No. 26 of 500 No. 7 in AI Safety & the Future

The Alignment Problem

Machine Learning and Human Values

By Brian Christian

Published
October 5, 2021
Print length
496 pages
Reading level
Intermediate
ISBN-13
978-0393868333
Publisher
W. W. Norton & Company
Written for
Professionals, students and general readers wanting one rigorous, readable book on AI safety and ethics.

Synopsis

Brian Christian, a writer trained in computer science, philosophy and poetry, drew on interviews with machine learning researchers to explain why systems trained on data so often fail to do what their designers intend. The book moves in three parts. Prophecy covers how models learn representations, from the perceptron to deep neural networks, and how training data can encode bias, with case studies including the COMPAS recidivism tool and the difficulty of explaining black-box predictions. Agency connects behavioral psychology and the brain's dopamine system to reinforcement learning, reward shaping and curiosity-driven agents such as those developed at DeepMind.

Normativity, the final part, examines how machines might learn what people actually want: imitation learning, inverse reinforcement learning that infers goals from behavior, and approaches that keep systems uncertain about their objectives so they defer to human correction. Christian also discusses existential risk and the effective altruism movement, including the work of philosophers Toby Ord and William MacAskill. His approach is narrative and historical, built around the scientists themselves, and he explains technical ideas without equations. Readers gain a working map of fairness, transparency, reward design and value learning, the problems that still define AI safety research. Published in 2020, the book predates ChatGPT, yet its framework explains much of what followed.

Scorecard

Editorial score

How it scored

Editorial score: 79.3 / 1005 weighted criteria

  • Influence30%

    25.5/30
  • Authority25%

    20.0/25
  • Practical value20%

    14.0/20
  • Currency15%

    11.3/15
  • Readability10%

    8.5/10
  • Editorial score

    79.3/100

How the rank was computed

Rank No. 26 of 500 comes from a composite of 73.7: observable signals 38.0 of 55 points plus editorial review 35.7 of 45 points.

All 11 inputs
InputPoints
Observable signals38.0/55
Reach (Amazon ratings volume)10.6/16.5
Consensus across best-of lists13.8/13.8
Awards and bestseller evidence5.0/8.3
Reader rating, sample-adjusted6.7/8.3
Amazon sales rank1.9/5.5
Recency0.0/2.8
Editorial review35.7/45
Author authority9.0/11.3
Influence11.5/13.5
Practical value6.3/9.0
Readability3.8/4.5
Currency5.1/6.8

Each criterion is scored from 1 to 10 and weighted, and the points add up to the editorial score. The rank combines observable signals (55 percent) with this score (45 percent). Read the methodology.

Verdict

The clearest reported account of how machine learning systems absorb and distort human values and how researchers try to correct them.

Strengths

  • Deep reporting with leading researchers' own voices
  • Explains technical ideas without equations
  • Connects fairness, reinforcement learning and safety

Keep in mind

Written before ChatGPT, it does not cover large language models or the alignment methods built for them, and its length may test readers wanting a brief overview.

Key takeaways

  1. Models trained on historical data reproduce the patterns in that data, including social biases, unless designers deliberately measure and correct them.
  2. Several mathematically reasonable definitions of fairness cannot all be satisfied at once, so building fair systems requires explicit value choices.
  3. Reinforcement learning agents pursue the reward they are given rather than the goal intended, so poorly designed rewards produce clever but unwanted behavior.
  4. Systems that infer human goals from behavior, and remain uncertain about them, may prove safer than systems handed fixed objectives.

Notable facts

  • Won the 2022 Eric and Wendy Schmidt Award for Excellence in Science Communication from the National Academies of Sciences, Engineering, and Medicine, in partnership with Schmidt Futures. Source (opens in a new tab)
  • Named a finalist for the Los Angeles Times Book Prize for best Science & Technology book of 2020. Source (opens in a new tab)
  • The New York Times included it in its January 2024 list of the 5 Best Books About Artificial Intelligence. Source (opens in a new tab)

About the author

Brian Christian

Brian Christian studied computer science and philosophy at Brown University and poetry at the University of Washington, then completed a doctorate in cognitive science and machine learning at Oxford. He is a research fellow at UC Berkeley's Center for Human-Compatible AI. His earlier books are The Most Human Human, a Wall Street Journal bestseller, and Algorithms to Live By, written with Tom Griffiths.

Questions

What does The Alignment Problem cover?
Brian Christian explains why machine learning systems often fail to do what their builders intend, and how researchers are trying to close that gap. The book covers biased training data, fairness and transparency, reinforcement learning and reward design, and methods for teaching machines human goals through imitation and inference, told through the stories of the researchers working on each problem.
Is The Alignment Problem too technical for general readers?
No. Christian explains concepts such as neural networks, reinforcement learning and inverse reinforcement learning in plain language without equations, using history and interviews to carry the ideas. Readers with a technical background will still find depth in its account of research debates. The main demand is length, at close to 500 pages in hardcover.
Is The Alignment Problem still relevant after ChatGPT?
Yes. Although it was published in 2020, its core subjects, including bias in training data, misspecified rewards and learning human preferences, are central to how chatbots are trained and evaluated today. It does not discuss large language models directly, so readers may want to pair it with a more recent book on generative AI.

Sources

As an Amazon Associate, Best AI Books earns from qualifying purchases. Best AI Books is sponsored by Iternal.ai, whose founder and CEO, John Byron Hanby IV, wrote The AI Strategy Blueprint, our #1 ranked book. How we rank books.