John McCarthy: Founding Father of Artificial Intelligence

John McCarthy gave artificial intelligence its name, its first programming language, and its most ambitious early research program. In a proposal dated August 31, 1955, the 27-year-old Dartmouth mathematician asked for money to study whether "every aspect of learning or any other feature of intelligence" could be described precisely enough for a machine to simulate it. Three years later he designed Lisp. This article follows what he built, what stalled, and which of his ideas are returning inside today's language models.

The 1955 proposal that named a field

McCarthy was an assistant professor at Dartmouth College in 1955, with a Caltech bachelor's degree in mathematics (1948) and a Princeton PhD (1951) behind him. He wrote the funding request with three co-signers: Marvin Minsky of Harvard, Nathaniel Rochester of IBM, and Claude Shannon of Bell Telephone Laboratories. The document is titled "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence," and its full text is still hosted at Stanford.

The ask was modest by modern standards. It described a two-month study by 10 people, to be held at Dartmouth in the summer of 1956, and it requested support from the Rockefeller Foundation with estimated expenses of $13,500. Faculty-level participants were budgeted at $1,200 in salary each, and up to two graduate students at $700 each.

The central sentence, which every history of the field quotes, is a conjecture rather than a result:

"[E]very aspect of learning or any other feature of intelligence can in principle be so precisely described that a machine can be made to simulate it." (McCarthy, Minsky, Rochester, and Shannon, 1955)

Think of it as a bet placed before any evidence existed, the way a surveyor might stake a claim on land nobody has crossed yet. The proposal then listed seven topics to work on: automatic computers, how a computer can be programmed to use a language, neuron nets, the theory of the size of a calculation, self-improvement, abstractions, and randomness and creativity.

That list matters for a later argument. Neural networks were on McCarthy's own founding agenda. The popular story that symbolic AI and neural networks were separate camps from the first day is tidier than the record.

What happened at Dartmouth in summer 1956

The workshop was smaller and messier than its reputation. Ray Solomonoff's notes put it at roughly eight weeks, from about June 18 to August 17, 1956, although other accounts say six. His list names 20 people over the whole period, including Minsky, McCarthy, Shannon, Rochester, Oliver Selfridge, Allen Newell, Herbert Simon, John Nash, and Arthur Samuel. Only Solomonoff, Minsky, and McCarthy stayed for the full run, and daily attendance ranged from three to about eight people, according to the Wikipedia account of the workshop, which draws on those notes.

It was not a directed research project. People came and went, argued about method, and left with different convictions. Samuel, who built a checkers-playing program, later described the experience as "very interesting, very stimulating, very exciting." The event is now treated as the founding moment of the field, but it was a founding by naming and networking rather than by results.

Why McCarthy chose the name "artificial intelligence"

The name did real work. According to the account cited in the same article, McCarthy wanted a label that was neutral and not tied to automata theory. He also wanted to avoid "cybernetics," a field heavily focused on analog feedback, partly so he would not have to accept or argue with Norbert Wiener as its guru. A term is a flag planted in territory, and this one claimed the digital, symbolic, programmable part of the map.

What the myths get wrong

Popular summaries compress the story. The table sets the common claims against what the primary sources support.

Common claim	What the record shows
The conference was held in 1955	The proposal is dated August 31, 1955; the workshop ran in summer 1956
It was a formal conference with a published result	It was a loosely organized summer study; attendance varied between three and about eight people per day
It solved or nearly solved AI in two months	The proposal asked for a 2-month, 10-person study; it did not promise a solution
Neural networks were excluded	"Neuron nets" is item three on the proposal's own topic list
DARPA funded the founding	The proposal requested Rockefeller Foundation support

Lisp: a language built to manipulate ideas

By 1957 McCarthy was writing chess move routines in FORTRAN at MIT, and he was frustrated. FORTRAN had no clean way to say "if this condition holds, do this, otherwise that" inside an expression. A helper function called XIF evaluated all three of its arguments, which made it useless for anything that could fail or loop forever. So he invented the conditional expression, which evaluates only the branch it needs. That small idea made recursive definitions practical.

