Artificial Intelligence: A Guide for Thinking Humans

Citation

Mitchell, M. (2020). Artificial Intelligence: A Guide for Thinking Humans. Pelican (Penguin Books). First published 2019 by Farrar, Straus and Giroux.

Chapter Summary

A book for general readers by a computer scientist (then at Portland State University, now at the Santa Fe Institute) whose research concerns abstraction, analogy-making and reasoning. After a prologue it has sixteen chapters in five parts: Background (Chapters 1–3), Looking and Seeing (4–7), Learning to Play (8–10), Artificial Intelligence Meets Natural Language (11–13) and The Barrier of Meaning (14–16). They follow the main areas of recent AI: history, vision, game playing, natural language and, finally, understanding. The publisher describes it as weaving together advances in AI with cognitive science and philosophy, and reviewers note that each area is first explained mechanically and then examined critically. The book was written in 2019, before today's large language model chatbots.

Prologue: Terrified

Mitchell recounts a 2014 meeting at Google that she attended with her former doctoral adviser Douglas Hofstadter, author of Gödel, Escher, Bach, who said he was "terrified" that programs such as champion chess players and music generators might turn human beings into "relics". His reaction frames the book's question of how intelligent current AI really is.

Chapters 1–3: The Roots of Artificial Intelligence; Neural Networks and the Ascent of Machine Learning; AI Spring

The history of the field from the 1956 Dartmouth workshop, the split between symbolic AI (rules and expert systems) and subsymbolic approaches beginning with the perceptron, the boom-and-bust cycles of optimism and "AI winter", and the rise of machine learning and deep learning. Reviewers mention her account of the Turing test and of predictions of a coming "singularity".

Chapters 4–6: Who, What, When, Where, Why; ConvNets and ImageNet; A Closer Look at Machines That Learn

Computer vision: why understanding a photograph is hard, how multilayer networks trained by backpropagation and convolutional neural networks classify images, and the ImageNet competition. The part then looks critically at what such systems learn, with examples reviewers single out: a network that learned to detect blurry backgrounds rather than animals, the role of human-set hyperparameters, and adversarial images that make a school bus look like an ostrich to the network.

Chapter 7: On Trustworthy and Ethical AI

A short chapter, about thirteen pages according to one reviewer, on the trustworthiness and ethics of AI systems.

Chapters 8–10: Rewards for Robots; Game On; Beyond Games

Reinforcement learning, introduced through a robot dog learning to kick a ball, then its combination with deep networks in DeepMind's Atari players, the search methods behind Deep Blue and AlphaGo, and whether skills learned in games transfer to the real world.

Chapters 11–13: Words, and the Company They Keep; Translation as Encoding and Decoding; Ask Me Anything

Natural language processing: sentiment classification, recurrent neural networks, word vectors such as word2vec, encoder–decoder machine translation, question answering and reading comprehension, including Winograd schemas as a test that requires commonsense knowledge.

Chapters 14–16: On Understanding; Knowledge, Abstraction, and Analogy in Artificial Intelligence; Questions, Answers, and Speculations

The final part asks what understanding would require: intuitive physics and other core knowledge, abstraction and analogy (the subject of the author's own research), and why current systems lack them. The final chapter, as its title says, turns to questions and speculations about AI's future.

Key Concepts

Part I: Background

• Symbolic versus subsymbolic AI: programming explicit rules and reasoning, as against learning from data with networks of simple units.
• Perceptron: an early artificial neuron that sums weighted inputs and fires if a threshold is passed.
• AI spring and winter: the recurring cycles of optimism, funding and disappointment in the field.

Part II: Looking and Seeing

• Convolutional neural network: a network that builds up features such as edges and shapes to classify images.
• Training data and hyperparameters: the labelled examples and human-chosen settings on which deep learning depends.
• Adversarial examples: inputs altered imperceptibly so that a network misclassifies them.

Part III: Learning to Play

• Reinforcement learning: learning which actions earn reward through trial and error, balancing exploration and exploitation.
• Transfer: whether skill learned in one task carries over to another, which the game-playing systems largely lack.

Part IV: Natural Language

• Word vectors: numerical representations in which words used in similar contexts lie close together.
• Encoder–decoder translation: networks that encode a sentence in one language and decode it into another.
• Winograd schema: a sentence pair whose correct reading depends on commonsense knowledge.

Part V: The Barrier of Meaning

• Common sense and core knowledge: intuitive knowledge of physics and biology, such as how objects fall or which things move on their own, that people use without noticing.
• Abstraction and analogy: the ability to see the essence of a situation and map it to another, which the author treats as central to understanding.

Critical Analysis

Strengths

• Clear technical explanation for non-specialists. Kirkus Reviews calls it a nonmathematical yet somewhat technical and mostly lucid introduction, and Barbara Spindel in The Christian Science Monitor praises how clearly it conveys its themes; perceptrons, backpropagation, convolutional networks, reinforcement learning and word vectors are all explained without equations.
• Calibrated scepticism. Each part pairs successes with their limits: shortcut learning and adversarial examples in vision, lack of transfer in games, the need for real understanding in language. Peter Kassan's review in Skeptic calls her assessment measured and cautious.
• History as context. Jim Hahn in Library Journal singles out the historical grounding, from post-war expert systems and reasoning to the early dismissal and later validation of deep learning, which gives students perspective on current claims. The author is also candid about uncertainty: asked whether AI will cause mass unemployment, she answers that she does not know.
• A researcher's view of understanding. Part V draws on the author's own work on analogy and abstraction, connecting AI to cognitive science.

Limitations

• Dated in fast-moving areas. Published in 2019, the language chapters cover recurrent networks and encoder–decoder translation and do not cover the transformer-based large language models that now dominate the field. Readers should check later developments before relying on its examples or benchmark figures.
• Not a textbook. There are no exercises or mathematics, and algorithms are explained conceptually; Kirkus also found some passages too abstruse for general readers, and the Christian Science Monitor review notes technical language that can challenge laypeople.
• Thin on ethics and society. Kassan's review describes the chapter on trustworthy and ethical AI as short and inadequate to the subject.

Where It Fits in the Course

In the machine-learning and AI course this is the non-technical orientation. Russell and Norvig's Artificial Intelligence: A Modern Approach covers the same history, search, game playing, learning, language and vision as a full textbook with algorithms; James, Witten, Hastie, Tibshirani and Taylor's An Introduction to Statistical Learning teaches the statistical methods with Python labs; Murphy's Probabilistic Machine Learning and Szeliski's Computer Vision go deep into machine learning theory and vision. Mitchell is best read first, or alongside the first weeks of those books, to understand what the techniques are for and where their limits lie.

Its questions about understanding, generalisation and evaluation remain useful when reading technical chapters on deep learning: students can test her claims about shortcut learning, adversarial examples and transfer against what the textbooks report.

How to Use This Book

The book assumes no mathematics or programming and suits first-year students in any discipline, as well as computer science students who want perspective before technical courses.

• Core: the Prologue and Parts I and II, which establish the history and the critical view of deep learning used in the rest of the book.
• By interest: Part III for reinforcement learning and games, Part IV for language.
• Essential conclusion: Part V, which states the book's central argument about meaning and understanding.

Read it quickly as background, then return to specific chapters when the course reaches convolutional networks, reinforcement learning or language models. For each technique, note the strength and the weakness Mitchell identifies and look for how later research has addressed it; in Part IV in particular, compare her account of language systems with current large language models.