A Note on AI Assistance

Every chapter of this book was researched, written, and rewritten by me. I also used AI tools throughout. What follows is more than the disclosure MIT Press requires of authors who use AI. It is a statement of the standard I held the book to, and a claim about what authorship means when the tools are this capable.

The standard is traceability. Derived quantities in these chapters are backed by Python computation cells whose source code ships with the book. Citations were checked against their sources, and cross-references are checked against stable anchors. The purpose-built infrastructure modeling engine used for the book’s system calculations is available at https://mlsysbook.ai/mlsysim. A pre-commit test suite checks the machine-testable parts of these invariants on every commit, covering unit consistency, inline-reference resolution, cross-reference integrity, notation canonicality, and citation hygiene. AI tools made parts of this rigor mechanically tractable. The enforcement infrastructure makes that work auditable.

Authorship is thinking. Which concepts belong in this book and which do not. How to sequence the chapters so that each idea arrives only after the reader has the tools to understand it. Which worked examples to trace end-to-end, and which numbers to compute so the reader can verify the argument independently. What level to pitch each explanation at, rigorous enough for a graduate engineer and concrete enough for a student encountering systems thinking for the first time. Where students will get stuck, because I have taught this material and watched them get stuck. These are judgment calls that no tool can make, because they require knowing the reader. I used AI tools throughout this process to brainstorm framings, explore alternatives, and pressure-test my reasoning. The choices are mine.

In practice, AI was genuinely useful for writing code: the computation cells, the pre-commit checks, and the worked examples that are essentially programs. It helped me survey literature I then read in the original, draft passages I then rewrote substantially, and audit the manuscript for consistency across hundreds of cross-references, thousands of index entries, and tens of thousands of lines of prose. Throughout, the tools proposed and I ratified.

This pattern is older than the present moment. In 1683 Joseph Moxon printed the first English-language manual on printing, Mechanick Exercises on the Whole Art of Printing, using the press it described. Three centuries later Donald Knuth wrote The TeXbook in TeX, the typesetting system he created after becoming dissatisfied with the typesetting of his own books. A book about ML systems infrastructure, written without the infrastructure it describes, would be a contradiction. The toolmaker documents the tool with the tool.

Introduction describes a verification gap: ML systems can not be certified the way traditional logic can. That gap applies to AI-assisted authoring with equal force. The standard I held this book to is the standard I want us to demand of every system we ship: the tool can propose, but only an engineer can ratify. MIT Press does not list AI tools as authors because they can not assume ethical and legal responsibility for their work. The same is true of every engineered system. Authorship is the assumption of that responsibility, and it belongs to the human.