Some Studies in Machine Learning Using the Game of Checkers

idea is right and the id was filed correctly. Two alternatives are tempting and both should be killed on the way past, because the popular version of this entry is assembled out of them.

The first is moment, and the moment on offer is the 1962 game against Robert Nealey — the single win that entered the folklore of computing. That game is covered at length below, in the section on what this entry gets wrong, and the reason it is not the entry is that it is the least reliable thing about the work: a result that was misdescribed at the time, corrected by its own author in print five years later, and still misquoted sixty years on. Filing it as a moment would make the canon's version of this entry the same shape as the folklore's. The second is prediction, offered by the paper's Conclusions, which are two sentences of forecast. They are graded below, but they are too thin to carry a whole entry, and they are undated, which is most of what is wrong with them.

What survives as idea is a loop: play a game against a copy of yourself, score the resulting positions with a parameterised function, use the difference between what the function said and what the game later showed to adjust the parameters, and repeat. That is what every self-play system since has been a variation on, and it is what the readings now cover under the words "post-training" and "reinforcement learning." Precedent in canon/ is settled: eliza-1966, mycin-1976, logic-theorist-1956 and rosenblatt-perceptron-1958 are all programs or devices filed as idea, while deep-blue-1997 — a contest with a date and a scoreline — is moment.

The title in proposals.md is wrong, and the id's year is four years late. The paper is Some Studies in Machine Learning Using the Game of Checkers, not "Studies in Machine Learning Using Checkers"; the header above carries the real one. More importantly, 1959 is the publication, not the work. The paper's own footnote 3 dates the programme precisely, and it is the single most useful sentence in the reference list:

> The first operating checker program for the IBM 701 was written in 1952. It > was recoded for the IBM 704 in 1954. The first program with learning was > completed in 1955 and demonstrated on television on February 24, 1956.

So the learning program was finished in 1955 and shown to the public in February 1956 — before the Dartmouth workshop of that summer, and before the Logic Theorist ran on a machine. Any claim in this entry about what was built should be dated 1952–56; only the write-up, the second series of tests and the analysis are 1959. This bears directly on logic-theorist-1956, which already concedes the point: Shrager's 2026 paper, quoted there, notes that Samuel's program "existed a couple of years before LT, and it too was heuristic," and "might arguably be the first machine learning program, and the first system to learn from self-play." That concession is correct and this entry supplies its dates.

A small bibliographic snag, recorded so nobody has to rediscover it. Catalogue records give the paper as IBM Journal of Research and Development Vol. 3, No. 3, July 1959, pp. 210–229. Samuel's own 1967 sequel cites it as "211–229." Both are right: page 210 is a full-page board diagram facing the title page. Cite 210–229 and expect to see 211 in older references.

descends_from is a checked citation, not an inference. Samuel names Shannon in the body — "The general methods for doing this were described by Shannon in 1950 as applied to chess rather than checkers" — and footnote 4 of the paper is "C. E. Shannon, 'Programming a Computer for Playing Chess,' Phil. Mag. 41, 256 (March 1950)," which is the exact paper that is shannon-chess-1950. The minimax back-up, the evaluation polynomial and the ply-limited look-ahead are all Shannon's, pointed at a smaller game. That is the whole of the header's claim.

Two other real ancestors are unavailable. The nearer one is Christopher Strachey's draughts program: Samuel writes that "the basic program used in these experiments is quite similar to the program described by Strachey in 1952," and footnote 9 gives Logical or Non-Mathematical Programmes, Proc. ACM Meeting at Toronto, 8–10 September 1952, pp. 46–49. There is no strachey-1952 in canon/ and no line for one in proposals.md; if someone writes it, this entry's header should be edited to descend from it as well. The other is rosenblatt-perceptron-1958, which is in canon/ and which this entry deliberately does not list as an ancestor, because the relationship is a fork rather than a descent — and the fork is one of the most useful things in the paper. It is set out in the next section, and it is graded.

What it is

A checkers program that adjusts its own evaluation function by playing against a copy of itself, published with enough detail to see exactly how it failed.

The abstract states the claim and its price in one breath:

> Two machine-learning procedures have been investigated in some detail using > the game of checkers. Enough work has been done to verify the fact that a > computer can be programmed so that it will learn to play a better game of > checkers than can be played by the person who wrote the program. Furthermore, > it can learn to do this in a remarkably short period of time (8 or 10 hours of > machine-playing time) when given only the rules of the game, a sense of > direction, and a redundant and incomplete list of parameters which are thought > to have something to do with the game, but whose correct signs and relative > weights are unknown and unspecified. The principles of machine learning > verified by these experiments are, of course, applicable to many other > situations.

