A book written with AI where every claim carries a truth label, reviewed ruthlessly by rival models

__hymn · reddit · 2026-08-31

The author just published an AI-assisted book built around a rule: every load-bearing claim carries a visible marker — BENCHMARK (externally established, cited), DISCLOSED (from the author's own work, provenance not proof), ASSUMPTION (proposal only), GAP (genuinely unknown), and REPORTED (a source turned out wrong; the correction is printed on the page).
Two REPORTED flags target the author themself: an earlier paper calling p=0.04 significant against its own 0.01 threshold, and a source claiming 12,847 words that measures 5,335.
The draft was also sent to rival labs' AI models with one instruction: be ruthless. Their named corrections are in the finished book; one model's epistemology critique became a front-matter list of the seven assumptions the argument depends on.
Key lesson: AI collaboration didn't make the book more persuasive — persuasive was the failure mode. Pointing every model at claims rather than prose bought checkability. The marker system is the part worth stealing for any trust-critical LLM work.

More from AGI Musings