Landmark Works on AI

Most writing about AI is out of date within a year. These works are not. Each one started an argument that is still running, or gave the field a word it still uses. People read them for what they say, not for what they were. The oldest is a note from 1843, and the newest an encyclical from May 2026. The list is short on purpose.

Where a free copy exists, the link opens it. A book links to its publisher, and a paper behind a paywall links to the journal’s page. A library can get you either.

Where it began

• Ada Lovelace, Notes to Menabrea’s “Sketch of the Analytical Engine” (1843) (opens in a new tab). The first argument about whether a machine can think, made in a translator’s notes. Note G says the Engine “has no pretensions whatever to originate any thing”: it does what it is ordered to do. Turing answers her a century later, and the exchange is still going. Free at Project Gutenberg.
• Alan Turing, “Computing Machinery and Intelligence,” Mind (1950) (opens in a new tab). Turing replaces “Can machines think?” with a game: can a machine pass for a person in conversation? Then he answers nine objections, Lovelace’s among them. Read it for the objections. Most arguments made since are already in them. Free to read on the journal’s page.
• John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon, “A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence” (1955) (opens in a new tab). The proposal that named the field and asked for two months and ten men. Every feature of intelligence, it says, can be described precisely enough for a machine to simulate it. Seventy years on, that sentence is still the bet. AI Magazine reprinted it in 2006, with a free PDF.
• Norbert Wiener, “Some Moral and Technical Consequences of Automation,” Science (1960) (opens in a new tab). The founder of cybernetics on machines that learn faster than their makers can follow. If we hand such a machine a purpose, he writes, we had better be sure it is the purpose we really want. That is the alignment problem, stated sixty years early, in four pages. Behind the journal’s paywall; a library has it.
• J. C. R. Licklider, “Man-Computer Symbiosis,” IRE Transactions on Human Factors in Electronics (1960) (opens in a new tab). Not a machine in place of a person, and not a person running a machine, but the two thinking together, each doing what it does best. It is the partnership these courses assume. Free at MIT.
• I. J. Good, “Speculations Concerning the First Ultraintelligent Machine,” Advances in Computers (1965) (opens in a new tab). A Bletchley Park codebreaker reasons that a machine able to improve its own design would set off what he calls an intelligence explosion. It would be “the last invention that man need ever make.” Everything written since about superintelligence starts here. Behind the publisher’s paywall; a library has it.

What a machine can and cannot do

• Joseph Weizenbaum, Computer Power and Human Reason: From Judgment to Calculation (W. H. Freeman, 1976) (opens in a new tab). Weizenbaum wrote ELIZA, the 1966 program that played a therapist, and watched people confide in it. This book is his answer. Some decisions call for judgment, which grows out of a human life, and no computer should be asked to make them, however well it could. Start with the introduction, where he tells the ELIZA story.
• John Searle, “Minds, Brains, and Programs,” Behavioral and Brain Sciences (1980) (opens in a new tab). The Chinese Room. A man who knows no Chinese follows a rulebook to answer questions in Chinese, and his answers pass. Does he understand Chinese? Searle says no, and that a program is in the same position. Decades of replies have not closed the question. Free, in the author’s deposited copy.
• Hubert Dreyfus, “Why Heideggerian AI Failed and How Fixing It Would Require Making It More Heideggerian,” Artificial Intelligence (2007) (opens in a new tab). Dreyfus began arguing in 1965 that a machine working from rules and symbols would never do what a person does without thinking. Forty years on, this essay is his summary: why those programs stalled, and what a machine that lives in a world, not a model of one, would need. Start with the opening pages, where he tells the story of the field as he watched it. Free on the journal’s site.

How today’s machines work

• Ashish Vaswani and seven coauthors, “Attention Is All You Need” (2017) (opens in a new tab). The paper behind every chatbot: the Transformer, a design that lets a model weigh each word in a passage against every other. It is written for specialists. Read the first two pages and look at Figure 1, and you will know what the T in GPT stands for. Free at arXiv.
• Richard Sutton, “The Bitter Lesson” (2019) (opens in a new tab). Two pages from a founder of reinforcement learning. Seventy years of AI research teach one lesson, he says, and researchers keep refusing to learn it. General methods that use more computation beat methods built on human knowledge, and by a wide margin. It explains why today’s models are built the way they are. Free, on his own site.