"A redundant and incomplete list of parameters … whose correct signs and relative weights are unknown and unspecified" is the 1959 statement of the modern bargain: supply the features and the objective, let the machine find the weights. Every word of the qualification matters — redundant, incomplete, signs unknown — and the paper spends most of its length on what goes wrong when you take that bargain.

The fork, stated on the first page. Before any checkers, Samuel divides the field in two:

> At the outset it might be well to distinguish sharply between two general > approaches to the problem of machine learning. One method, which might be > called the Neural-Net Approach, deals with the possibility of inducing > learned behavior into a randomly connected switching net (or its simulation on > a digital computer) as a result of a reward-and-punishment routine. A second, > and much more efficient approach, is to produce the equivalent of a highly > organized network which has been designed to learn only certain specific > things. The first method should lead to the development of general-purpose > learning machines. A comparison between the size of the switching nets that > can be reasonably constructed or simulated at the present time and the size of > the neural nets used by animals, suggests that we have a long way to go before > we obtain practical devices. The second procedure requires reprogramming for > each new application, but it is capable of realization at the present time. > The experiments to be described here were based on this second approach.

Footnote 2, attached to the Neural-Net Approach, cites McCulloch (1949), Farley and Clark, Rochester–Holland–Haibt–Duda, and "F. Rosenblatt ('The Perceptron; A Probabilistic Model for Information Storage and Organization in the Brain,' Psych. Rev. 6, 65, November 1958)" — the paper that is rosenblatt-perceptron-1958, cited by name as the road Samuel was not taking. The reasoning is not dismissal. It is a capacity argument with a date on it: the general-purpose route is the right one and the hardware is not there yet. This is a rare thing in the record — a 1959 paper stating clearly which of two approaches it expects to win in the long run, and choosing the other one because it works now.

The playing machinery. Board positions live in four 36-bit IBM 704 words, one bit per playing square. The program looks ahead, scores the terminal positions of the search with a linear polynomial over hand-defined features, and backs the scores up by minimax — "Carrying this 'minimax' procedure back to the starting point results in the selection of a 'best move.'" The look-ahead depth is not fixed: a minimum of three moves, extended when the next move is a jump, the last move was a jump, or an exchange offer is possible, stopping at ply 5 unless a jump is pending, terminating at ply 11 if one side is more than two kings ahead, and hard-stopping at ply 20 "regardless of all conditions (since the memory space for the look-ahead moves is then exhausted)." Machine playing time "has usually been limited to approximately 30 seconds per move." Evaluating a position with the sixteen-term polynomial cost 7.5 milliseconds.

Two of the features are worth naming because they show what "hand-crafted" meant. MOB is total mobility; DYKE is "credited with 1 for each string of passive pieces that occupy three adjacent diagonal squares." Appendix C defines some thirty of these. Samuel is candid that this was not a triumph:

> Many of the terms in the existing list are related in some vague way to the > parameters used by checker experts. Some of the concepts which checker experts > appear to use have eluded the writer's attempts at definition, and he has been > unable to program them.

Learning method one: rote. Save every board position seen, with its backed-up score, and look it up next time. Because a saved score was itself backed up through three levels, finding it at ply 3 of a later search yields "a 6-ply score rather than a simple 3-ply score," and the amplification compounds across games. The memory tape held "something over 53,000 board positions (averaging 3.8 words each)." Positions carry an age, halved on use and incremented at each merge; when the age hits a limit the position is expunged — "This is a form of forgetting," adopted, Samuel says, "on the basis of reflections as to the frailty of human memories."

The rote learner also required something Samuel calls a sense of direction, and which is a discount factor by another name. Without it, a program that knows every winning king-versus-king variation has no reason to walk toward the win rather than around it. The fix: decrease the magnitude of the score "a small amount each time it is backed up a ply level during the analyses," so that "if the program is now faced with a choice of board positions whose scores differ only by the ply number, it will automatically make the most advantageous choice, choosing a low-ply alternative if winning and a high-ply alternative if losing." He adds: "The significance of this concept of a direction sense should not be overlooked. Even without 'learning,' it is very important. Several of the early attempts at learning failed because the direction sense was not properly taken into account."

Rote learning worked, in a narrow way, and Samuel graded it in the paper's own words: "This program now qualifies as a rather better-than-average novice, but definitely not as an expert." He also priced the extrapolation, which is the kind of number this canon exists to keep: "rough estimates … indicate that a library tape containing at least 20 times the present number of board positions would be needed to improve the midgame play significantly. At the present rate of acquisition of new positions this would require an inordinate amount of play and, consequently, of machine time." A scaling argument, computed by the person who would have benefited from it being wrong.

