Artificial Intelligence: A Guide for Thinking Humans Summary and Analysis

Artificial Intelligence: A Guide for Thinking Humans by Melanie Mitchell is a clear, thoughtful guide to what AI can and cannot do. Rather than treating artificial intelligence as magic or doom, Mitchell explains its history, methods, achievements, limits, and ethical risks with calm judgment.

She moves from early symbolic AI to neural networks, computer vision, reinforcement learning, language systems, and debates about artificial general intelligence. The book’s central concern is meaning: machines can classify images, play games, translate text, and answer questions, but they still struggle with common sense, abstraction, real-world context, and true understanding.

Summary

Artificial Intelligence begins by setting out the strange tension that surrounds the field. Artificial intelligence has existed for decades, yet public expectations keep rising and falling.

Researchers, entrepreneurs, and commentators have repeatedly predicted that human-level machine intelligence is close, but each wave of progress has also exposed how difficult ordinary human abilities are to reproduce. Melanie Mitchell frames the book around this gap between impressive performance and genuine understanding.

She does not deny the power of modern AI; instead, she asks what kind of intelligence these systems really have, and what they still lack.

The book opens with the unease caused by rapid progress. Mitchell recalls a visit to Google, where Douglas Hofstadter, her former mentor and a major influence on her thinking, told a room of AI researchers that he was terrified.

His fear was not simply that machines might become hostile. He was unsettled by the possibility that qualities people associate with human depth, such as creativity, musical feeling, insight, and thought, might be reduced to mechanical procedures.

This moment introduces one of the book’s main questions: if a machine can perform a task that once seemed deeply human, does that mean it understands the task, or has it only found a powerful shortcut?

Mitchell then traces the roots of AI to the mid-twentieth century, when computers inspired scientists to imagine thought as symbol manipulation. Alan Turing, John von Neumann, Marvin Minsky, John McCarthy, Claude Shannon, and others believed that intelligence might be described precisely enough to be simulated on a machine.

Early AI was driven by bold optimism. Researchers created programs that solved puzzles, played games, and manipulated symbols according to rules.

These systems suggested that reasoning could be formalized, but they also revealed major limitations. The things that seemed intellectually advanced, such as chess or logical puzzles, were often easier for machines than things children do naturally, such as recognizing objects, moving through the world, or understanding a simple story.

The book explains how symbolic AI gave way, in many areas, to machine learning. Instead of programming every rule by hand, researchers began designing systems that could learn patterns from examples.

Neural networks became central to this shift. Mitchell describes them as systems made of simple connected units, loosely inspired by the brain.

A network receives input, processes it through layers, and produces an output, such as a prediction about whether an image contains a handwritten digit. Through training, the system adjusts the strength of its internal connections.

Backpropagation, a method for using errors to improve the network, made multilayer networks far more practical. Over time, increasing computing power, larger datasets, and better methods turned neural networks from a disputed idea into the engine of modern AI.

This shift produced what Mitchell calls a new AI spring. Systems began succeeding at tasks that had long resisted automation.

Google’s neural networks learned to recognize cats in videos. Image classifiers improved rapidly.

Speech recognition, translation, facial recognition, search, virtual assistants, and recommendation systems entered everyday life. IBM’s Watson defeated human champions on Jeopardy!, and DeepMind’s AlphaGo beat one of the world’s strongest Go players.

These events created excitement and investment, but Mitchell repeatedly distinguishes narrow success from general intelligence. A system may outperform people in one domain while remaining helpless outside it.

A large part of Artificial Intelligence focuses on computer vision. Humans do not merely see objects; they interpret scenes.

When people look at a photograph, they infer who is present, what is happening, what probably happened before, what might happen next, and which details matter. Machines, by contrast, must turn raw pixels into categories.

Convolutional neural networks brought major improvements by learning visual features through many layers. Early layers may detect edges or simple shapes, while later layers respond to more complex patterns.

ImageNet, a huge labeled image dataset, played a decisive role in this progress. When AlexNet performed far better than competing systems in the ImageNet challenge, deep learning became the dominant force in computer vision.

Yet Mitchell emphasizes that strong classification is not the same as human sight. Vision systems require enormous labeled datasets, careful architecture choices, and expert tuning.

They can make errors that no person would make, and they may depend on superficial cues rather than meaningful concepts. A system might associate a blurry background with an animal, or misclassify an image after tiny changes invisible to human observers.