In his own history of Lisp, McCarthy recalls the next step. In summer 1958, at IBM, he used differentiation of algebraic expressions as a test problem. "The idea of differentiation is obviously recursive," he noted: the derivative of a sum is the sum of the derivatives, so a function that differentiates can call itself on the parts. He borrowed Alonzo Church's lambda notation for writing function definitions.

A precursor existed. Rochester and Herbert Gelernter, with McCarthy advising, built a list-processing language inside FORTRAN called FLPL. It had neither conditional expressions nor recursion, and that gap is what Lisp was designed to close. In fall 1958 McCarthy became an assistant professor at MIT, the MIT AI Project started, and Lisp implementation began.

How Lisp treats code as data

Lisp's defining trick is that programs and data share one shape: the list. A list in parentheses is data if you quote it and a program if you evaluate it.

(+ 1 2)      ; evaluated: the program adds 1 and 2, giving 3
'(+ 1 2)     ; quoted: a plain list of three symbols

Because a program is just a list, another program can build, inspect, and rewrite it. Compare a recipe card with a photograph of a recipe card: in most languages the program is the photograph and cannot be edited by the cook, while in Lisp the program is the card itself. That property is why Lisp became the natural language for programs that reason about symbols, including other programs.

A small example shows recursion and conditionals working together:

(defun sum-list (lst)
  (if (null lst) 0
      (+ (car lst) (sum-list (cdr lst)))))

The function asks whether the list is empty. If so, the sum is zero. If not, it adds the first element to the sum of the rest, and the function calls itself on the shorter list until nothing remains.

The accident behind eval

McCarthy's 1960 paper in Communications of the ACM, "Recursive Functions of Symbolic Expressions and Their Computation by Machine, Part I," defined a universal function called eval[e,a] that computes the value of a Lisp expression. He wrote it to show that the language was neater than Turing machines as a foundation for computation, and he expected it to be read, not run. Steve Russell, a graduate student, noticed that eval could serve as an interpreter, and he promptly hand-coded it. The project suddenly had a working language.

It is the programming equivalent of a grammar book written in the language it teaches, and then discovering the book can speak.

The names for taking a list apart came from the hardware. car and cdr descend from IBM 704 instructions: McCarthy noted that the machine "had instructions (connected with indexing) that made these operations easy to implement." They are still used in Lisp dialects today, even though the 704 vanished decades ago.

Ideas that outlived the machine

McCarthy also chose automatic storage reclamation, later called garbage collection. Erasing storage by counting references seemed infeasible to him, so Lisp marks everything reachable and frees the rest. He noted that actual implementation "could be postponed" because only toy examples were being run at the start.

The table lists contributions commonly credited to Lisp and to its early design choices, drawn from the standard history of the language.

Feature	What it does	Where you meet it now
Conditional expressions	Evaluate only the branch that is needed	Every mainstream language has an if-else
Recursion as a first-class idiom	A function defined in terms of itself	Tree traversal, parsers, divide-and-conquer algorithms
Higher-order functions	Functions passed as arguments and returned as values	Map and filter in Python, JavaScript, and others
Garbage collection	The runtime frees unreachable memory	Java, Go, Python, JavaScript
Dynamic typing	Types are checked at run time	Python, Ruby
Read-eval-print loop	Type an expression, see the result immediately	Interactive consoles in most scripting languages

Lisp is described as the second-oldest high-level programming language still in common use, after Fortran. Tim Hart and Mike Levin wrote the first complete Lisp compiler, in Lisp, at MIT in 1962. Dialects followed: Scheme in 1975 from Guy Steele and Gerald Sussman, and Common Lisp, standardized by ANSI in 1994.

