John McCarthy
Last reviewed
Apr 28, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 ยท 4,500 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Apr 28, 2026
Sources
16 citations
Review status
Source-backed
Revision
v1 ยท 4,500 words
Add missing citations, update stale details, or suggest a clearer explanation.
John McCarthy (September 4, 1927 to October 24, 2011) was an American computer scientist and cognitive scientist widely regarded as one of the founders of artificial intelligence as a research discipline. McCarthy coined the phrase "artificial intelligence" in 1955 when he drafted, together with Marvin Minsky, Nathaniel Rochester, and Claude Shannon, the proposal for the 1956 Dartmouth Conference [1][7]. He went on to invent the Lisp programming language in 1958, propose the first practical scheme for general-purpose time-sharing computing in a 1959 memorandum, develop automatic memory management through garbage collection, and lay much of the formal groundwork for logic-based and knowledge-based AI through papers such as Programs with Common Sense (1959), Some Philosophical Problems from the Standpoint of Artificial Intelligence (1969), and Circumscription, A Form of Non-Monotonic Reasoning (1980) [2][3][4][5].
McCarthy founded the AI Project at the Massachusetts Institute of Technology in 1958, which formally became the MIT Artificial Intelligence Laboratory in 1959, co-led with Marvin Minsky. After moving west in 1962 he founded the Stanford Artificial Intelligence Laboratory (SAIL) at Stanford University [8][13]. He received the ACM A.M. Turing Award in 1971, the Kyoto Prize in 1988, the United States National Medal of Science in 1990, and the Benjamin Franklin Medal in 2003 [7][14][15]. For more than half a century he championed a particular vision of intelligent machines: thinking can be captured by symbolic logic, and an intelligent program ought to be told facts about the world in a declarative formal language and then reason from those facts using rules of inference. This commitment helped define the tradition that became known as symbolic AI, or Good Old-Fashioned AI (GOFAI), in deliberate contrast to the connectionist and statistical methods that dominate modern AI [6].
McCarthy spent the bulk of his academic life on the Stanford faculty, retiring formally in 2000 but continuing as professor emeritus until his death in 2011 [8][10]. He returned again and again to the same large question that had drawn him into the field as a young graduate student: what would it take to build a machine with genuine common sense?
John McCarthy was born in Boston, Massachusetts, on September 4, 1927. His father, John Patrick McCarthy, was an Irish Catholic immigrant who worked as a carpenter, fisherman, and union organizer. His mother, Ida Glatt McCarthy, was a Lithuanian Jewish immigrant who worked as a journalist for the Federated Press wire service. Both parents were politically active on the left. The family moved several times during the Depression years, first to New York City and then to Los Angeles [12][13].
McCarthy showed exceptional mathematical ability as a teenager. While still in high school in Los Angeles he taught himself college-level calculus from the textbooks used at the California Institute of Technology (Caltech), so that when he enrolled at Caltech in 1944 he was permitted to skip the first two years of the undergraduate mathematics sequence. He completed his B.S. in mathematics at Caltech in 1948 [12].
A pivotal moment came in September 1948, when McCarthy attended the Hixon Symposium on Cerebral Mechanisms in Behavior at Caltech. Among the speakers was John von Neumann, whose talk on automata and the analogy between brains and computing machines deeply impressed the young mathematician. McCarthy later said that this symposium planted the idea that one might design a machine that could think [12][13]. He went on to Princeton University and earned his Ph.D. in mathematics there in 1951 with a dissertation entitled Projection Operators and Partial Differential Equations, written under Solomon Lefschetz with guidance from Donald C. Spencer [12]. At Princeton he befriended Marvin Minsky, then a fellow graduate student in mathematics, and the two began discussing how machines might one day be made to reason.
After completing his doctorate, McCarthy held positions at Princeton (1951 to 1953), Stanford (1953 to 1955), and Dartmouth (1955 to 1958), before returning east to MIT in 1958 as an assistant professor of communication sciences. In 1962 he accepted a full professorship at Stanford and relocated permanently to California, where he remained for the rest of his life [8][12].
