, what guidance has been shown to fix it, and what is still to be tested.

An AI draft of a report, a vignette or a proof arrives with the spelling, the grammar and the formatting close to flawless, and with the emphasis in the wrong places: pages on the routine steps, a paragraph on the step that decided the result, no comparison with what was already published, and no overview that lets a reader decide whether to read on. This project asks what instructions move the emphasis back to where a human reader needs it, and how to tell when an instruction has worked. Behind that sits a second question, on what happens to a field once producing the result is cheap and reading it is not; Tao asks it for mathematics, and pharmacometrics runs parallel to it stage for stage.

Two starting points. The writing guide in this repository is one: a set of rules derived from one reader’s edits of AI drafts, already in use. Terence Tao’s account of AI proof exposition is the other: the same failures, observed in a different field by a different reader.

Documents

• Working specification. The question, the failures named so far and their form in a modeling report, what the writing guide already covers, the instructions to test, and how to test them.
• References. The sources, the reading queue, and a marker on every entry recording whether the claim drawn from it has been checked against the source.