Learning method two: generalization, and the self-play loop. The second method is the one that matters. The program keeps a pool of 38 parameters, 16 of them active in the polynomial at any time and 22 in reserve. Every eight moves or so, the active term with the lowest correlation coefficient is sent to the bottom of the reserve list and replaced from the head of it.

To train those weights the program is split in two:

> After a number of relatively unsuccessful attempts to have the program > generalize while playing both sides of the game, the program was arranged to > act as two different players, for convenience called Alpha and Beta. Alpha > generalizes on its experience after each move by adjusting the coefficients in > its evaluation polynomial and by replacing terms which appear to be > unimportant by new parameters drawn from a reserve list. Beta, on the > contrary, uses the same evaluation polynomial for the duration of any one > game.

Alpha is the learner; Beta is the frozen checkpoint. "At the end of each self-play game a determination is made of the relative playing ability of Alpha, as compared with Beta, by a neutral portion of the program." Win, and Alpha's scoring system is copied to Beta. Lose, and Alpha collects a black mark; at three black marks "it is assumed to be on the wrong track, and a fairly drastic and arbitrary change is made in its scoring polynomial (by reducing the coefficient of the leading term to zero)."

Anyone who has read a 2020s paper on self-play league training will recognise every component: a learner, a fixed opponent snapshot, a promotion rule, an arbiter that is not either player, and a reset when the learner degenerates.

The update itself is a temporal-difference rule written before the term existed. Alpha records the polynomial's score for each position as it occurs, then compares it against the backed-up score computed at the next move:

> At each play by Alpha the initial board score, as saved from the previous Alpha > move, is compared with the backed-up score for the current position. The > difference between these scores, defined as delta, is used to check the > scoring polynomial. If delta is positive it is reasonable to assume that the > initial board evaluation was in error and those terms which contributed > positively should have been given more weight, while those that contributed > negatively should have been given less weight.

That is the value function being made consistent with its own later self, which is the definition of a temporal-difference method. Sutton and Barto, in the section of Reinforcement Learning: An Introduction devoted to this program, put it plainly: "Samuel's method was the same in concept as that used much later by Tesauro in TD-Gammon."

The paper's best paragraph is the one where it names its own logical problem. Samuel is trying to improve the polynomial, and the only instrument he has for judging the polynomial is the polynomial:

> If Alpha is to make changes in its scoring polynomial, it must be given some > trustworthy criteria for measuring performance. A logical difficulty presents > itself, since the only measuring parameter available is this same scoring > polynomial which the process is designed to improve.

His workaround is the look-ahead: a backed-up score is more trustworthy than a static one, so the static one can be corrected toward it. It is a real answer and an incomplete one, and the incompleteness is what the entry's downstream value rests on.

The failure catalogue. This is the part of the paper that has aged best, and it reads like a list of 2026 problems with 1959 names. Instability: "oscillations quite analogous to the instability induced in electrical circuits by long delays in a feedback loop," fixed with a minimum delta below which no change is made. New terms taking over on nothing: "the too frequent introduction of new terms into the scoring polynomial and the tendency for these new terms to assume dominant positions on the basis of insufficient evidence," fixed by slowing introduction from one every eight moves to one every 32. Credit assignment: "the misassignment of credit to those board positions which permitted spectacular moves when the credit rightfully belonged to earlier board positions which had permitted the necessary groundlaying moves" — and, plainly, "no precise way has yet been devised to insure the correct assignment of credit." Local optima: "the entire learning process is an attempt to find the highest point in multidimensional scoring space in the presence of many secondary maxima on which the program can become trapped." Gaming the opponent: Beta was made to require "a majority of wins on Alpha's part" before adopting Alpha's polynomial, specifically "to avoid the danger of having Beta adopt Alpha's polynomial as a result of a chance win."

The results, stated as Samuel states them, including the confound. First series, 28 games: Alpha won 16 and lost 12, and the quality of play "was extremely poor" at the start. "Some fairly good amateur players who played the machine during this period agreed that it was 'tricky but beatable'." By the end the program "appeared to be approaching a quality of play which caused it to be described as 'a better-than-average player'."

Second series, after the fixes, first 28 games: "18 won by Alpha, and four lost, with six draws. In all cases the games were terminated, if not finished, in 70 moves and a judgment made in terms of the final positions." And then, without being asked, the author knocks his own headline number down:

> Unfortunately, these figures are not strictly comparable because of the > decreased frequency with which Beta adopted Alpha's polynomial during the > second series, both by design and because a programming error immobilized the > adoption procedure during part of the tests.