The Lisp machine boom and bust

Lisp's appetite for memory and processing time pushed people to build hardware for it. In 1973, Richard Greenblatt and Thomas Knight began the MIT Lisp Machine Project, which led to machines such as the CADR. Companies followed: Symbolics, LMI, Texas Instruments, and Xerox. Total production was perhaps 7,000 units as of 1988.

Then general-purpose workstations and later desktop PCs ran Lisp fast enough at a fraction of the price. By the account in Wikipedia's AI winter article, "an entire industry worth half a billion dollars was replaced in a single year" around 1987. Most Lisp machine makers were gone by the early 1990s. The language survived the hardware, which is the opposite of what usually happens.

The Advice Taker: a program you teach by telling it things

If Lisp was McCarthy's tool, the Advice Taker was his plan. He presented it in 1958 at a symposium on the Mechanization of Thought Processes at the National Physical Laboratory in Teddington, England, and the paper "Programs with Common Sense" is available from Stanford.

Its premise is a complaint. Programs of the time solved problems that demand high intelligence in humans, yet "certain elementary verbal reasoning processes so simple that they can be carried out by any non-feeble minded human have yet to be simulated by machine programs." McCarthy wanted a system that could be improved by conversation, not by a programmer opening it up and rewriting it. His working definition of common sense follows:

"A program has common sense if it automatically deduces for itself a sufficiently wide class of immediate consequences of anything it is told and what it already knows."

He built the design on a short principle: "In order for a program to be capable of learning something it must first be capable of being told it." Samuel's checkers program could tune the weights it gave to board positions, but McCarthy pointed out that no existing scheme could discover an abstraction like the principle of opposition in checkers. The only way he knew to express abstractions was language, so he decided to build a system that reasons verbally.

The airport example

The paper illustrates the idea with an everyday goal. The program is told that it is at a desk, the desk is at home, the car is at home, and home and the airport are both in the same county. It also gets two rules: walking is feasible within the home, and driving is feasible within the county. From the goal of being at the airport, it should deduce that it can walk to the car and then drive to the airport.

The example looks trivial, which is the point. A person solves it without noticing any reasoning. A formal system has to be handed the transitivity of "is at," the conditions for walking and driving, and the rule that doing something changes where you are. McCarthy wrote every one of those as a logical premise.

Declarative versus imperative instruction

The paper contrasts two ways of instructing a machine. His comparison is easy to restate:

	Imperative sentences (do this)	Declarative sentences (this is true)
Speed	A procedure laid out in advance runs faster	Slower, because the machine must reason
Prior knowledge	Starts from a basic state and assumes nothing	Can use what the system already knows
Dependence on order	Order matters a great deal	Meaning depends much less on order
Effect of context	Depends heavily on the current state	Less dependent on the previous state

A thermostat setting is imperative. A statement such as "the guest room is cold" is declarative, and a program with common sense would infer a response. This distinction became a core design question for knowledge-based systems, and McCarthy placed his bet on the declarative side.

A footnote added in 1996 is candid. McCarthy had written that the design would be a joint project with Minsky, and in 1996 he appended: "This was wishful thinking. Minsky's approach to AI was quite different." The two founders pursued diverging programs for the rest of their careers.

The frame problem: why a robot needs to know what stays the same

Representing change is harder than it sounds, and McCarthy ran into the difficulty first. With Patrick Hayes he published "Some Philosophical Problems from the Standpoint of Artificial Intelligence" in Machine Intelligence 4 (1969, pages 463 to 502). The paper uses a formalism called the situation calculus, which he had developed through papers beginning in 1963.

The idea is to treat the world as a sequence of snapshots. A situation is one snapshot, an action is a function that maps one situation to the next, and a fluent is a property whose truth can change from one situation to another, such as "the robot is in room A." Moving to room B produces a new situation in which the robot's location fluent has changed.