In the summer of 1955, McCarthy was a young assistant professor of mathematics at Dartmouth College in Hanover, New Hampshire. Together with Marvin Minsky (then at Harvard), Nathaniel Rochester (manager of information research at IBM and chief designer of the IBM 701), and Claude Shannon (the founder of information theory, then at Bell Telephone Laboratories), he drafted a funding proposal to the Rockefeller Foundation titled A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence, dated August 31, 1955 [1].
The choice of the phrase "artificial intelligence" was McCarthy's. He used it deliberately to distinguish the new field from cybernetics, Norbert Wiener's competing intellectual program, and from the term "automata studies," which he had used the year before but came to feel was too narrow. McCarthy wanted a name that did not commit the field to any particular technique or philosophical school. The proposal opens with one of the most quoted sentences in the history of computer science: "The study is to proceed on the basis of the conjecture that every 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" [1].
The Rockefeller Foundation funded a scaled-down version of the proposal, and the Dartmouth Summer Research Project on Artificial Intelligence took place from June 18 through August 17, 1956. Roughly twenty researchers attended at various points, including Allen Newell, Herbert A. Simon, Arthur Samuel, Ray Solomonoff, Trenchard More, and Oliver Selfridge, in addition to the four organizers. The workshop did not produce immediate technical breakthroughs, but it established artificial intelligence as a named field, brought its founding personalities into contact with each other, and gave them a shared vocabulary. The 1956 workshop is now universally recognized as the birth of AI as a discipline [1][13].
In 1958, shortly after arriving at MIT, McCarthy began working on a programming language suited to symbolic computation rather than to numerical calculation. The dominant scientific language of the day was FORTRAN, designed at IBM in 1957 for floating-point arithmetic on the IBM 704. McCarthy wanted something that could manipulate algebraic expressions, lists of symbols, and logical formulas, since these were the data structures he believed an AI program would need. The result was Lisp, short for LISt Processor. Lisp is the second-oldest high-level programming language still in use, after FORTRAN, and many of its concepts have since become standard [3][16].
McCarthy first sketched Lisp in a series of MIT AI Lab memos in 1958 and 1959. The design was published in his classic paper Recursive Functions of Symbolic Expressions and Their Computation by Machine, Part I, which appeared in the Communications of the ACM in April 1960 [3]. The paper introduced several ideas that have since become commonplace:
The first working Lisp implementation was written by McCarthy's graduate student Steve Russell, who hand-translated the eval function from the paper into IBM 704 machine code. Once Russell finished, Lisp had a working interpreter and immediately became the language of choice for AI research at MIT and around the world [16].
In the late 1950s, computers were used in batch mode: a programmer submitted a deck of punched cards, waited hours or days, and received a printout. McCarthy found this cycle intolerably slow for AI research, where one wanted to try an idea and adjust quickly.
In a memo to MIT colleague Philip Morse dated January 1, 1959, McCarthy described what he believed was needed: a system in which several users would share a single computer simultaneously, each with the illusion that the machine was responding to them alone [8]. McCarthy's memo is generally credited as the first written description of general-purpose computer time-sharing. It inspired Fernando Corbato and his team at the MIT Computation Center to build the Compatible Time-Sharing System (CTSS), demonstrated in 1961 and operational by 1963. CTSS in turn led to Project MAC at MIT, to the Multics operating system, and ultimately to Unix [8][13].
In a 1961 talk at MIT's centennial, McCarthy predicted that "computation may someday be organized as a public utility," much as electricity and the telephone were. The prediction is now routinely cited as an early articulation of the idea behind cloud computing [8].
In 1958, McCarthy and Marvin Minsky began the MIT AI Project, which was reorganized in 1959 as the Artificial Intelligence Laboratory and remained one of the world's premier AI research centers for decades. The MIT AI Lab was the original home of Lisp and of much foundational work on machine vision and robotics [12][13].
In 1962, McCarthy left MIT for Stanford and transferred to the newly formed Department of Computer Science when it was established in 1965. With Lester Earnest he founded the Stanford Artificial Intelligence Project, later renamed the Stanford Artificial Intelligence Laboratory (SAIL). He served as its director from 1965 until 1980. SAIL was housed in a former rocket research site in the foothills above the main campus and became famous for work on machine vision, speech understanding, robot manipulation, autonomous vehicles, and computer music [8][10].
