Artificial Intelligence: A Guide for Thinking Humans

In the Prologue, Mitchell recounts a 2014 visit to Google’s headquarters, when she ironically got lost inside the Google Maps building while heading to an AI meeting. She sketches Google’s evolution from a search engine into an “applied AI” company, highlighting its AI-focused acquisitions, the creation of Google Brain, and how the company embraced Ray Kurzweil’s Singularity vision via DeepMind’s mission to use AI to “solve everything.”

Mitchell then notes how Douglas Hofstadter’s 1979 book Gödel, Escher, Bach inspired her interest in AI. She describes her obsession with the book, her pivot into computer science, and how she became one of Hofstadter’s graduate students and long-term collaborators. This backstory leads into the Google meeting, which was convened so that company researchers could hear Hofstadter’s thoughts on AI.

At the meeting, Hofstadter shocked the room by confessing, “I am terrified. Terrified” (6). He revisited his earlier belief that human-level AI was far off and explained how advances in computer chess, culminating in Deep Blue’s victory over Garry Kasparov, first shook his confidence. A more decisive blow came from David Cope’s EMI, a program that composed music in the style of Bach or Chopin. Hofstadter was disturbed when expert listeners mistook EMI’s composition for genuine Chopin, an outcome that undermined his sense of music as a direct channel to the human soul.

Hofstadter criticized Google’s drive toward AI and Kurzweil’s Singularity scenarios, fearing that humans could become “relics.” Noting that the Google engineers seemed puzzled by his terror, Mitchell clarifies that Hofstadter’s fear was not of hostile superintelligence but of the possibility that AI might reduce cherished human qualities to mere algorithmic tricks, thereby trivializing the human spirit.

Mitchell traces AI’s origins to mid-20th-century efforts to model human thought as “symbol manipulation” on digital computers. She explains how figures like Alan Turing and John von Neumann saw analogies between computers and brains and regarded human intelligence as something that, in principle, software could replicate. The formal birth of AI, she notes, was the 1956 Dartmouth workshop organized by John McCarthy, Marvin Minsky, Claude Shannon, and Nathaniel Rochester, which coined the term artificial intelligence and ambitiously proposed that it could describe every aspect of learning and intelligence well enough to simulate.

The chapter outlines the optimism of early AI pioneers, who predicted fully intelligent machines and universal machine competence within a few decades. Mitchell then raises the problem of definition: “Intelligence” is a “suitcase word” (19) that contains many overlapping meanings (binary, continuous, and multidimensional), yet AI largely pushes past these nuances to focus on two aims: scientifically modeling natural intelligence and practically building systems that perform as well as or better than humans in performing specific tasks.

Mitchell describes the resulting “anarchy of methods” within AI: symbolic approaches, which are based on logical rules and human-readable symbols; probabilistic and inductive approaches, which extract patterns from data; and biologically inspired “subsymbolic” methods. As an example of symbolic AI, she describes the General Problem Solver and its representation of the “Missionaries and Cannibals” puzzle (22) via explicit symbolic states and operators. She then contrasts this approach with Frank Rosenblatt’s perceptron, a simple neural model that learned to classify handwritten digits via supervised learning and the perceptron-learning algorithm.

In closing the chapter, Mitchell recounts cycles of AI “spring” and “winter,” noting that grand promises have often outpaced results. She cites John McCarthy’s admission that “‘AI was harder than we thought’” (33) and Marvin Minsky’s paradoxical observation that “‘[e]asy things are hard’” (33), highlighting how childlike capabilities remain stubbornly difficult to replicate in machines.

Mitchell notes that multilayer neural networks (which some researchers once dismissed as a dead end) are at the core of modern AI. She introduces neural networks as systems built from many simple units, loosely inspired by neurons and arranged in layers. To make this concrete, she gives the example of handwritten digit recognition: The network takes in raw pixel values from an image, processes them through one or more hidden layers, and produces output indicating how likely the image is to be each possible digit. Each unit in the network performs a small calculation, combining its inputs into a number between 0 and 1, and each output represents a confidence score.