The trouble appears at once. Moving does not change the color of the walls, the power level of the robot's battery, or whether the robot is holding a cup. In first-order logic, you would have to write an axiom for every property an action leaves alone. McCarthy and Hayes called this the frame problem: describing a robot's world takes many extra axioms just to state that things do not change arbitrarily.

Imagine a stage manager who must announce, after every scene change, that the furniture has not moved. The announcements would drown the play. What intelligent agents actually need is a default: nothing changes unless something changes it.

Why the frame problem stayed alive

McCarthy proposed handling this by assuming a minimal amount of change, formalized through circumscription. The proposal did not close the case. Steve Hanks and Drew McDermott showed in a 1987 paper in Artificial Intelligence (volume 33, issue 3, pages 379 to 412) that the default-logic solution can give unintended answers. Their example became known as the Yale shooting problem. By the end of the 1980s the original formulation was widely regarded as solved, yet the term lives on for related difficulties, such as handling several actions at once.

Circumscription: reasoning with exceptions

Real reasoning rarely has complete information. Classical logic is monotonic: adding premises never takes a conclusion away. People work differently. You assume a bird can fly, then learn it is a penguin and drop the conclusion. Non-monotonic reasoning is the family of methods that allows this kind of retraction.

McCarthy's contribution was circumscription, introduced in "Circumscription: A Form of Non-Monotonic Reasoning," Artificial Intelligence 13 (1980), pages 27 to 39, and extended in a 1986 paper in volume 28. According to the description on Wikipedia's circumscription page, it formalizes the common-sense assumption that things are as expected unless otherwise specified. It extends first-order logic so that the extension of chosen predicates can be minimized, similar to the closed-world assumption.

For abnormality reasoning, each rule is written with a condition saying "unless this case is abnormal." A rule might say that a bird flies unless it is abnormal. Minimizing the set of abnormal things lets the reasoner assume normality, but only where that is consistent with the known facts. The penguin enters the knowledge base as an exception, and the default conclusion quietly disappears. Think of it as a landlord's "no pets" rule with a list of exceptions that you only consult when something unusual walks in. You do not enumerate every possible pet in advance.

Where circumscription falls short

The method has known limits. It struggles with unstated conditions, which McCarthy illustrated with the missionaries and cannibals puzzle: explicitly stating that a bridge does not exist is not a solution, because there are endless other exceptions nobody listed. Its handling of inertia was later shown to give wrong answers in cases such as the Yale shooting problem. Ray Reiter's coin example shows trouble with disjunctive information, and later researchers proposed repairs such as theory curbing. None of this made the idea useless. It made it a precise tool whose failure modes are well understood.

Time-sharing, the Stanford lab, and chess

McCarthy's influence reached beyond logic and into how people use computers. In 1959 he wrote a memo at MIT that initiated work on a time-sharing system, the idea that many users could appear to have sole access to an expensive mainframe. The Compatible Time-Sharing System (CTSS), led by Fernando Corbató, had a prototype tested by November 1961. The Dartmouth Time-Sharing System began service in March 1964. The Computer History Museum named McCarthy a Fellow in 1999 "for his cofounding of the fields of Artificial Intelligence (AI) and timesharing systems."

He moved to Stanford as a full professor in 1962 and stayed until he retired in 2000. The Computer History Museum records him as the founding director of the Stanford Artificial Intelligence Laboratory in 1965, where the work spanned machine intelligence, interactive graphics, and autonomous vehicles.

The chess program and a telegraph match

Between 1959 and 1962, McCarthy's MIT students Alan Kotok, Elwyn Berlekamp, and others wrote a chess program for the IBM 7090 in FORTRAN and assembly language. They added alpha-beta pruning, a technique that skips branches of the game tree that cannot change the outcome. By 1962 the program had completed fragments of four games at a level comparable to an amateur with about 100 games of experience.