In 1965 and 1966, McCarthy organized one of the first international computer chess matches. The Kotok-McCarthy chess program, written by his MIT students, played a four-game correspondence match against a Soviet program developed at the Institute of Theoretical and Experimental Physics in Moscow. Moves were exchanged by telegraph over a nine-month period in 1966 and 1967. The Soviet program won three games and drew one [10][13].
In September 1958, at a conference held at the National Physical Laboratory in Teddington, England, McCarthy presented a paper now regarded as the founding document of logical, knowledge-based AI: Programs with Common Sense [2]. The paper described a hypothetical program McCarthy called the Advice Taker.
The Advice Taker was not a working program but a sketch of an architecture. McCarthy proposed that an intelligent program should represent its knowledge of the world as a collection of declarative sentences in a formal language (most likely a fragment of first-order predicate calculus), and that it should draw conclusions from those sentences using a built-in deduction procedure. Crucially, the program would be improvable by being told new things in the same formal language, rather than being reprogrammed. McCarthy gave a worked example involving a person at home who wanted to get to the airport: the program was supposed to deduce, from general statements about transportation and local geography, the action sequence "walk to the car, drive to the airport" [2].
Programs with Common Sense may be the first published proposal to use formal logic as the representation language of a thinking machine. It argued that common-sense reasoning, the everyday knowledge humans use without effort, would be the central technical problem of AI. Six decades later, the problem of giving machines common sense remains an open challenge, and the descendants of McCarthy's Advice Taker include knowledge graphs, the ontology-based expert system tradition, and projects such as Cyc.
In 1969, McCarthy and his Stanford colleague Patrick J. Hayes published Some Philosophical Problems from the Standpoint of Artificial Intelligence in the volume Machine Intelligence 4 [4]. The paper introduced the situation calculus, a formalism for representing actions and their effects in first-order logic.
In situation calculus, the world is described by a sequence of situations, each a snapshot of how things are. Actions are functions that map one situation to another, and properties of the world (which McCarthy called fluents) are predicates whose truth depends on the situation. With this machinery, one can write in pure logic statements such as "if a robot is at location A in situation s and performs move(A, B), then in the resulting situation it is at location B." The situation calculus has remained a central representation in logical AI and in research on planning and the semantics of action.
The same paper introduced the now-famous frame problem: the difficulty of saying, in logic, which features of a situation do not change when an action is performed. If a robot moves from one room to another, the colors of the walls and the positions of the chairs in the original room do not change. Stating all the things that do not change after every action, in classical logic, requires an enormous number of axioms, and there is no elegant way around the problem within first-order logic. The frame problem became one of the most-discussed technical and philosophical issues in AI during the 1970s and 1980s [4].
Classical first-order logic is monotonic: adding new premises can only add new conclusions, never retract old ones. But everyday reasoning is not like that. If someone tells you Tweety is a bird, you conclude that Tweety can fly. If they then tell you Tweety is a penguin, you withdraw the conclusion. McCarthy's response was a new formal device he called circumscription, introduced in his 1980 paper Circumscription, A Form of Non-Monotonic Reasoning, published in the journal Artificial Intelligence [5].
Circumscription is, in essence, a way of saying "and nothing else." Given a logical theory and a designated predicate, the circumscription of that predicate is the strengthened theory in which the extension of the predicate is taken to be as small as the original theory permits. By circumscribing a predicate such as "abnormal," one can express defeasible defaults of the form "birds that are not abnormal can fly," together with the implicit assumption that nothing is abnormal unless evidence forces us to think otherwise. If new information later marks Tweety as abnormal, the conclusion that Tweety can fly is automatically retracted [5].
The 1980 issue of Artificial Intelligence containing McCarthy's circumscription paper also included foundational papers by Drew McDermott and Jon Doyle on non-monotonic logic and by Raymond Reiter on default logic. That issue is widely viewed as the moment non-monotonic reasoning matured into an accepted subfield. McCarthy followed up in 1986 with Applications of Circumscription to Formalizing Common Sense Knowledge, applying the formalism to the frame problem, the qualification problem, and other common-sense puzzles.
McCarthy and Hayes's 1969 paper introduced a distinction that became central to McCarthy's thinking for the rest of his career: the distinction between the epistemological and the heuristic aspects of an intelligent program [4].