While the earliest layers respond to simple features like light and dark regions, deeper layers can learn increasingly abstract patterns. However, deciding how many layers a network should have, or how large they should be, often involves experimentation rather than theory. She then introduces “backpropagation,” the learning algorithm that made training deep networks practical by allowing the use of output errors to adjust connections throughout the system. Although early critics such as Marvin Minsky and Seymour Papert cast doubt on neural networks, backpropagation enables multilayer systems to outperform simpler models in tasks like image recognition (though it makes their internal reasoning harder to explain).

The chapter then places these developments within the broader movement known as “connectionism,” which gained momentum in the 1980s through the work of David Rumelhart and James McClelland. Connectionists argue that intelligence emerges from patterns of interaction among many simple units rather than from explicit, human-written rules. Mitchell contrasts this approach with symbolic expert systems, which rely on carefully programmed rules and tend to break down outside narrow situations. In contrast, neural networks learn directly from data and adapt more flexibly, but they do so in ways that are difficult to interpret or control.

Mitchell widens the lens to machine learning as a whole. Machine learning is a branch of AI in which systems improve through experience. Like AI more broadly, machine learning has periods of enthusiasm and disappointment. In earlier decades, the lack of data and computing power often limited the results it could produce. As digital data and computing power exploded and hardware became faster and cheaper, these constraints loosened, setting the stage for the large-scale learning systems that define today’s AI boom.

Mitchell recalls how, in 2012, Google researchers trained a massive neural network on millions of YouTube video frames, and the network developed a unit that reliably responded to images of cats—without ever being told what a cat was. She presents this moment as emblematic of a new wave of deep-learning successes that earned AI mainstream attention, contrasting with AI trials in earlier decades, when everyday AI encounters were more often frustrating or gimmicky (e.g., automated phone menus, Furbies, or Microsoft’s Clippy). In addition, she revisits the shock of IBM’s Deep Blue defeating chess champion Garry Kasparov in 1997, noting that public unease faded once chess came to be seen as a narrow problem that brute-force computation could solve.

By the mid-2000s, AI applications rapidly began to appear in daily life: Google Translate, self-driving car prototypes, virtual assistants, automatic captioning, facial recognition, and image labeling. High-profile systems such as IBM’s Watson and DeepMind’s AlphaGo further reinforced the sense that machines could achieve humanlike capabilities in complex domains. Mitchell observes that these advances fueled a booming AI industry through heavy corporate investment and a surge of start-ups. As a result, many commentators once again predicted the imminent arrival of “general AI” (or AGI), even though the systems driving this optimism remained forms of “narrow” or “weak” AI, each designed for a specific task.

Mitchell then surveys competing views about where this momentum leads. Optimists such as DeepMind’s cofounder Shane Legg and Facebook’s Mark Zuckerberg suggest that human-level AI may arrive within decades, while skeptics like Steven Pinker argue that similar predictions have repeatedly failed in the past. To frame these disagreements, Mitchell revisits long-standing philosophical debates about machine intelligence, including Alan Turing’s imitation game, John Searle’s distinction between “strong” and “weak” AI, and enduring questions about whether successful simulation should count as genuine thinking.

Additionally, the chapter examines modern versions of the Turing Test and their shortcomings. For example, in the 2014 Eugene Goostman episode, a chatbot posing as a non-native English-speaking teen briefly convinced some judges that it was human—an outcome that many critics viewed as exploiting loopholes rather than demonstrating intelligence. Mitchell then turns to Ray Kurzweil’s Singularity predictions, which extrapolate exponential technological growth into a future of rapidly accelerating intelligence. Presenting both enthusiastic endorsements and pointed skepticism, she closes the chapter by asking whether the current AI spring represents a true turning point or simply another peak before an AI winter.