These adversarial examples show that neural networks often learn statistical patterns rather than robust understanding. They can also absorb social biases from their data, causing unfair or dangerous outcomes in systems such as facial recognition.

Mitchell then turns to trust and ethics. AI already influences medicine, navigation, policing, hiring, loans, sentencing, news feeds, and many other areas.

The question is not only whether AI works, but when people should rely on it. She describes benefits such as translation, accessibility tools, fraud detection, and scientific analysis, while also stressing privacy risks, bias, opacity, and overconfidence.

Unlike electricity, AI systems do not behave predictably in all settings. Their failures can be hard to explain because deep networks cannot easily show their reasoning.

Mitchell argues that trustworthy AI requires oversight from governments, companies, universities, and public-interest groups. The challenge is not only technical; it is moral and social.

The book also explains reinforcement learning, in which an agent learns by acting in an environment and receiving rewards. Mitchell uses simple examples to show how an agent tries actions, receives feedback, and gradually learns which choices lead to better outcomes.

This approach has produced striking results in games. DeepMind combined reinforcement learning with neural networks to create systems that learned Atari games from pixels and rewards.

AlphaGo and AlphaZero showed even greater success in Go, chess, and other games. These achievements are important because games offer clear rules, measurable rewards, and repeatable environments.

But Mitchell again cautions against overreading the results. Game-playing systems are often highly specialized.

They usually cannot transfer what they learn in one setting to another. AlphaGo can master Go, but it does not thereby understand strategy in a broad human sense, nor can it use that knowledge to load a dishwasher, navigate a social situation, or interpret a story.

Real environments are messy, partly observable, physically complex, and filled with rare cases. Reinforcement learning has shown great promise in controlled domains, but it remains far from flexible, everyday intelligence.

Language is another major test. Mitchell shows that understanding language requires much more than matching words.

A simple story about a man receiving a burned hamburger and leaving without paying requires readers to infer that he probably did not eat it. That inference depends on common sense about restaurants, anger, food, payment, and human behavior.

Natural language processing has improved greatly through statistical methods, word embeddings, recurrent networks, encoder-decoder systems, attention, and large-scale training. These methods allow systems to transcribe speech, translate text, answer some questions, and generate fluent sentences.

Still, Mitchell argues that fluency can be misleading. Translation systems may perform well on standard metrics while missing context, idiom, tone, or implied meaning.

Image-captioning systems may produce plausible descriptions while misidentifying objects or actions. Watson’s Jeopardy!

victory looked like a triumph of machine intelligence, but it depended on searching huge stores of text, ranking possible answers, and exploiting the structure of the game. When moved into medicine and other real-world areas, the branding around Watson often exceeded its practical capabilities.

Reading-comprehension benchmarks can also be narrow: a system may extract an answer from a passage without truly understanding the passage.

The central idea of Artificial Intelligence is what Mitchell calls the barrier of meaning. Humans understand the world through intuitive physics, intuitive biology, intuitive psychology, embodiment, memory, analogy, abstraction, and lived experience.

People form concepts from few examples, use metaphors grounded in bodily life, imagine alternatives, infer other minds, and transfer knowledge across domains. Current AI systems, despite their power, do not possess this kind of broad, grounded common sense.

Mitchell reviews attempts to give machines knowledge and abstraction, from symbolic commonsense projects such as Cyc to modern systems trained through data, simulation, or visual puzzles. Each approach has made partial progress but has not solved the deeper problem.

Human knowledge is not merely a list of facts. Much of it is implicit, contextual, physical, social, and analogical.

Machines can store information and detect correlations, but they struggle to form flexible concepts and apply them in new situations.

The book closes by considering the future. Mitchell discusses self-driving cars, employment, creativity, and artificial general intelligence.

She treats confident timelines with skepticism, noting that predictions of near human-level AI have often failed. She is less worried about an immediate superintelligent takeover than about brittle systems used too widely, manipulated too easily, or trusted too much.

AI may transform society, but its core scientific problems remain unresolved. Artificial Intelligence ultimately presents the field as both powerful and incomplete: a technology capable of remarkable feats, yet still missing the common sense and understanding at the heart of human intelligence.