A 16–12 record improving to 18–4–6 is the sort of figure a press release is built on. Samuel reports it and immediately says a bug in his own training loop means the two series cannot be compared. That sentence is a large part of why this paper belongs in the canon.

He closes the section with a caveat of the kind that has since gone out of fashion:

> As a final precautionary note, it should be stated that these experiments have > not encompassed a sufficiently large series of games to demonstrate > unambiguously that the learning procedure is completely stable or that it will > necessarily lead to the best possible choice of parameters and coefficients.

The two learners had opposite shapes, which is a small finding in its own right and is stated cleanly: rote learning "soon learned to imitate master play during the opening moves," was "always quite poor during the middle game," and handled endgames well; generalization "never learned to play in a conventional manner and its openings are apt to be weak," but "it soon learned to play a good middle game." Memorisation gets you the parts of the space that are enumerable; generalization gets you the parts that are not. Samuel's summary: "rote learning is of the greatest help, either under conditions when the results of any specific action are long delayed, or in those situations where highly specialized techniques are required."

The Conclusions, in full, because they are two sentences:

> As a result of these experiments one can say with some certainty that it is now > possible to devise learning schemes which will greatly outperform an average > person and that such learning schemes may eventually be economically feasible > as applied to real-life problems.

And the sequel, which is where the honest self-grade lives. Samuel returned to the problem in Some Studies in Machine Learning Using the Game of Checkers. II—Recent Progress, IBM Journal of Research and Development Vol. 11, No. 6, November 1967, pp. 601–617, received 5 June 1967, written from Stanford. It replaces the linear polynomial with hierarchical "signature tables," adds real alpha-beta pruning and heavy book learning from master games, and reports a correlation with book moves rising from 0.26 (polynomial) to 0.48 (signature tables) after some 175,000 book moves analysed. Its abstract does not oversell: "While still unable to outplay checker masters, the program's playing ability has been greatly improved." Its introduction is blunter: "the most basic limitation of the known machine learning techniques, as previously outlined, has not yet been overcome nor has the program been able to outplay the best human checker players." Its conclusion, eight years on from the paper this entry is about, is one of the cleanest self-gradings in the literature:

> While the goal outlined in Ref. 1, that of getting the program to generate its > own parameters, remains as far in the future as it seemed to be in 1959, we can > conclude that techniques are now in hand for dealing with many of the tree > pruning and parameter interaction problems which were certainly much less well > understood at the time of the earlier paper.

Why a reading would cite it

Four occasions, all of which the readings have already produced or shortly will.

1. When a lab reports that the gains came from post-training. This is the live one and it recurs weekly. The 14 Aug reading logs Zhipu's GLM-5.3 with the note that it uses the same base as GLM-5.2 and "the improvement is post-training." Samuel is where that claim begins: same machine, same rules, same feature set, a different set of weights arrived at by playing. He is also where the discipline for grading it begins, because he supplies three questions his own results answer and most modern announcements do not. What was the training signal, and was it independent of the thing being trained? Alpha's promotions were judged "by a neutral portion of the program," not by Alpha. How much compute? "8 or 10 hours of machine-playing time," stated in the abstract. Is the headline comparison sound? Samuel's own answer was no, and he said so, because a bug had immobilised part of the adoption procedure. A reading that puts those three questions to a 2026 post-training claim is doing what a 1959 IBM paper did to itself, and may say so.

2. When a system's only judge is a version of itself. This is the deepest use and the one most specific to this entry. Samuel wrote the problem down — "the only measuring parameter available is this same scoring polynomial which the process is designed to improve" — and then built a partial answer to it: a frozen opponent, a neutral arbiter, a promotion threshold requiring a majority rather than a single win, and a hard reset when the learner accumulated three black marks. The 15 Aug midday reading turns on a lab raising its own misalignment risk rating while conceding that its safety benchmarks are saturating and its R&D-acceleration measurement is degrading, with the model too capable to release already in heavy internal use for coding, agentic work and data generation. That is Samuel's logical difficulty at frontier scale: the instrument and the subject have converged, and the party holding both is the only one who can see either. The 1959 paper is not a rebuke. It is a dated demonstration that the problem is structural rather than a lapse, that it was named by the first person to hit it, and that the mitigations available are architectural — hold something fixed, judge from outside the loop, require a margin, reset on degeneration. goodharts-law-1975 is proposed and unwritten and is the natural companion; when it exists, this paragraph should cite it.

