Preface

Most organizations now run AI systems that talk to customers, write code, draft documents or take actions on someone’s behalf. Each of those systems was tested by someone before it went live. Then it changed: the provider shipped a new version, an engineer edited the instructions, a cheaper copy replaced the original, a new tool was connected. This book is about what happens to the behavior you checked when the system around it changes, and about the other ways deployed AI systems end up doing things their owners did not intend.

It is written first for an executive or team lead who has to decide how the company keeps its AI systems behaving as intended, and who does not have a background in machine learning. Nothing here assumes one. Where a point needs technical depth, it sits in a short box marked for practitioners, so engineers get the settings and numbers they need without the main text depending on them. Students and researchers should find it useful as a map of how deployed systems fail and where the open questions are.

The book stays close to the public record. Its examples are real incidents, told from company postmortems, regulator and court records, research papers and news reporting, and every factual claim carries a numbered endnote with a link to its source.

How the book is organized

Chapter 1 tells four recent cases in which a change to a deployed system broke behavior that had been checked, and draws out what the public record supports. Part One explains, in plain terms, how AI models are made, how they are taught to behave, how they are tested and why tests mislead, and what changes when a model is given tools and allowed to act. Part Two is the core of the book: seven chapters on how deployed systems go wrong, from training that rewards the wrong thing to agents that act on text written by an attacker to robots in the physical world, each built around the cases behind it. Part Three turns to practice, with a routine for any change, checklists for six common kinds of deployment, and questions an executive can put to a team or a vendor. Part Four sets out what nobody knows yet. The reference section holds a record of the public incidents, a table of who is most exposed to each failure, a glossary, a map to standards and regulations, and suggestions for further reading.

The chapters can be read in order or by need. A reader responsible for one deployment can start with the opener to Part Two, which says where each kind of team should begin.

What the book does not do

It does not teach how to train models; Part One gives enough to follow the rest and points to the people who teach it properly. It makes no forecasts, rates no company’s overall safety, and offers no legal or compliance advice, though it shows where its advice meets existing standards. Deliberate misuse of AI by people who intend harm is a different problem from systems that misbehave, and the book treats it only briefly and labels it as such.

License and interests

Align Your AI is published by Norbury Labs and is free to read, download and share under the Creative Commons Attribution 4.0 International license (CC BY 4.0).¹ The book is free, and nothing in it is for sale.

Notes

1. Creative Commons. “Attribution 4.0 International (CC BY 4.0).” creativecommons.org, license deed. https://creativecommons.org/licenses/by/4.0/. The license under which the book may be shared and adapted with attribution. ↩︎