Key Figures

Melanie Mitchell

Melanie Mitchell is the central guiding figure of Artificial Intelligence, acting as the thinker whose questions, judgments, and intellectual history shape the book. She presents herself as both an insider and a skeptic: someone deeply familiar with AI’s technical progress, yet unwilling to accept exaggerated claims about machine intelligence.

Her voice is measured, curious, and corrective. She does not dismiss AI achievements, but she repeatedly asks what those achievements actually prove.

Mitchell’s role is especially important because she stands between public excitement and technical reality. She helps the reader see why systems that classify images, translate sentences, or win games can still lack common sense, abstraction, and real understanding.

Her intellectual honesty gives the book its balance. She admires the ingenuity behind AI while remaining alert to its brittleness, ethical risks, and conceptual limits.

Through her, the book becomes less a celebration or warning and more a careful investigation into what intelligence means.

Douglas Hofstadter

Douglas Hofstadter appears as one of the book’s most influential intellectual presences. He is not merely a background figure; he represents a deep, humanistic vision of intelligence.

His fear at Google gives the book one of its most memorable emotional moments. Hofstadter is disturbed not only by the possibility that machines may become powerful, but by the idea that human creativity, musical feeling, analogy, and thought might be imitated so well that they appear less mysterious or less special.

His importance also comes from his influence on Mitchell herself. His work inspired her to pursue AI, and his emphasis on analogy, meaning, and consciousness echoes throughout the book.

Hofstadter stands for a view of intelligence that cannot be reduced to task performance. He reminds the reader that the deepest questions in AI are not just about whether machines can win games or generate language, but whether they can truly understand, create, and experience meaning in anything like the human way.

Alan Turing

Alan Turing is presented as one of the foundational figures in the history of AI. His importance lies in the way he helped make machine intelligence thinkable.

By proposing the imitation game, later known as the Turing Test, he shifted the question from whether machines possess an inner mind to whether their behavior could become indistinguishable from human behavior in conversation. In the book, Turing’s role is both historical and philosophical.

He belongs to the early generation that saw computation as a possible model for thought, but his ideas also remain active in modern debates about whether fluent performance should count as intelligence. Mitchell uses Turing’s legacy to show both the power and limitation of behavioral tests.

A machine may imitate human responses convincingly in narrow settings, yet still fail to possess common sense or genuine understanding. Turing therefore becomes a figure whose ideas opened the field while also leaving behind questions that remain unresolved.

John McCarthy

John McCarthy is important as one of the formal founders of artificial intelligence. His role in organizing the Dartmouth workshop marks a turning point, because that event gave the field its name and its early ambition.

McCarthy represents the confidence of early AI: the belief that intelligence could be described with enough precision for machines to reproduce it. In the book, he also embodies the later humility that followed AI’s early optimism.

His admission that AI turned out to be harder than expected captures the central historical rhythm Mitchell describes: excitement, bold prediction, disappointment, renewed progress, and renewed caution. McCarthy’s significance comes from this dual position.

He helped launch the dream of building intelligent machines, but his later recognition of the field’s difficulty shows how much remained unknown. He is a figure of ambition tempered by experience.

Marvin Minsky

Marvin Minsky appears as another defining figure of early AI. He represents the symbolic tradition, in which intelligence was often understood as the manipulation of explicit symbols and rules.

Minsky’s early confidence reflected the field’s belief that human reasoning could be formalized and reproduced through computation. Yet Mitchell also uses his observations to highlight one of AI’s central paradoxes: tasks that seem difficult to adults, such as formal reasoning or chess, can be easier for machines than tasks that seem effortless to children.

His phrase that easy things are hard captures the book’s recurring lesson. Minsky’s role is also complex because he was skeptical of early neural networks, yet neural networks later became central to modern AI.

He therefore stands at an important crossroads in the field’s history, representing both the brilliance of early AI thinking and some of the assumptions that later approaches challenged.

Claude Shannon

Claude Shannon appears as a key early figure who helped connect information, computation, games, and intelligence. His work on information theory and his interest in chess-playing machines made him part of the intellectual environment from which AI emerged.

In the book, Shannon helps represent the mid-century belief that human thought might be studied mathematically and mechanically. His early predictions about intelligent machines also show how long the dream of human-level AI has been present.

