Draft complete but not yet reviewed; expect revisions

Preface

My previous book, on Bayesian optimization, ended with a question it could not settle: what a human preference is, and whether a system that asks a person questions can find it or only shapes it. Alignment research asks the same question at a larger scale, about a learner that may come to outgrow the person it learns from. That is where this book begins.

It has two aims. The first is a systematic account of the field, from the optimization problem at its core to the institutions around the developers. The second is to read that field beside the disciplines that studied delegation long before machine learning: economics, law, organization studies, political science, philosophy, biology, safety engineering, and security. Those fields cannot answer alignment's questions for it, but they can say which of its problems are old, which are new, and where an analogy stops.

What the book argues #

The book treats alignment as four stacked problems: specification, saying what we want; instillation, getting a learner to want it rather than its measurement; verification, checking whether it did; and maintenance under pressure, keeping it true as capability grows and the system is deployed among people, adversaries, and other systems. The parts follow the four problems in that order. The safety papers three frontier laboratories published from 2022 to mid-2024 were mostly on the second and third (Section 3.1.3; Part IV; Part V; Part VI; Part VII). The first and fourth have long literatures under other names, which the book brings into the parts where they bear (Part III; Part VIII; Part IX; Part X).

Its claim is that alignment is a delegation problem under an asymmetry of competence and visibility, and that what is new is the opacity of the agent, its speed and scale, and the possibility that it outgrows its principal's ability to check it. Chapter 1 states the claim with the negative space fixed in advance, the observations that would show where the older theories of delegation fail, and Chapter 69 reports what the evidence did to it.

Who this book is for #

The book is written for a strong software engineer who is new to machine learning. Part II builds the tools the rest uses, from probability and reinforcement learning to the economics of delegation that the argument rests on (Chapter 14). Researchers in machine learning can start at the main lines of research (Chapter 3). Readers from economics, law, policy, and the sciences can start at the field guide (Appendix D), which gives each discipline a path through the book.

How to read it #

The landing page offers three ways in: from the beginning, through the technical frontier, and from another field. The book, each part, and each chapter open with a brief: the ideas, the evidence with its status, and what is open, each bullet pointing to the section that develops it, so a reader can see the argument before reading it or find the section to read. Each chapter states its question in its lead and reports the evidence section by section; each research chapter then says what the evidence means for delegation and ends with what is settled, what is contested, and what is missing. Every empirical claim is cited in the sentence that makes it, and a badge marks works that are preprints or other publications that were not peer reviewed. A sentence that states the book's own inference rather than a source's ends with (inference). A statement that no study was found means none was found in this search, as of October 2026.

The bibliography distinguishes the spine, the works a reader needs to follow the argument, from side reading. The reading map (Appendix F) lists every work by field, spine first. Particulars that date quickly, such as the provisions of statutes, the versions of developers' frameworks, the incident record, and scores of named models, are collected in a dated appendix (Appendix I); a chapter that needs one states it as of a date and points there, so that the chapters stay right when those particulars change.

Sources and conventions #

Much of the research this book reports is recent, published as preprints or posts, and published by the laboratories whose systems it studies. The book marks this in the prose and in the badges, and it applies one standard of evidence throughout: a result counts as settled only when a group without a stake in it has reproduced it (Section 5.7).

The laboratory layer of the bibliography comes from five research indexes as they stood on 24 October 2026: four of Anthropic's (its alignment and frontier red team research pages, its Alignment Science blog, and the Transformer Circuits thread) and OpenAI's alignment research blog. The collection took every work those indexes list and every source those works cite, then added the works from other fields that the chapters needed (Appendix E). Every work was matched to its record on a primary page before it could be cited, and the works that could not be matched are listed with the reason (Appendix G). The sources were read between 5 and 9 October 2026, with later rounds of additions and corrections on 13 and 14 October and on 23 and 24 October, when the indexes were collected again; nothing published after 24 October 2026 is reflected.

Position #

A book that judges the laboratories' evidence should say where its author stands. I hold a Ph.D. in computer science from LMU Munich (2023), with research on human-in-the-loop systems, human-AI interaction, and the calibration of trust in agentic AI. I was a staff engineer at Sixt SE, leading enterprise AI work, and I founded Latere AI, which builds AI infrastructure with human intelligence in the loop.

I have no employment, funding, or advisory role with Anthropic, OpenAI, Google DeepMind, Meta, or Microsoft. Latere AI uses these laboratories' models in its commercial products and is not tied to any one of them.

The first pass of this book, the literature collection, the outlines, and the drafts, was automated with Claude models, made by Anthropic, whose research indexes supply four of the five sources of the laboratory layer. That is a reason for suspicion, and the book's process is built around it. Every chapter is reviewed by me before it leaves draft status, and the drafts are reviewed by models from developers other than Anthropic, so that the developer whose research the book examines does not also supply the last check on its text. As of October 2026 every chapter is a draft: complete and checked against its sources, and not yet through my review. The outside review, by OpenAI's GPT-6.1 Sol, read every chapter, part page, and written appendix except the chapter on cruxes, whose review that model's own safety filter stopped; the three appendices generated from the bibliography were not part of it. Each of its findings was checked against the sources before the text changed.

What this could bias: which works form the spine; how generously the book reads laboratories' reports about their own systems; its vocabulary, much of which the laboratories coined and which the glossary credits to its coiners; and the view of the laboratories held by an author whose company builds on their models. The reader is entitled to discount accordingly, and the citations are there so that the discount can be checked.

Use of language models #

This book was written by large language models, prompted and steered by the author. That includes the text, the interactive figures and the code behind them, the bibliography, and much of the checking; no sentence was written or edited by hand. The book's four problems apply to its own production, and it is fair to say how each was handled.

Specification: the author set the book's scope, structure, and argument in a written plan and in conventions for drafting, and decided every question of structure and every claim that depended on judgment. Instillation: each chapter was drafted by a separate model instance from the plan, the conventions, an exemplar chapter, and the sources, and each decision a drafter made that touched another chapter was recorded and checked. Verification: every cited work was matched to its record on a primary page; every claim about a work was checked against the work's text, at the access the book actually had, and logged with the passage that supports it; claims about works the book could read only as catalog records were removed; and the plan, the structure, and the exemplar chapters were reviewed by separate model instances, each assigned a lens such as accuracy, the reader, or the argument. Maintenance: errors are corrected as they are found, and the dated appendix states when its particulars were collected.

The author takes full responsibility for the accuracy, integrity, and conclusions of the book.

How this book could be wrong #

The argument could fail in at least three ways. The delegation thesis may fit a deployed model poorly: a model is not a person with interests, and every analogy the book draws from human delegation stops somewhere, which the chapters try to mark. The laboratory layer may tilt the spine toward what two laboratories chose to publish, and toward the problems they find tractable. And much of the recent evidence has not been reproduced by anyone without a stake in it, so some of what the book reports as findings will not survive.

The readers most likely to find these errors are researchers outside the large laboratories, critics of the safety framing, and specialists in the fields the book borrows from. Their corrections are the ones this book most needs.