3. When self-training is reported to degrade. Samuel's program got worse. He reported it, and he reported the crude intervention that fixed it — zeroing the leading coefficient. Sutton and Barto's diagnosis, given at length in their Samuel section, is the one worth carrying, because it explains the mechanism rather than the symptom:

> But also needed is a way of tying the value function to the true value of the > states. We have enforced this via rewards and by discounting or giving a fixed > value to the terminal state. But Samuel's method included no rewards and no > special treatment of the terminal positions of games. As Samuel himself pointed > out, his value function could have become consistent merely by giving a > constant value to all positions.

And their conclusion: "Since Samuel's learning procedure was not constrained to find useful evaluation functions, it should have been possible for it to become worse with experience. In fact, Samuel reported observing this during extensive self-play training sessions." A self-improvement loop with nothing anchoring it to the outside world can converge on being beautifully self-consistent and useless. That is a mechanism, from 1959 and 1998 respectively, for a family of 2026 results about models trained on their own output collapsing, and about optimisation against a learned reward drifting away from what the reward was supposed to stand for. A reading meeting either has an ancestor with a named cause, and the cause is not "the model is deceptive" — it is that the loop had no fixed point outside itself.

4. When one win over one named human is offered as a capability claim. The 1962 Nealey game is the cleanest historical case of a single result outrunning the system that produced it, and the correction is unusually well documented: Samuel's own footnote in the 1967 paper records the world champion beating the program four games to nil by mail. Cite it beside deep-blue-1997 and watson-jeopardy-2011 — both of which are in canon/ and both of which are contests the machine actually won — to make the distinction that matters. The question to ask of a demonstration is not whether it happened but what happened next, and this entry is where the answer to "what happened next" was published by the winner.

What this entry should not be cited for. It bears on no limit result. It says nothing about whether machines think; Samuel does not raise the question and footnote 1 waves at "the philosophical aspects" only to set them aside. And it is not evidence that self-play works in general — it is evidence that self-play worked once, on a decidable two-player game with a cheap, dense correlate of winning (piece advantage) available as an anchor. Samuel says so himself in choosing checkers over chess: "the simplicity of its rules permits greater emphasis to be placed on learning techniques." A reading that transfers the result to a domain with no such correlate is doing something the paper does not license.

What it got right, and what it got wrong

Not required for idea. Included because the central bet has now been graded twice — once by its author in 1967, and once by the field between 2012 and 2017 — and because the resulting arc is the same arc as shannon-chess-1950 and logic-theorist-1956, which at three instances stops being a coincidence.

Claims dated 1959, except where the 1967 sequel is named. Graded 2026-08-16.

Right, and it is still the architecture. A learner, a frozen opponent, a value function over features, an update driven by the gap between an early estimate and a later one, an arbiter outside both players, and a reset rule. That is a description of Samuel's Alpha and Beta and it is a description of self-play reinforcement learning as practised now. The lineage is not a narrative convenience; it is in the citations. DeepMind's A general reinforcement learning algorithm that masters chess, shogi and Go through self-play opens by saying "A long-standing ambition of artificial intelligence has been to create programs that can instead learn for themselves from first principles (5, 6)," and reference 5 of that paper is, verbatim, "A. L. Samuel, IBM Journal of Research and Development 11, 601 (1967)." Reference 6 is Tesauro's TD-Gammon. Two ancestors are named for the ambition AlphaZero claims to fulfil, and Samuel is the first of them. I read that reference list to check it.

Right, and under-credited: the paper reports its own confounds. The immobilised adoption procedure, the 70-move termination rule, the stability caveat, the note that the second series is "not strictly comparable" to the first. Every one of those makes the result look smaller. The readings' standing complaint — that capability numbers arrive vendor-reported with no independent evaluation, a caveat carried in some form by all three digests written so far — has a 1959 counterexample, and it is a counterexample from a corporate research lab publishing about its own product.

Right, with a caveat the legend removes: "a better game of checkers than can be played by the person who wrote the program." True as stated. But the 1967 paper describes that person as "the programmer, who is not a very good checker player," and the 1959 paper says he was "unable to program" concepts good players use. Beating Samuel at checkers was not a high bar and he never pretended it was. The claim is right; the sentence people build out of it — that the program surpassed its creator in some general sense — is not in the paper.

Wrong, and the author was the one who said so: the parameters stayed hand-made. The 1959 paper poses the goal and admits the gap: "It might be argued that this procedure of having the program select terms for the evaluation polynomial from a supplied list is much too simple and that the program should generate the terms for itself. Unfortunately, no satisfactory scheme for doing this has yet been devised." Eight years later the goal "remains as far in the future as it seemed to be in 1959."