Shannon’s importance is not only technical; he symbolizes the moment when computation began to seem like a general framework for understanding mind, communication, and problem-solving. His presence helps Mitchell show that AI did not arise from one invention or one laboratory, but from a convergence of mathematics, engineering, psychology, and philosophy.

John von Neumann

John von Neumann serves as one of the book’s major background figures in the formation of computational thinking. His importance lies in the analogy between computers and brains.

Like Turing and other early theorists, von Neumann helped create the intellectual foundation for treating thought as something that might be modeled by machines. In the book, he represents the scientific confidence of an era when digital computers were new and their implications seemed enormous.

His presence reminds readers that AI grew out of broader questions about whether mental processes could be understood as information processing. Although he is not developed as personally as Hofstadter or Mitchell, von Neumann’s role is essential because he belongs to the group of thinkers who made it plausible to imagine intelligence as a computational phenomenon.

Frank Rosenblatt

Frank Rosenblatt is central to the book’s account of neural networks. His perceptron represents one of the earliest attempts to build a machine that could learn from examples rather than follow only hand-coded rules.

In Mitchell’s treatment, Rosenblatt stands for a different vision of AI from the symbolic approach. Instead of treating intelligence as explicit logic, the perceptron suggested that learning could emerge from adjusted connections among simple units.

His work anticipated many later developments, even though early perceptrons had serious limitations and were criticized by figures such as Minsky and Seymour Papert. Rosenblatt’s importance lies in how his ideas returned with force decades later.

Modern deep learning is far more complex than the perceptron, but it still carries forward the basic hope that machines can learn patterns from data through networks of weighted connections.

Seymour Papert

Seymour Papert appears mainly through his connection to early critiques of neural networks. Alongside Marvin Minsky, he helped shape skepticism about what perceptrons could achieve.

In the book, Papert’s role is important because he shows how intellectual judgments can influence the direction of a field. The criticism of perceptrons contributed to a period in which neural-network research lost prestige and attention.

Yet Papert was not simply a negative figure; his presence helps illustrate how AI develops through debate, correction, and changing evidence. His role reminds readers that methods rise and fall not only because of technical success, but because of the assumptions researchers hold about what intelligence is and how machines might reproduce it.

Geoffrey Hinton

Geoffrey Hinton is one of the major figures behind the rise of modern deep learning. In the book, he represents persistence during a period when neural networks were often treated as unfashionable or limited.

Hinton and other connectionists believed that intelligence could emerge from distributed patterns across many simple units rather than from explicit rules. His influence becomes especially visible through the success of AlexNet, trained by Alex Krizhevsky under Hinton’s direction.

That breakthrough helped transform AI by showing that deep neural networks, given enough data and computing power, could outperform older methods in image recognition. Hinton therefore stands for the return of neural networks from the margins to the center of AI.

His role also shows how long-term conviction can reshape a field when technical conditions finally catch up to theoretical belief.

Yann LeCun

Yann LeCun is presented as one of the key developers of convolutional neural networks. His early work on LeNet and handwritten digit recognition showed that neural networks could perform practical visual tasks, even before the deep-learning boom.

LeCun’s importance comes from his commitment to ConvNets when they were not yet dominant. In the book, he represents the long development of an idea before its public breakthrough.

ConvNets did not suddenly appear in 2012; they were the result of years of research, refinement, and belief that better data and hardware would eventually make them powerful. LeCun’s role also helps Mitchell explain that modern AI progress depends not only on new algorithms, but on the alignment of older ideas with massive datasets, stronger computers, and institutional support.

Fei-Fei Li

Fei-Fei Li is one of the most important figures in the book’s account of computer vision because of her role in creating ImageNet. Her insight was that progress in object recognition required a much larger and more varied dataset than earlier competitions provided.

By organizing ImageNet around a vast hierarchy of labeled images, she helped create the conditions for deep learning’s breakthrough in vision. In Artificial Intelligence, Li represents the often overlooked importance of data.

AI progress is not only a matter of clever algorithms; it also depends on the labor, organization, and conceptual design behind training resources. Her work also raises questions about the human effort hidden inside machine learning, since large datasets depend on labeling, categorization, and social choices about what counts as an object, a category, or a correct answer.

Alex Krizhevsky

Alex Krizhevsky is significant because of AlexNet, the system that dramatically changed the ImageNet competition and accelerated the deep-learning revolution. His role in the book is tied to a specific turning point: the moment when a deep convolutional network outperformed competing image-recognition systems by a striking margin.