A representation is epistemologically adequate, in McCarthy's sense, if it can express the facts the program needs to know to behave intelligently. A representation is heuristically adequate if the reasoning processes that operate over it are computationally tractable enough to solve problems in real time. McCarthy argued that AI research had spent too much effort on heuristic adequacy and not enough on epistemological adequacy. Until we can write down, in some formal language, what an intelligent agent needs to know about the everyday world (objects, events, time, causation, knowledge, belief, intention, ability), no amount of cleverness in the search procedure will produce genuine common-sense reasoning [6].
This emphasis on representation, declarative knowledge, and explicit logical inference defined what came to be called logical AI and placed McCarthy at the head of a tradition often called GOFAI (Good Old-Fashioned AI), or more neutrally symbolic AI. It stood in deliberate contrast to the connectionist tradition associated with Frank Rosenblatt's perceptrons and, much later, with Geoffrey Hinton, Yann LeCun, Yoshua Bengio, and deep learning. McCarthy was respectful of statistical and learning approaches but consistently argued that they could not by themselves give a machine the ability to reason about a novel situation it had not seen during training. He defended this position in Generality in Artificial Intelligence, his 1971 Turing Award lecture, eventually published in 1987 in Communications of the ACM [6].
McCarthy's interests ranged beyond core AI into programming language theory and philosophy. A Basis for a Mathematical Theory of Computation (1963) was an early attempt to give rigorous semantics to computer programs, helping establish the mathematical study of programming languages. As a member of the ACM committee that designed ALGOL 60, he successfully argued for the inclusion of recursive procedure calls and conditional expressions. Ascribing Mental Qualities to Machines (1979) argued that it can be useful to describe a complex computer program in mentalistic vocabulary (saying it "believes" or "knows" or "wants" something), even when the machine is plainly not conscious. He also proposed an early formal Common Business Communication Language (CBCL), decades before XML, and wrote a science-fiction story, The Robot and the Baby, dramatizing his views on machine ethics.
McCarthy received nearly every major recognition available to a computer scientist. The most important honors include the following [7][14][15].
| Year | Award | Awarding Body |
|---|---|---|
| 1971 | A.M. Turing Award | Association for Computing Machinery |
| 1985 | Research Excellence Award (first recipient) | International Joint Conference on Artificial Intelligence (IJCAI) |
| 1988 | Kyoto Prize in Information Science | Inamori Foundation |
| 1990 | National Medal of Science (Mathematical, Statistical, and Computational Sciences) | United States National Science Foundation |
| 2003 | Benjamin Franklin Medal in Computer and Cognitive Science | The Franklin Institute |
| 2011 | AI's Hall of Fame (inaugural class) | IEEE Intelligent Systems / IEEE Computer Society |
In addition, McCarthy was elected a member of the National Academy of Sciences, the National Academy of Engineering, the American Academy of Arts and Sciences, and a fellow of the AAAI. His Turing Award lecture, Generality in Artificial Intelligence, eventually delivered in expanded written form in 1987, is one of the canonical position papers of the logical AI tradition [6][7].
The table below summarizes McCarthy's most consequential technical contributions and their lasting significance.
| Year | Contribution | Significance |
|---|---|---|
| 1955 | Coined the term "artificial intelligence" in the Dartmouth proposal | Named and helped define the field. |
| 1956 | Co-organized the Dartmouth Summer Research Project | Founding event of AI as a discipline. |
| 1958 | Designed the Lisp programming language | Second-oldest high-level language; introduced S-expressions, conditional expressions, recursion as primary control. |
| 1958 | Proposed conditional expressions for ALGOL | Ancestor of the if-then-else expression in every modern language. |
| 1959 | Proposed time-sharing computing | Direct inspiration for CTSS, Multics, and interactive computing. |
| 1959 | Co-founded the MIT AI Project (later Lab) | First university AI research lab in the U.S., with Minsky. |
| 1959 | Wrote Programs with Common Sense | First proposal for declarative, knowledge-based AI. |
| 1960 | Invented automatic garbage collection | Now standard in Java, Python, JavaScript, Go, C#, Ruby, and most modern languages. |
| 1962 | Founded Stanford AI Project (later SAIL) | Director until 1980; one of the most influential AI labs in the world. |
| 1969 | Introduced situation calculus and the frame problem (with Hayes) | Foundational formalism for logical reasoning about action. |
| 1980 | Defined circumscription as a non-monotonic logic | Made common-sense default reasoning expressible in logic; helped launch the subfield of non-monotonic reasoning. |
| 1986 | Applied circumscription to common-sense knowledge | Showed how non-monotonic logic captures defeasible inferences. |
McCarthy's published output spans more than fifty years. The papers below are most often cited as his defining contributions.