What could go wrong

• Nick Bostrom, Superintelligence: Paths, Dangers, Strategies (Oxford University Press, 2014) (opens in a new tab). The book that made the control problem respectable: if a machine comes to outthink us, how do we keep it doing what we meant? Its paperclip maximizer, a machine that turns everything into paperclips because that is all it was asked to do, is now a figure of speech.
• Stuart Russell, Human Compatible: Artificial Intelligence and the Problem of Control (Penguin Books, 2019) (opens in a new tab). Russell wrote the textbook most AI researchers learned from. Here he argues that its standard model, a machine pursuing a fixed objective, is the mistake. A machine should stay uncertain about what people want and keep asking. The clearest statement of the control problem from inside the field.
• Cathy O’Neil, Weapons of Math Destruction: How Big Data Increases Inequality and Threatens Democracy (Crown, 2016) (opens in a new tab). A mathematician turned Wall Street quant on the models that score teachers, sort résumés, price insurance, and set parole. Her test for a harmful one: it is opaque, it works at scale, and it damages the people it scores. The book that taught the public to ask what a model is for.
• Julia Angwin, Jeff Larson, Surya Mattu, and Lauren Kirchner, “Machine Bias,” ProPublica (2016) (opens in a new tab). The investigation of COMPAS, a score courts used to predict who would offend again. The reporters compared the scores with what defendants did next and found the errors ran in opposite directions for Black and white defendants. Every argument since about fairness in an algorithm has had to answer it. The reporters published their data and their method with it.
• Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell, “On the Dangers of Stochastic Parrots: Can Language Models Be Too Big?” FAccT (2021) (opens in a new tab). The paper that gave the era its phrase. A language model, it argues, stitches together forms of language it has seen without any grasp of what they mean. The bigger it gets, the more that costs: in the data it swallows, the energy it burns, and the harm to the people whose words it learned from. Open access at the ACM.

The human person

• Dicastery for the Doctrine of the Faith and Dicastery for Culture and Education, Antiqua et nova: Note on the Relationship Between Artificial Intelligence and Human Intelligence (2025) (opens in a new tab). The Vatican’s doctrinal note on AI, in 117 numbered paragraphs, released January 28, 2025. The first half draws the line between what a machine does and what a person does, which includes judgment, relationship, and responsibility. The second half applies it, field by field: education, work, health, relationships, war. The groundwork for the encyclical that follows.
• Pope Leo XIV, Magnifica humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence (2026) (opens in a new tab). Leo XIV’s first encyclical, dated May 15, 2026, and released ten days later, 135 years after Rerum novarum, which it takes as its model. Five chapters on what AI does to work, education, family life, politics, war, and public truth, and on who holds the power behind it. Its answer is shared standards of social justice, so that AI serves human dignity and the common good instead of concentrating power. Start with the introduction, then Chapter Four, on safeguarding humanity at a time of transformation.

Stories

• Mary Shelley, Frankenstein; or, The Modern Prometheus (1818) (opens in a new tab). Every story about a made mind starts here. Victor Frankenstein builds a being, recoils from it, and refuses it the one thing it asks for. The harm follows from the refusal, not the making. The 1818 text, free at Project Gutenberg.
• E. M. Forster, “The Machine Stops” (1909) (opens in a new tab). People live alone in cells underground, served by a Machine that feeds them, connects them, and thinks for them. Then it stops. Written in 1909; read it for how familiar it is. The opening story in The Eternal Moment and Other Stories, free at Project Gutenberg.
• Karel Čapek, R.U.R. (Rossum’s Universal Robots) (1920) (opens in a new tab). The play that gave English the word robot, from the Czech for forced labor. A factory makes artificial workers. The workers are cheap, tireless, and then unwilling. Free at Project Gutenberg, in the Selver and Playfair translation.
• Isaac Asimov, “Runaround” (1942), in I, Robot (1950) (opens in a new tab). The story that states the Three Laws of Robotics, and the first in which they fail: a robot caught between two of the laws runs in circles. Every proposal to write rules for machines is arguing with Asimov, whether it knows it or not. Collected in I, Robot, with eight more.