Krizhevsky represents the technical breakthrough that converted years of neural-network research into visible dominance. His work also demonstrates how progress in AI often depends on combining several elements at once: architecture, data, graphics-processing hardware, training methods, and expert judgment.

Through him, Mitchell shows that a single competition result can redirect an entire field, attract corporate attention, and reshape public expectations about machine intelligence.

Demis Hassabis

Demis Hassabis appears as a major figure in the book’s discussion of reinforcement learning and game-playing AI. As a cofounder of DeepMind, he represents the ambition to use games as controlled environments for developing more general forms of intelligence.

DeepMind’s work on Atari games, AlphaGo, and AlphaZero made Hassabis central to the modern AI story. In the book, however, his role is not treated as simple triumph.

Mitchell uses DeepMind’s achievements to ask whether success in games truly leads toward general intelligence. Hassabis therefore stands for both the extraordinary promise and the unresolved limitation of current AI.

His systems can master complex games, but they remain dependent on carefully defined environments, huge training processes, and specialized design.

Ray Kurzweil

Ray Kurzweil is the book’s major representative of Singularity thinking. His predictions about accelerating technological progress and the eventual arrival of superintelligent machines give shape to one of the most optimistic, and to Mitchell one of the most questionable, visions of AI’s future.

Kurzweil’s role is important because he articulates a belief that human-level or beyond-human intelligence is not only possible but approaching through exponential growth. Mitchell does not present him merely as a fantasist; she takes his influence seriously, especially because his ideas have shaped public and corporate imagination.

At the same time, his presence allows the book to examine the danger of projecting smooth future progress from uneven past achievements. Kurzweil represents the predictive confidence that Mitchell repeatedly tests against AI’s actual technical limits.

John Searle

John Searle appears as a key philosophical critic of strong AI. His distinction between simulating understanding and actually understanding it is central to the book’s larger argument.

Searle’s ideas help Mitchell frame the difference between outward performance and inner comprehension. A machine might manipulate symbols, generate correct answers, or appear conversational, but that does not automatically mean it understands meaning in the human sense.

His role is especially relevant in discussions of language, where fluent output can easily be mistaken for thought. Searle represents the philosophical pressure behind the book’s technical analysis.

He pushes readers to ask whether behavior alone is enough, or whether intelligence requires some deeper relation to meaning, intention, and experience.

David Cope

David Cope is important because of his Experiments in Musical Intelligence, a program that generated music in the style of composers such as Bach and Chopin. In the book, Cope’s work deeply unsettles Hofstadter because listeners could mistake machine-generated music for human composition.

Cope therefore represents one of the most provocative challenges to ideas about creativity. If a program can produce music that experts find convincing, what does that imply about artistic originality and emotional expression?

Mitchell uses this example not to settle the question, but to complicate it. Cope’s system shows that machines can reproduce stylistic patterns with impressive results, yet it also raises the issue of whether such production involves self-judgment, intention, or understanding.

His role makes creativity one of the book’s most philosophically charged subjects.

Andrew Ng

Andrew Ng appears as a figure associated with optimism about AI’s transformative potential. His comparison of AI to electricity captures a popular belief that AI will become a general-purpose technology reshaping nearly every industry.

In the book, Ng’s role is to express the excitement surrounding practical AI applications. Mitchell treats this view with respect but also with caution.

Electricity is predictable and governed by well-understood principles, while AI systems can behave unexpectedly, fail silently, or reproduce hidden biases. Ng therefore represents the productive but risky language of technological revolution.

His presence helps Mitchell examine how metaphors can inspire investment and innovation while also simplifying the real limits of current systems.

Gary Marcus

Gary Marcus serves as one of the book’s critical voices against exaggerated claims for deep learning. His importance lies in his insistence that neural networks often learn shallow patterns rather than robust concepts.

Mitchell uses his critiques to show why systems that perform well in training environments can fail when conditions change. Marcus represents a demand for stronger forms of generalization, reasoning, and built-in structure.

In the book, he is part of a broader skeptical tradition that does not deny AI progress but questions whether current methods are enough for humanlike intelligence. His role helps sharpen the distinction between statistical success and conceptual understanding.

