Artificial Intelligence
What AI can and cannot do
Description
In 2016, a program called AlphaGo beat Lee Sedol, one of the strongest Go players alive, four games to one. Go had long been the game AI researchers pointed to as out of reach — too vast, too intuitive, too human. When the machine won, the headlines did what headlines do: the singularity was near, human-level intelligence had arrived, the last citadel had fallen. Melanie Mitchell, a computer scientist who studied under Douglas Hofstadter, watched the reaction with a mix of admiration and unease. The system was a genuine achievement. It was also, she knew, spectacularly narrow — a program that could not tell you what a game of Go was, or why anyone would play one.
That gap — between what these systems do and what we imagine they understand — is the thread Mitchell pulls through her book. She has spent a career inside the field, close enough to know its real capabilities and skeptical enough to distrust its promises. The same decade that produced AlphaGo also produced self-driving cars that stalled on unexpected snow, image classifiers fooled by a few altered pixels, and chatbots that assembled fluent sentences with no idea what they meant. The successes were real. So were the failures. The trouble was telling which was which from the outside.
Mitchell's aim is neither to hype nor to dismiss. It is to look clearly at a technology that has become an economic force and a cultural obsession, and to ask a plainer question than the headlines allow: what is actually happening inside these machines, and how far does it really go?
The question we’re asking : When a machine beats us at a task, what has it actually understood — and how do we tell competence from comprehension?What we’ll see : How these systems learn, where they break, and why the harder problem may be understanding ourselves.
Table of contents
01Chapter 1 — The word that meant everything and nothing
The term "artificial intelligence" was coined in 1956, at a summer workshop in Dartmouth organized by John McCarthy and a handful of others who believed a machine could be made to simulate any feature of human thought. Mitchell notes the confidence of that founding moment — the participants expected serious progress within a generation. What followed instead was a cycle the field would repeat for decades: a burst of optimism, a wave of funding, a collision with problems harder than anyone guessed, then a collapse of interest that researchers came to call an AI winter.
Part of the trouble was that nobody agreed on what intelligence was. Mitchell points out that the field never settled its own central term. Early researchers assumed the hard part of thinking was formal reasoning — chess, logic, theorem-proving — because that is what educated humans found difficult. The easy part, they figured, was perception and common sense, the things a child does effortlessly. They had it backwards. Machines mastered chess decades before they could reliably tell a photo of a dog from a photo of a cat, and even now they stumble over the ordinary knowledge a toddler takes for granted.
02Chapter 2 — How the machine learned to guess
The engine behind the recent boom is the neural network, an idea loosely inspired by the brain and dating back, in primitive form, to the 1950s. A network is built from layers of simple units, each passing numbers to the next, with adjustable weights on every connection. Mitchell is careful about the "brain-inspired" framing: the resemblance to actual neurons is thin, and treating it as more is one of the field's oldest sources of confusion. What the network really does is tune millions of numbers until its outputs match the examples it was trained on.
The training itself is a form of guided guessing. You show the system a labeled image — this is a cat — it makes a prediction, you measure how wrong it was, and an algorithm nudges every weight a little in the direction that reduces the error. Repeat this across millions of examples and the network gradually gets good at the task. Mitchell stresses what this process is not: there is no understanding of cats, no concept of an animal, only a statistical mapping from pixels to labels, refined by relentless correction.
03Chapter 3 — The barrier of meaning
The failures are where Mitchell's argument sharpens. A network trained to recognize objects can be fooled by changing a handful of pixels invisible to a human eye — a school bus confidently reclassified as an ostrich. These "adversarial examples" are not rare glitches; they reveal that the system never learned what a bus is. It learned surface statistics that happen to correlate with the label, and those correlations can be gamed. The machine sees textures and patterns where we see things.
The same limit surfaces in language. Systems that translate and generate text with startling fluency have, Mitchell argues, no grip on meaning. They predict which words tend to follow which, having absorbed vast amounts of human writing, but they do not know what the words refer to. She uses the case of Winograd schemas — sentences where a pronoun's reference depends on knowing how the world works. "The trophy didn't fit in the suitcase because it was too big." What was too big? Any person knows instantly; the machine has no body of common sense to consult.
04Chapter 4 — A mirror we keep mistaking for a mind
Step back and the recurring error is not in the machines but in us. Mitchell keeps returning to how readily people credit these systems with understanding they do not have. When AlphaGo won, we read intention into it. When a chatbot answers smoothly, we assume someone is home. This is the same instinct that lets us see faces in clouds, projected onto silicon — and it matters, because it shapes how we deploy technology we have misjudged.
The deeper problem, she suggests, is that we do not understand our own intelligence well enough to build it or even to measure it. Every benchmark AI conquers turns out to test something narrower than we meant. We thought Go demanded intuition; it demanded search and pattern-matching at scale. We thought language demanded understanding; a great deal of it can be faked by prediction. Each victory redraws the line and shows us how much of what we call thinking we had never actually examined. AI keeps functioning as an accidental probe into human cognition.
05Conclusion
The Go match that opened the story remains a genuine landmark, and nothing Mitchell writes takes that away. AlphaGo did something extraordinary. It also could not explain, want, or understand a single move it made — and that combination, not the victory alone, is the real state of artificial intelligence. The field has produced tools of remarkable narrow power while remaining, on the questions that first animated it, almost as far from human-level intelligence as it was in that Dartmouth summer of 1956.