How this book was made.

The Anti-AI Brain argues for the dosed use of AI, not its refusal. This page documents the dose. Every decision below was a deliberate choice, and every choice is on the record so you can evaluate the book by the same standard it asks you to apply to any other AI-touched artifact.

Research process

The book synthesizes findings from 100+ peer-reviewed studies. The core anchors are the MIT Media Lab EEG study (Kosmyna et al., 2025) on brain-activation during AI-assisted writing, Microsoft Research's 2025 "cognitive debt" paper on AI trust and critical-thinking erosion among knowledge workers, Sparrow et al.'s "Google Effect" (Science, 2011), Wolf's deep-reading work, van der Weel & van der Meer's handwriting neuroimaging, and research programs at Princeton and UPenn on attention, neuroplasticity, and learned helplessness. Every claim that rests on a specific study is sourced in VERIFICATION — Claims Evidence Table.md, which cross-indexes each statement in the manuscript to at least one primary reference. The full bibliography lives on the research page.

Fact-check methodology

Every numeric claim was verified against the primary source, not a secondary summary. Where a secondary summary was the only accessible version, that is flagged in the evidence table. Where a finding comes from a preprint rather than a peer-reviewed publication, that is also flagged. Where a result has been disputed or failed to replicate, the manuscript either acknowledges the dispute or omits the claim entirely. A separate revision pass (VERIFICATION — Revision Checklist.md) confirms that nothing in the finished manuscript relies on a contested or withdrawn result.

Which AI tools were used where

Claude Opus 4.6 and 4.7 (Anthropic) were used as sparring partners for argumentation — challenging the author's reasoning, proposing counter-examples, stress-testing chapter logic under Socratic, steelman, and falsification prompt modes. Opus was not used to generate prose that appears unedited in the final book. Every paragraph that survived into v11 was either written originally by the author or rewritten line by line after a model draft, voice-checked against earlier work to maintain authorial signature.

Google Gemini 3.1 Pro was used as a second, independent critic for each completed chapter — explicitly to catch the failure modes that a single-model workflow tends to miss (shared hallucinations, echoed framing, over-smoothed rhetoric). Every critique that produced a revision is preserved in GEMINI CRITIC PASS — vNN.md. No Gemini-generated prose appears in the manuscript.

Claude Cowork (Anthropic's desktop agent environment) was the build environment for the entire project — the manuscript editing loop, the illustration pipeline, the research table, and this launch site itself. The Cowork session transcripts are part of the evidence archive; every non-trivial author decision is logged with a timestamp, a prompt, and an unedited model response, so that any specific claim about how the book was made can be audited against source.

Image-generation models were used only for the illustrations (cover 3D render, tier art, patent-style briefs) under explicit creative direction, with final crops and compositing reviewed by the author. The current illustration pipeline runs on Google Nano Banana 2 (Gemini 3.1 Flash Image, released February 2026). No AI-generated photograph of a real person appears anywhere in the book.

What was never delegated: the book's central thesis, the five-layer model (Flesh, Archive, Lens, Loop, Voice), the Anti-AI Seven framework, the Ten-Minute Wall concept, the 30-day protocol structure, the tier-naming scheme, the voice, the ordering of chapters, and the editorial decisions about what to cut. These are author decisions on the record.

Human decision-process

The book went through eleven named manuscript versions between initial draft and publication. Versions v1, v2, and v3 each received a full-pass structural critique from an external super-critic prompt (preserved in SUPER CRITIC PROMPT.md). Two full reviews by the author on paper, with handwritten margin notes, happened between v8 and v11 — explicitly because the book argues for handwriting as a cognitive discipline and the author's final pass on the manuscript had to obey its own rules.

Chapter 1's opening sentence, the Cognitive Partisan / Outsourced Mind distinction, the Pharmakon framing, and the decision to name the four tiers of the quiz after specific archetypes rather than numeric bands — all of these were author decisions made early and held through every AI-assisted editing pass.

Illustration pipeline

Patent-style illustrations were generated on Google Nano Banana 2 (Gemini 3.1 Flash Image) from explicit prompt briefs preserved in ILLUSTRATIONS — Patent Style Briefs.md. Each brief specifies the diagrammatic subject, the historical visual reference, and the compositional constraints. The four quiz-tier illustrations were generated by the author and then selected from a larger candidate set; the Ghost tier was flipped horizontally in post-production to make the figure face the card text. Nothing photographic of a real person was AI-generated.

Why pharmakon

The ancient-Greek word pharmakon means both remedy and poison — the same substance, different doses. The book uses this as its central framing for AI: the problem is not the tool, it is the quantity, the timing, and the task. Writing the book while actively using the tools it analyzes is not an inconsistency; it is the thesis. A book about dosage must be written by someone who is still dosing.

If you find an error, an uncited claim, or a passage that reads as under-edited, please write to nik@nikmcfly.com. Corrections land in the next revision and are credited in the acknowledgments.