The two programs most often named as the first working artificial intelligence are the Logic Theorist, built by Allen Newell, Herbert Simon, and Cliff Shaw at the RAND Corporation in 1955 and 1956, and Arthur Samuel's checkers player, reported in the IBM Journal of Research and Development in July 1959. The first proved theorems of symbolic logic by searching with rules of thumb. The second improved its own play by storing positions and adjusting a scoring formula. Both were narrow, and the record shows that the strongest claims made for them need dates attached.
Key takeaways
- The Logic Theorist was not yet running on a computer at the Dartmouth workshop in summer 1956. It was described on paper and tested by hand simulation, and machine runs came in late 1956 and early 1957.
- The widely quoted result, 38 of the first 52 theorems in Chapter 2 of Principia Mathematica, comes from those 1957 machine runs.
- Samuel reported that his program learned to play better than its author after 8 or 10 hours of machine-playing time, yet he rated it a better-than-average novice, not an expert.
- A 2026 restoration of the Logic Theorist from 1963 source code proved 16 of 23 attempted theorems, which shows how much early AI behavior can be checked only by running the original code.
- Calling either program "the first" depends on the definition used, and the sources themselves disagree on which came earlier.
The Logic Theorist: proofs by search with subgoals
Newell, Simon, and Shaw set out to simulate complex thought by manipulating symbols. Their 1956 RAND report P-868, published the same year in IRE Transactions on Information Theory, described a program that finds proofs in propositional logic. A 2026 historical study by Jeff Shrager says the program introduced heuristic search strategies, namely means-ends analysis and subgoal decomposition, which the team derived from studies of how people solve problems. To write it, the team created a programming language, the Information Processing Language, which introduced list handling and dynamic memory and which Shrager calls a direct ancestor of Lisp.
The timeline is easy to blur. The study states that when the Dartmouth workshop met in summer 1956, the program was not yet running on a machine. Newell and Simon described it as essentially specified, and hand simulations suggested it would prove the 60-odd theorems of Chapter 2 of Principia Mathematica. Family members and students acted as program components in those simulations. Machine-executable proofs appeared only with the implementation on the JOHNNIAC computer in late 1956 and early 1957, and the 38-of-52 figure stems from those runs. The program also found a shorter proof of Theorem 2.85 than the one in the book.
Samuel's checkers player: learning from stored positions
Samuel chose checkers over chess because simpler rules let him concentrate on learning. He listed what a good test activity needs: no practical algorithm that guarantees a win, a definite goal, known rules, a body of knowledge to test progress against, and familiarity to many people. He noted that the general methods for programming game play were described by Claude Shannon in 1950 for chess and that his basic program resembled one described by Christopher Strachey in 1952. His machine was an IBM 704.
Two learning procedures were studied. In rote learning, the program saves positions it has seen along with their scores. An illustrative example, following Samuel's own description: if the program normally looks three moves ahead and later reaches a saved position whose score was already computed three moves ahead, the stored score counts as a six-move score. Each saved result extends the effective search depth over time. The second procedure adjusted the coefficients of a scoring formula built from board features. Samuel's first formula had four terms: piece advantage, denial of occupancy, mobility, and a term combining center control and piece advancement. Later experiments used 16.
His abstract reports that the program learned to play a better game than its writer in a remarkably short time, 8 or 10 hours of machine-playing time, when given only the rules, a sense of direction, and a redundant list of features. The detailed results are more modest. After rote learning, the memory tape held over 53,000 board positions, and the program had learned a good opening game and to spot many winning and losing endings, but its midgame play was not greatly improved. Samuel judged it a better-than-average novice and not an expert, and estimated that at least 20 times as many stored positions would be needed to improve the midgame significantly.
The two programs side by side
| Feature | Logic Theorist | Samuel's checkers player |
|---|---|---|
| Task | Prove theorems of propositional logic | Play checkers and improve with experience |
| Main idea | Heuristic search with subgoals | Look-ahead search plus stored positions and tuned scoring |
| Machine | Hand simulation in 1956, JOHNNIAC in 1957 | IBM 704 |
| Headline result | 38 of the first 52 Chapter 2 theorems (1957 runs) | Better than its author after 8 or 10 hours of machine play, as reported by Samuel |
| Stated limit | Early versions were specifications and hand simulations before machine runs | Midgame play not greatly improved by rote learning; learning in later tests was erratic and none too stable |
What "first" can and cannot mean
The 2026 study calls the Logic Theorist widely regarded as the first AI program, yet a footnote in the same paper observes that Samuel's checkers program existed a couple of years before it, and that Samuel built arguably the first machine learning program and the first system to learn from self-play. The difference lies in aim. Samuel wanted a checkers player, while Newell, Simon, and Shaw wanted to model human thinking across many domains. A reader who needs one answer should say which definition is in use: first program to prove theorems, first to learn from play, or first described as AI at a named meeting.
The same study offers a warning about verification. Its author rebuilt a working Logic Theorist from code transcribed from a 1963 RAND report, and the revived program proved 16 of 23 attempted theorems from Chapter 2. That code is a later re-coding of the 1956 conceptual model, not the 1957 program, so the figure supports the general behavior but is not a replication of the 38-of-52 result.
FAQ
Did the Logic Theorist run at the Dartmouth workshop?
According to the 2026 historical study, no. The program was essentially specified and had been simulated by hand, and machine proofs appeared on the JOHNNIAC in late 1956 and early 1957.
Was Samuel's program an example of machine learning?
Yes. Samuel titled his paper Some Studies in Machine Learning Using the Game of Checkers, and it reports two learning procedures, rote learning and adjusting the weights of a scoring formula.
Could these programs do anything outside their domain?
Not by design. The Logic Theorist handled formal logic, and Samuel's program handled checkers. Samuel argued that the principles might apply to other situations, but he reported results only for checkers.
Why does the difference between hand simulation and machine runs matter?
It separates an idea from a measured result. A claim that a program proved 38 theorems refers to a specific implementation on specific hardware at a specific time, and earlier descriptions did not have that evidence.
Related reading
- Where AI technology began
- The people behind early AI
- Why researchers set out to build AI
- How Canadian parliamentary records trained a 1990 translator
More in AI Basics.
Last verified: October 2026
Sources
- Executable Archaeology: Reanimating the Logic Theorist from its IPL-V Source, Shrager J, arXiv:2603.13514, March 13, 2026 (arXiv record; preprint)
- The Logic Theory Machine: A Complex Information Processing System, Newell A, Simon HA, RAND report P-868, 1956; IRE Transactions on Information Theory IT-2(3), pp. 61-79, September 1956
- Some Studies in Machine Learning Using the Game of Checkers, Samuel AL, IBM Journal of Research and Development 3(3), pp. 210-229, July 1959