Andrej Karpathy

Andrej Karpathy appears as a figure who helps illuminate the complexity of human understanding. His analysis of a humorous photograph shows how much background knowledge, physical intuition, social reasoning, and contextual inference humans bring to even a casual image.

In the book, Karpathy’s importance comes from his ability to make visible what is usually invisible: the enormous amount of understanding behind ordinary perception. His reflections support Mitchell’s growing view that embodiment and lived interaction with the world may be essential for intelligence.

Karpathy therefore serves as a bridge between technical AI practice and the book’s deeper concern with meaning.

Themes

The Gap Between Performance and Understanding

Mitchell repeatedly separates doing well on a task from understanding the task. This distinction shapes the whole argument of Artificial Intelligence.

A computer can win at chess, classify images, translate text, answer quiz-show clues, or generate fluent language without necessarily grasping the world as humans do. The book shows that AI achievements often look more human from the outside than they are on the inside.

Deep Blue’s chess victory was historically important, yet it relied on search, evaluation, and specialized hardware rather than human strategic awareness. Watson’s Jeopardy!

performance seemed conversational, but its strange mistakes revealed that it did not understand categories or context in a human way. Image classifiers can beat benchmarks while failing on altered images that people recognize instantly.

This theme matters because public discussion often treats output as proof of intelligence. Mitchell asks readers to be more careful.

Success in a narrow environment may reflect statistical pattern recognition, massive training, and clever engineering, not genuine comprehension. The book’s central warning is that mistaking performance for understanding can lead people to overtrust systems whose abilities are real but limited.

Common Sense as the Missing Core of AI

Common sense appears as one of the most stubborn barriers facing artificial intelligence. Humans use common sense constantly and usually without effort.

They know that dogs move on their own, glass can break, people act from beliefs and desires, burned food may make someone angry, and a driver may behave differently if a child runs into the road. AI systems, however, often lack this broad background knowledge.

They may process images or sentences successfully while missing the ordinary assumptions that make those images or sentences meaningful. Mitchell presents common sense not as a small add-on, but as a foundation for flexible intelligence.

It includes intuitive physics, intuitive psychology, social expectations, cause and effect, and the ability to imagine what might happen next. The difficulty is that common sense is hard to write down as rules and hard to learn from static datasets alone.

Much of it comes from embodied experience, interaction, perception, and analogy. This is why systems trained on huge amounts of data can still fail in surprising ways.

They have patterns, but not the lived model of the world that humans bring to ordinary situations.

The Power and Fragility of Machine Learning

Machine learning is shown as both the engine of modern AI and one of its major sources of weakness. Neural networks, supervised learning, reinforcement learning, ConvNets, and language models have enabled enormous progress because they allow systems to learn from examples rather than depend entirely on hand-coded rules.

This shift made possible advances in vision, translation, speech recognition, game playing, and many other fields. Yet Mitchell is careful to show that these systems are fragile.

They depend heavily on large datasets, human labeling, architecture choices, hyperparameter tuning, and training conditions. When the world differs from the data, performance can collapse.

A vision system may rely on background cues instead of object understanding. A game-playing system may master one environment but fail when small visual details change.

A translation system may handle common sentence patterns but lose meaning when context or idiom matters. This theme challenges the idea that more data and larger models automatically solve intelligence.

Scale matters, but it does not by itself guarantee abstraction, transfer, reliability, or meaning. Machine learning has made AI powerful, but it has also made its failures harder to predict and explain.

Ethics, Trust, and the Human Consequences of AI

The book treats AI as a social force, not just a technical field. Systems that classify, predict, recommend, and automate decisions already affect people’s lives in medicine, policing, hiring, finance, transportation, communication, and public information.

Mitchell’s concern is not only whether these systems are accurate, but whether they are trustworthy. A model can appear objective while carrying biases from its training data.

A face-recognition system can work better for some groups than others. A self-driving car can perform well in ordinary situations but fail in rare cases.

A language model can sound confident while producing false claims. These risks become more serious when people treat AI systems as neutral authorities.

Trust requires more than high benchmark scores; it requires transparency, accountability, testing, regulation, and an honest understanding of limits. Mitchell also shows that ethical AI cannot be solved by engineers alone.

Governments, companies, universities, researchers, and the public all have roles to play. The theme is ultimately about responsibility.

AI systems are made by people, trained on human data, deployed in human institutions, and judged by their effects on human lives.