Claim made 1959, restated as unmet 1967, cashed out in the 2010s. Feature learning arrived when the features stopped being written by hand — the ImageNet result of 2012 (proposed as alexnet-2012, unwritten) for perception, and AlphaGo Zero and AlphaZero in 2017 for exactly Samuel's problem: a value function learned from self-play with no supplied feature list and no domain knowledge beyond the rules. Roughly fifty-eight years late, by a mechanism Samuel identified as necessary and could not supply.

Read that beside the other two entries and the shape repeats exactly. Shannon specifies an evaluation function in 1950 and cannot write a good one. Newell, Shaw and Simon hand-write heuristics in 1956–57 and lose to a complete decision procedure by 1960. Samuel hand-writes 38 parameters in 1959 and says in 1967 that automating them is as far away as ever. In all three the hand-written component is the bottleneck, in all three the authors say so, and in all three the approach is reinstated decades later once the hand-written part is learned instead. bitter-lesson-2019 is proposed and unwritten and is the entry that would state this as a thesis; until it exists, these three entries are the evidence for it and should be cited together.

Wrong, and it is the fork: the road not taken was the road. Samuel's 1959 judgment was that the Neural-Net Approach "should lead to the development of general-purpose learning machines" but that the hardware gap made it impractical, so he took the engineered route because it "is capable of realization at the present time." Grade both halves. The strategic call was right — nothing resembling a general-purpose learning machine was buildable on an IBM 704, and rosenblatt-perceptron-1958's own trajectory over the following decade supports him. The long-run call was also right, and that is the point: he named the winner and then spent fifteen years on the other one, for a stated and correct reason. This is the most useful thing in the paper for a reading adjudicating a bet-on-the-future argument, because it is a case where the forecast and the decision correctly pointed in opposite directions. "The right long-term approach" and "the approach to work on this year" are different questions, and a 1959 paper answers both, differently, and is right twice.

Wrong, and famously: the 1962 game. In 1962 the program won a single exhibition game against Robert Nealey, and the result entered computing folklore as a machine defeating a champion. I could not establish the day or month; the sources I reached give the year only, and a reading should say "in 1962" rather than invent a date. What is actually established, from the Chinook project's account maintained by Jonathan Schaeffer at the University of Alberta: "Although a self-proclaimed master, Nealey's tournament results never justified such a claim"; "At the time of the game, he was not a former Connecticut state champion, although he did win the title in 1966, four years after the game with Samuel's program"; Nealey "didn't play in any of the premier checkers events, such as the U.S. championship, and apparently acquired his reputation by beating local players." Schaeffer's annotation identifies Nealey's 16th move as the loss, "a move that will live in infamy," and notes he had fallen into a well-documented trap.

The rematch is the part that never gets quoted. Per the same source, a six-game return match followed, played by post with Nealey moving from home — "The match started last fall and lasted more than five months" — and it went to Nealey, "one win and five draws." One game won by a program in 1962, five draws and a loss across a longer contest immediately afterward: the second result is the better estimate of the program's strength and it is the one that vanished.

And the strongest correction is Samuel's own, in footnote 2 on the first page of the 1967 paper:

> In a 1965 match with the program, the World Champion, Mr. W. F. Hellman, won > all four games played by mail but was played to a draw in one hurriedly played > cross-board game. Recently Mr. K. D. Hanson, the Pacific Coast Champion, has > beaten current versions of the program on two separate occasions.

Schaeffer's page additionally records that "In 1966, Samuel took his program to the world championship match between Walter Hellman (defending champion, U.S.A.) and Derek Oldbury (England)," where four games were played against each opponent and the program lost all eight "rather easily."

These two accounts do not quite reconcile and a reading should not pretend they do. Samuel dates the Hellman games to 1965 and records four mail losses plus one cross-board draw; Schaeffer dates a Hellman set to 1966, adds four against Oldbury, and reports all eight as losses with no draw. They may be the same Hellman match dated differently, or two separate encounters. What both agree on, including the participant with every incentive to report otherwise, is that the world's best players beat the program without difficulty and that no game against a genuine champion was ever won. That is the load-bearing fact and it is not in dispute.

Claim implied 1962, corrected by the author 1967, still misquoted 2026.

Wrong in a way that cost the field twenty-five years. Schaeffer's account holds that the 1962 win created a general belief that checkers was finished, that researchers left it for chess for over 25 years, and that publications kept repeating the claim. The actual answer arrived on 19 July 2007, when Schaeffer and colleagues published Checkers Is Solved on the website of Science: perfect play by both sides is a draw. The scale of what "solved" required — about 5 × 10²⁰ possible positions, of which roughly 10¹⁴ had to be evaluated, after nearly two decades of work — is the measure of the distance between a single won game in 1962 and the thing the 1962 game was taken to mean. A reading that meets the word "solved" applied to a task in 2026 has a clean case of the same word being used forty-five years early by people who were not lying.

