Part 8 — Nonfiction, research, and truth

Every factual sentence you publish asks the reader to take something on trust. Take three invented ones. The dam was finished in 1931. The senator changed her vote after that meeting. Most patients improve. The reader cannot visit the archive or telephone the senator. She can only decide whether to believe you, and she is likely to decide partly by feel, by the texture of the sourcing, the proportion of the doubt, and the way the writer treats opponents and the people in the pages. Nonfiction is a long run of such decisions.

The practices here lean toward the essay, the reported book, the history and the biography, but they apply wherever a text makes claims about the world or about people who exist, which includes the memoir, the novel set in a real city in a real decade, and the poem that names a living sister.

Read in order, the Part follows a project from gathering to publication. It can also be entered by problem. A pile of notes that has outrun your understanding needs 71, on how you know a thing, and 72, on how much checking it needs. For a thesis that has become suspiciously easy there are 73 and 74, on the claims you are eager to believe and the people who disagree with you. A draft that reads better than the material deserves is the subject of 75 and 76, on how arrangement can quietly assert things the record does not. A manuscript about to go to an editor wants 77 on quotation and 78 on the grammar of doubt; one in which anyone might recognize themselves wants 79 and 80, and those two are best read before the draft is finished.

Some writers research for a year before drafting a sentence and risk arriving at the page already committed to a reading. Others draft first, find out what they need to look up, and risk building on guesses. Neither is more rigorous, so long as every claim is checked before it leaves your hands.

An AI system can summarize a document you hand it, suggest where to look, propose objections and draft interview questions. It can also produce citations to papers no one wrote, quotations no one spoke, and biographical details that fit a plausible pattern. The principle that runs through this book, that the author remains responsible for the published Work, becomes in this Part a rule without exceptions. Any claim, citation, quotation or biographical detail that comes from an AI tool is a lead, usable only once you have confirmed it in a source outside the tool. The chapters take up AI tools only where a tool changes their own problem: how to rank claims for checking (72), agreement that flatters you (73), invented quotations (77), and private details pasted into a service (80).

Defamation, privacy and copyright law differ by country, by state and by form, and a book of practices cannot tell you what a court would do. Where the risk of publication is serious, get a lawyer who knows your jurisdiction. That lawyer will not answer the ethical questions in the last two chapters, either.