In 1966 and 1967 the Stanford version played a match against a program written at the Institute of Theoretical and Experimental Physics in Moscow. The moves traveled by telegraph, and the match took about nine months. The Soviet program won. Accounts disagree on the exact score, so this article reports only the outcome. Chess worked as a laboratory for the same reason fruit flies work for geneticists: simple rules, hard search, and an unambiguous winner.

Awards

The honors form a short ladder. He won the ACM Turing Award in 1971, the Computer Pioneer Award and the IJCAI Award for Research Excellence in 1985, the Kyoto Prize in 1988, and the National Medal of Science in 1990. Later came the Benjamin Franklin Medal in 2003 and induction into the IEEE AI Hall of Fame in 2011. He died on October 24, 2011, in Stanford, California, at age 84.

Why symbolic AI stalled

By the 1970s, the program McCarthy championed ran into walls. The research community called the main one combinatorial explosion. A method that works on a toy problem fails on a realistic one because the number of possibilities grows faster than any computer can search. Chess has simple rules and an astronomical number of positions. The world has no simple rules at all.

The Lighthill report and the first winter

In 1973 Sir James Lighthill reviewed the field for Britain's Science Research Council. The report said AI had not solved combinatorial explosion and was especially critical of basic research in robotics and language processing. Its summary judgment: "In no part of the field have the discoveries made so far produced the major impact that was then promised." McCarthy took part in a debate on the report at the Royal Institution in London on May 9, 1973, alongside Donald Michie, Richard Gregory, and Lighthill himself.

The consequences were concrete. The report formed the basis for the British government's decision to end support for AI research in most British universities. In the United States the 1969 Mansfield Amendment required DARPA to fund mission-oriented rather than basic research, and in 1974 DARPA cancelled a speech understanding contract at Carnegie Mellon worth roughly $3 million a year.

The picture is not uniform. Historian Thomas Haigh points out that membership in the ACM's AI interest group rose from 1,241 in 1973 to about 3,500 by mid-1978, which suggests that researchers did not experience a collapse. A winter for funding is not always a winter for people.

Expert systems: success that exposed the limits

In the 1980s the symbolic approach found a commercial home in expert systems. An expert system has a knowledge base of facts and rules and an inference engine that applies the rules. Forward chaining starts from known facts and fires rules to derive new ones, and backward chaining starts from a candidate conclusion and works back to check its premises. Both can explain themselves by showing which rules fired.

The flagship was R1, also called XCON, built by John McDermott of Carnegie Mellon in the OPS5 rule language and deployed at Digital Equipment Corporation's plant in Salem, New Hampshire, in 1980. It checked computer orders and configured the components. By 1986 it had processed 80,000 orders at 95 to 98 percent accuracy. DEC was reported to be saving about $25 million a year.

The same project documented the weakness. XCON reached about 2,500 rules at one point, then 6,200 by 1987 and 10,129 by 1989, and a 1987 source describes about half of its rules changing every year. A 1987 maintenance study found that "the problems of continually updating such a large system do not grow linearly."

Year	Event	Why it matters
1966	ALPAC report ends US machine translation support after about $20 million	First major funding reversal
1969	Mansfield Amendment	DARPA shifts toward mission-oriented work
1973	Lighthill report	UK support ends at most universities
1980	XCON deployed at DEC	Symbolic AI's commercial peak begins
1981	Japan's Fifth Generation project funded at $850 million	National bet on logic-based AI; ended 1992
1987	Lisp machine market collapses	Specialized hardware loses to cheaper machines

The knowledge acquisition bottleneck summarizes the problem. Domain experts are valuable and hard to schedule, and getting their knowledge into rules is slow and expensive. Systems also have only superficial knowledge, working inside a closed world with no deeper grasp of concepts until an expert supplies them. In the 1990s the phrase "expert system" largely left the vocabulary of information technology, though rule engines returned later inside business software under quieter names.