Commonly misused as

Not required for idea. Five misuses are frequent enough, and load-bearing enough in arguments the readings adjudicate, to be worth policing. The first is the most common sentence about machine learning on the internet.

"Machine learning is 'the field of study that gives computers the ability to learn without being explicitly programmed' — Arthur Samuel, 1959." The quote is not in the 1959 paper. I read the paper end to end for this entry and it does not appear, in that form or any near variant. It is not in the 1967 paper either. The two sentences it appears to have been distilled from are both real: the abstract's "a computer can be programmed so that it will learn to play a better game of checkers than can be played by the person who wrote the program," and the introduction's "Programming computers to learn from experience should eventually eliminate the need for much of this detailed programming effort." Neither says what the famous sentence says, and the famous sentence's crisp definitional form — the field of study that… — is doing rhetorical work the originals do not do. I could not trace the phrasing to any primary document, and others who have gone looking report the same. It should be quoted as a later paraphrase or not quoted at all; a reading that needs Samuel's definition of the field should use his actual words, which are more modest and more interesting.

"Samuel coined the term 'machine learning' in 1959." Weaker than it sounds, and the paper is the evidence. Samuel does not introduce the term, define it, or mark it as new. It appears in his second and third sentences as established vocabulary — "would be described as involving the process of learning," then "the importance of machine-learning procedures," then "computers with adequate data-handling ability and with sufficient computational speed to make use of machine-learning techniques, but our knowledge of the basic principles of these techniques is still rudimentary." That is a writer using a phrase his readers already know, not a writer coining one. I could not find a documented earlier printed use, so the coinage claim is not refuted here — but it is not supported by the document usually cited for it. "Popularised" is the defensible verb, and it is the verb proposals.md uses. The safe formulation is that the 1959 title is where the term entered general circulation.

"Samuel's program invented alpha-beta pruning." It did not, and Samuel is the source for that too. The 1959 paper's search economies are ply limits and table lookups, not alpha-beta. Alpha-beta appears in the 1967 paper, and footnote 6 there assigns credit away from himself in both directions at once:

> So named by Prof. John McCarthy. This procedure was extensively investigated by > Prof. McCarthy and his students at M.I.T. but it has never been adequately > described in the literature, being a simple algorithmic procedure and actually > only a special case of the more general "branch and bound" technique which has > been rediscovered many times and which is currently being exploited in integer > programming research.

A participant crediting a naming to a colleague and the underlying method to a different literature entirely. When a reading is adjudicating who invented something, this is the standard to hold the claim to.

"Samuel's program beat a champion" — in any of its forms. Covered above; the short version is that the 1962 opponent was not a champion at the time, the rematch was lost, and the world champion won four games to nil by mail three years later, which is on the first page of Samuel's own next paper. The general correction transfers: when a system is reported to have beaten a named human, ask what that human's actual standing was on the date of the game, and ask what the rematch showed.

"Samuel showed that machines learn like people, or that learning is mechanical." The paper makes no such claim and the mechanism argues against it. What learned was sixteen coefficients over a hand-supplied feature list, under a correction rule that Sutton and Barto show could have converged on a constant function without ever touching the game. Samuel's own description of what the program lacked — no way to generate its own parameters, no correct assignment of credit, no guarantee of stability — is a list of the respects in which it was not doing what a person does. godel-incompleteness-1931 and turing-halting-1936 are untouched by anything here in either direction.

Sources

Primary, read directly:

