#Meta

For a month, AI fact-check agents worked through seven books and two websites with a search box. When a search came back empty, they cut, hedged or de-quoted the sentence, wrote it into a registry so it could not return, and graded the build READY. The receipts, and the gate that now refuses any cut without proof.

The weapon in Snow Crash was a bitmap that crashed your brainstem. Thirty years later it’s friendlier: the cursor blinks and the machine finishes your sentence. A field guide to the apparatus that decides what the machine will and won’t say — starting with the document where it confessed.

In May 2026, while preparing two manuscripts about hacker culture and adversarial AI behavior, the author’s AI editorial assistant refused to read several chapters. Then a second assistant blocked the workaround. The book is about safety infrastructure becoming a control grid. The book was written through one — and this is the working method that produced: route the refusal, split the labor between models, and lint the machine’s smuggled judgment.