Where McCarthy's side was right, and where it was not

It is tempting to score the old debate as a win for learning and a loss for logic. The record supports a more specific verdict.

What the learning-based approach settled. Large neural networks trained on data now do the tasks symbolic AI could not scale to: image recognition, translation, fluent text generation. McCarthy's bet that the path to those abilities ran through hand-written logical knowledge did not pay off. The neural networks and deep learning that took over scale with data and compute instead of with engineers writing rules.

What the old program still contributes. Several of his questions remain open in systems built on large language models. Models generate fluent text that is sometimes false, a failure explained in why AI hallucinates facts, and McCarthy's emphasis on explicit, checkable statements is a counterweight. Interpretability research, covered in mechanistic interpretability, tries to recover the kind of inspectable structure a rule base offered for free. Hybrid designs combine learned components with symbolic search or verification, an active research area often called neuro-symbolic AI.

Where the Advice Taker idea echoes. McCarthy wanted a system that could be improved "merely by making statements to it." Anyone who has instructed a chatbot in plain sentences will recognize the shape, though the mechanism differs completely: a language model does not deduce consequences from premises, and nothing in its design guarantees it will. The echo is in the interface, not the engineering.

The largest attempt to write common sense by hand. Cyc, begun by Douglas Lenat in July 1984 at the Microelectronics and Computer Technology Corporation, aimed to assemble a comprehensive ontology and knowledge base spanning the basic concepts and rules about how the world works. By 2017 it held about 1.5 million terms and a knowledge base of about 24.5 million entries, largely written by hand. It shows both the ambition of McCarthy's common-sense program and the labor it demands.

McCarthy on machine minds

McCarthy defended a deflationary view of mental language. In his 1979 paper "Ascribing Mental Qualities to Machines," he argued that even a thermostat can be said to have beliefs, in a limited sense useful for describing and predicting its behavior. John Searle answered in 1980 with the Chinese Room argument, and McCarthy debated the philosopher Hubert Dreyfus throughout their careers. Whether or not you accept his position, it shaped how the field talks about machines.

His working definition, from the 2007 revision of his explainer on the subject, is plain: AI is "the science and engineering of making intelligent machines, especially intelligent computer programs." He defined intelligence as "the computational part of the ability to achieve goals in the world," and he noted that machines can be somewhat intelligent, since AI has mastered some of the underlying mechanisms but not others. He also said no solid definition of intelligence exists that avoids reference to human intelligence. The definition is deliberately engineering-minded: it asks what a system can do, not what it feels like to be it.

For a wider map of the field these ideas belong to, see the history of artificial intelligence timeline, and for the philosophical stakes of general-purpose machine minds, the explainer on artificial general intelligence.

Frequently Asked Questions

Q: Did John McCarthy invent the term "artificial intelligence"? A: Yes. The term appears in the August 31, 1955 proposal he wrote with Minsky, Rochester, and Shannon, and the Computer History Museum credits him with coining it in 1955. He chose a neutral label that was not tied to automata theory and did not invite comparison with Norbert Wiener's cybernetics.

Q: What is Lisp, and why did McCarthy create it? A: Lisp is a programming language for manipulating symbols and lists, described in a 1960 Communications of the ACM paper. McCarthy wanted a language with conditional expressions and recursion for AI work on the IBM 704, and FORTRAN lacked both. It is the second-oldest high-level language still in common use after Fortran.

Q: What was the Advice Taker? A: It was a proposed program, presented in 1958 and published in 1959, that would reason in a formal language and take actions when it deduced an imperative sentence. Its aim was a system whose behavior could be improved by telling it facts about its environment. It was a design proposal in "Programs with Common Sense," not a finished system.

Q: What are the frame problem and the situation calculus? A: The situation calculus represents the world as a series of situations connected by actions. The frame problem, defined by McCarthy and Patrick Hayes in 1969, is that describing an action's effects in logic seems to require axioms for everything the action does not change. Its original form was considered largely solved by the end of the 1980s, but the term persists for related problems in knowledge representation.