| Year | Title | Venue | Topic |
|---|---|---|---|
| 1955 | A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence (with Minsky, Rochester, Shannon) | Rockefeller Foundation proposal | Founding document of AI |
| 1959 | Programs with Common Sense | Teddington Conference on the Mechanization of Thought Processes | Advice Taker, declarative knowledge representation |
| 1960 | Recursive Functions of Symbolic Expressions and Their Computation by Machine, Part I | Communications of the ACM, vol. 3, no. 4 | Lisp, S-expressions, garbage collection |
| 1963 | A Basis for a Mathematical Theory of Computation | Computer Programming and Formal Systems, North-Holland | Mathematical semantics of programs |
| 1969 | Some Philosophical Problems from the Standpoint of Artificial Intelligence (with Hayes) | Machine Intelligence 4, Edinburgh University Press | Situation calculus, frame problem |
| 1979 | Ascribing Mental Qualities to Machines | Philosophical Perspectives in Artificial Intelligence | Intentional stance, machine mentality |
| 1980 | Circumscription, A Form of Non-Monotonic Reasoning | Artificial Intelligence, vol. 13, nos. 1 and 2 | Non-monotonic logic, defaults |
| 1986 | Applications of Circumscription to Formalizing Common Sense Knowledge | Artificial Intelligence, vol. 28 | Frame problem, common-sense reasoning |
| 1987 | Generality in Artificial Intelligence | Communications of the ACM, vol. 30, no. 12 | Written version of 1971 Turing Award lecture |
| 1990 | Formalizing Common Sense (Lifschitz, ed.) | Ablex Publishing | Collected papers on logical AI |
The history of AI since McCarthy's death has gone in directions he did not entirely anticipate. The dramatic successes of deep learning, large language models, and generative AI have come almost entirely from statistical methods. At the same time, the persistent weaknesses of those systems (hallucination, fragility on truly novel reasoning problems, inability to give explicit causal accounts of their own behavior) are exactly the failure modes McCarthy predicted decades earlier for any approach lacking an explicit world model. Contemporary research on neuro-symbolic AI, on grounding language models in formal knowledge bases, and on giving AI systems genuine common sense returns, often without acknowledging it, to McCarthy's central conviction that representation matters [11].
McCarthy was married three times. His second wife, Vera Watson, was a programmer at IBM and an accomplished mountaineer; she died on Annapurna I in 1978 as part of the American Women's Himalayan Expedition. He was survived by his third wife, Carolyn Talcott (a computer scientist at SRI International), and by his daughters Susan and Sarah and son Timothy [9][10].
McCarthy was known for direct and sometimes blunt opinions on political and scientific matters and was a longtime advocate of continued material and technological progress. He continued to publish, supervise students, and attend conferences into his eighties. He officially retired from Stanford in 2000 but remained professor emeritus. He died at his home in Stanford, California, on October 24, 2011, at age 84. His death was covered in the New York Times, IEEE Spectrum, the Guardian, and the technical and popular press around the world [9][10][11].
It is hard to overstate McCarthy's effect on computing. The phrase he coined in 1955 is now used by hundreds of millions of people. The programming language he designed in 1958 is still in active use, and the ideas he put into it (recursion, conditional expressions, garbage collection, the equivalence of code and data) are now part of every working programmer's vocabulary. The interactive style of computing he proposed in 1959 is the style every modern user takes for granted. The research questions he raised in 1959, 1969, and 1980 are still active problems in the field he helped found [10][11][12].
McCarthy left a particular intellectual stance that continues to influence how people think about machine intelligence: one that takes seriously the question of what it would mean for a machine to know something, to believe something, or to reason about a situation it had never encountered. As of the mid-2020s, the deep learning revolution has produced systems that can carry on convincing conversations and generate striking images and code, but it has also rediscovered, the hard way, that representation, knowledge, and explicit reasoning are not optional extras for genuine intelligence. McCarthy spent his career insisting on exactly that point.