• A. L. Samuel, "Some Studies in Machine Learning Using the Game of Checkers," IBM Journal of Research and Development, Vol. 3, No. 3, July 1959, pp. 210–229. Read from the scan at http://www.bitsavers.org/pdf/ibm/IBM_Journal_of_Research_and_Development/033/ibmrd0303B.pdf. Every quotation attributed to 1959 above is from this text: the abstract, the Neural-Net Approach passage and its footnote 2, the footnote 3 chronology, the Shannon and Strachey references, the minimax and ply-limit description, rote learning and forgetting, the sense of direction, the Alpha/Beta arrangement and the black marks, the "logical difficulty" paragraph, the delta rule, the instability and credit-assignment passages, the term-replacement scheme, both test series and the immobilised-adoption confound, the rote-versus- generalization comparison, the precautionary note, and the Conclusions. The scan I read runs through p. 228; the tail of the Appendix C parameter definitions continues to p. 229 and I did not reach it. Nothing above depends on it.
• A. L. Samuel, "Some Studies in Machine Learning Using the Game of Checkers. II—Recent Progress," IBM Journal of Research and Development, Vol. 11, No. 6, November 1967, pp. 601–617; received 5 June 1967. Read from http://www.bitsavers.org/pdf/ibm/IBM_Journal_of_Research_and_Development/116/ibmrd1106C.pdf. I read pp. 601–603 and pp. 615–617 in full and did not read the signature-table machinery on pp. 604–614. The abstract, the "most basic limitation" sentence, the Hellman and Hanson footnote, the McCarthy alpha-beta footnote, the correlation figures of 0.26 and 0.48, and the Conclusions are all from the pages I read.
• David Silver et al., A general reinforcement learning algorithm that masters chess, shogi and Go through self-play, DeepMind and University College London; read from the preprint deposited at https://discovery.ucl.ac.uk/id/eprint/10069050/1/alphazero_preprint.pdf. I read the opening and the reference list specifically to check the Samuel citation: the "learn for themselves from first principles (5, 6)" sentence is on p. 1, and references 5 and 6 on p. 6 are Samuel 1967 and Tesauro 1994.

Secondary and corroborating:

• Richard S. Sutton and Andrew G. Barto, Reinforcement Learning: An Introduction, section 11.2, "Samuel's Checkers Player." Read in full from a 2011 PDF print of the first edition's web version, hosted at https://webdocs.cs.ualberta.ca/~jonathan/publications/ai_publications/samuel.pdf. The TD-Gammon comparison, the missing-rewards diagnosis, the constant-function argument and the "worse with experience" observation are from this text. This is the first edition; the corresponding section of the second edition (2018) is numbered 16.2 and I did not read it, so anyone quoting section numbers should check which edition they mean.
• The Chinook project, "Arthur Samuel's Legacy," University of Alberta (https://webdocs.cs.ualberta.ca/~chinook/project/legacy.html), for Nealey's standing, the 1962 game's 16th move, the six-game postal return match, the 1966 matches against Hellman and Oldbury, and the claim that the field abandoned checkers for 25 years. Read through two summarising fetches rather than in full; the sentences quoted above were returned as verbatim extracts and I have not seen them in their original page context. The second fetch was made specifically to separate the 1962 single game from the return match, because the first had run them together. AI Impacts' draughts timeline (https://aiimpacts.org/time-for-ai-to-cross-the-human-range-in-english-draughts/) independently carries the same sentences about Nealey, which is corroboration of the text but not of the underlying research, since both trace to Schaeffer.
• Martin Fierz's page for the game itself (http://www.fierz.ch/samuel.htm) carries the full move list of the 1962 Nealey–Samuel game with Samuel's own annotations, and confirms that White resigned. It gives no dates and no match record, so it corroborates the game and nothing around it.
• Suhas Sreedhar, "Checkers, Solved!", IEEE Spectrum, 19 July 2007, for the publication date, the draw result, the 5 × 10²⁰ position count and the 10¹⁴ positions actually evaluated.

Noted failures, per the discipline:

• Science's page for Schaeffer et al., "Checkers Is Solved," doi:10.1126/science.1144079, returned HTTP 403 on 2026-08-16. The solve figures above are from the contemporaneous IEEE Spectrum report, not from the paper, and should be attributed that way.
• incompleteideas.net, Sutton's own site and the canonical home of the Reinforcement Learning text, returned a self-signed-certificate error on 2026-08-16 on two attempts. The Alberta mirror was used instead.
• I could not establish any origin for "the field of study that gives computers the ability to learn without being explicitly programmed." Absence from the two Samuel papers is a checked fact; the phrase's actual first appearance is not, and finding it would improve this entry.
• I could not find a documented printed use of "machine learning" earlier than 1959, so the coinage claim above is left open rather than refuted. The paper's internal evidence is what it is.
• I could not establish the calendar date of the 1962 Nealey game, only the year, and I could not reconcile Samuel's 1965 dating of the Hellman match with Schaeffer's 1966 dating. Both gaps are marked in the prose where they occur.
• Jonathan Schaeffer, One Jump Ahead, which is the primary secondary source for the whole Nealey episode, was not read. Everything about 1962, 1963 and 1966 here reaches this entry through the Chinook project's web pages.
• I did not read Samuel's 1960 IRE Transactions or 1963 Computers and Thought reprint of the 1959 paper. The reprint is noted in the literature as carrying "minor additions and corrections" plus an extra annotated game; if a quotation from it ever conflicts with one above, the reprint is the later text.

This entry places no needle, no score and no landmark, and deposits nothing in the evidence ledger.