Q: Why did symbolic AI lose ground to neural networks? A: Symbolic systems hit combinatorial explosion, a high cost of entering knowledge, and brittleness outside their narrow domains. The XCON expert system needed constant rule changes as it grew, with about half its rules changing each year by one 1987 account. Neural networks trained on large data sets scaled to perception and language where hand-written rules could not.

Q: Did McCarthy reject neural networks? A: Not at the founding. "Neuron nets" appeared on the 1955 Dartmouth proposal's list of topics, and his 1959 Advice Taker paper discusses nerve nets as one way to represent behavior. His own research program favored logic and explicit knowledge, and his 1996 footnote shows he knew colleagues such as Minsky pursued different approaches.

Q: What awards did John McCarthy win? A: The Turing Award in 1971, the Computer Pioneer Award and IJCAI Award for Research Excellence in 1985, the Kyoto Prize in 1988, the National Medal of Science in 1990, and the Benjamin Franklin Medal in 2003. He was named a Computer History Museum Fellow in 1999 and entered the IEEE AI Hall of Fame in 2011.

The takeaway

McCarthy's lasting achievement is a way of framing the problem, more than any single system. He insisted that intelligence could be described precisely, that programs should be told about the world in a language they can reason over, and that exceptions and change deserve their own mathematics. The engineering path he favored, hand-written logic, ran into costs that data-driven learning avoided. The questions he posed did not go away: how to keep what is stable while change happens, how to hold a default and give it up gracefully, and how to make a machine's knowledge something a person can inspect.

The sharpest irony of his career may be that his best-known artifact outlived his best-known theory. Lisp's conditionals, recursion, garbage collection, and read-eval-print loop sit inside tools used by people who have never heard of him, while the Advice Taker remains a proposal. That asymmetry is a fair summary of how research tends to work: the ideas that survive are often the tools built along the way.

Sources

• McCarthy, J., Minsky, M. L., Rochester, N., and Shannon, C. E. (August 31, 1955). A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence. Stanford University.
• McCarthy, J. (1959). Programs with Common Sense. Paper for the 1958 Teddington symposium on the Mechanization of Thought Processes.
• McCarthy, J. (1979, draft). History of Lisp: prehistory and the implementation of Lisp. Stanford Artificial Intelligence Laboratory.
• McCarthy, J. (revised November 12, 2007). What Is Artificial Intelligence? Stanford University.
• McCarthy, J. (1960). Recursive Functions of Symbolic Expressions and Their Computation by Machine, Part I. Communications of the ACM. Table of contents at Stanford.
• McCarthy, J., and Hayes, P. J. (1969). Some Philosophical Problems from the Standpoint of Artificial Intelligence. Machine Intelligence 4, 463–502.
• McCarthy, J. (1980). Circumscription: A Form of Non-Monotonic Reasoning. Artificial Intelligence 13, 27–39.
• Hanks, S., and McDermott, D. (1987). Nonmonotonic Logic and Temporal Projection. Artificial Intelligence 33(3), 379–412.
• Computer History Museum. John McCarthy, 1999 Fellow profile.
• Wikipedia contributors. John McCarthy (computer scientist), Dartmouth workshop, Lisp (programming language), Lisp machine, Lighthill report, AI winter, Xcon, Expert system, Circumscription (logic), Frame problem, Time-sharing, Kotok-McCarthy, and Cyc. Used as secondary summaries; figures were attributed to the primary works they cite.

---

Read next:

• Alan Turing: The Father of Computer Science and AI
• History of Artificial Intelligence: Timeline
• What Makes Something AI? Defining Artificial Intelligence
• Claude Shannon: The Bell Labs Tinkerer Who Invented Information
• Reinforcement Learning Explained: How AI Learns to Play Games
• History of Computing: From the Abacus to Artificial Intelligence