The boundary between what computers can do and what only humans can do has moved many times across the past 80 years, but it has rarely moved in a way that anyone could see happening in real time. Most domains of human cognition have crossed over to machines gradually, in increments that took years to register: pattern recognition, language translation, image classification, scientific data analysis. The crossing rarely came with a specific date, a specific room, a specific witness who saw the moment of transition. The exception, and arguably still the cleanest single example in the history of artificial intelligence, was 11 May 1997. The thing that crossed over was chess, the specific cognitive domain that human civilisation had treated for centuries as the canonical example of pure intellectual skill — the activity that, if any activity could, demonstrated the depth and creativity of human thought. The machine that crossed over the boundary was an IBM-designed parallel computer with custom chess processors built by a team led by the Taiwanese-American computer scientist Feng-hsiung Hsu. The human being on the losing side of the match was Garry Kasparov, then 34 years old, ranked No. 1 in the world, considered by general consensus among grandmasters to be the strongest chess player who had ever lived.
According to a History.com account of the final game of the Deep Blue versus Kasparov rematch and its broader historical context, the rematch was the second of two six-game matches between Kasparov and successive iterations of the same IBM machine. The first match, held in Philadelphia in February 1996, had been won by Kasparov 4-2, with Kasparov losing the first game (the first ever loss of a reigning world champion to a computer under classical time controls) but recovering to win the match decisively. IBM had spent the subsequent 14 months substantially upgrading Deep Blue’s hardware and chess-specific knowledge — doubling its evaluation speed to approximately 200 million chess positions per second, expanding the opening book curated by Grandmaster Joel Benjamin, refining the position-evaluation function used to assess the relative strength of different board configurations. The 1997 rematch began on 3 May. Kasparov won the first game in 45 moves. Deep Blue won the second game in 45 moves. Games three, four, and five were drawn. By the morning of 11 May, the match was tied at 2.5 to 2.5, with one game remaining. Kasparov needed only a draw to win the match. Deep Blue needed a win to win the match. Kasparov was widely expected to either draw or win the deciding game.
The 19 moves
What actually happened on the afternoon of 11 May was, by the standards of high-level chess, very strange. As reported by an EBSCO Research Starters analysis of the Deep Blue match and its significance for the history of computing, Kasparov elected to play the Caro-Kann Defence — a solid but defensive opening that was not part of his standard repertoire and that he had rarely employed in serious competition. The strategic reasoning was that the Caro-Kann would lead to positions outside Deep Blue’s heavily-prepared opening book, forcing the machine to evaluate positions independently rather than relying on database knowledge. On the seventh move, Kasparov played a moderately unusual variation (…h6) which several commentators subsequently described as either a deliberate anti-computer experiment or a small but consequential mistake — Kasparov himself has never publicly clarified which. The position that resulted offered Deep Blue the opportunity to play an aggressive knight sacrifice on the e6 square that, in human play of the era, would typically be considered too speculative for a machine to identify on its own.
Deep Blue identified it. The knight sacrifice on move 8 had, in fact, been deliberately programmed into Deep Blue’s opening book by Joel Benjamin after analysis of similar positions in master-level games. Kasparov, who later said he had not expected the machine to play the sacrifice, found himself in a position that was, by the cold assessment of subsequent computer analysis, objectively losing within approximately five moves. The position deteriorated rapidly across moves 10 through 19. Deep Blue applied pressure systematically. Kasparov failed to find any meaningful counterplay. At move 19, with his king position exposed, his pieces uncoordinated, and the machine’s threats accumulating, Kasparov resigned. The game had lasted approximately one hour. The match was over. Deep Blue, by 3.5 points to 2.5, had won.
What it meant and what it didn’t
The substantive question of what the match actually proved has been the subject of approximately 28 years of subsequent commentary, second-guessing, and revisionist analysis. As detailed in IBM’s own retrospective on Deep Blue and the technical architecture behind the 1997 victory, Deep Blue’s win was not, in the strict sense, a victory for artificial intelligence as the term is now used. Deep Blue contained no neural networks, no machine learning, no capacity for transfer to any other task. It used 32 processors to perform a set of coordinated, high-speed computations in parallel, evaluating approximately 200 million chess positions per second at a processing speed of 11.38 billion floating-point operations per second — but it could not play any other game, recognise any image, parse any sentence, or do anything other than evaluate chess positions through extremely fast brute-force search. The cognitive achievement involved was, in computer-science terms, narrow rather than general. What Deep Blue had demonstrated was that, in the specific domain of chess, the combination of approximately 200 million evaluated positions per second with a moderately sophisticated handcrafted evaluation function was sufficient to defeat the strongest human player who had ever lived. The implication was significant but bounded. Other domains of human cognition — Go, natural language, image recognition, protein folding, generalised problem-solving — would each take additional decades of independent research to fall, and several of them are arguably still falling rather than fallen.
Kasparov’s own subsequent assessment, expressed in his 2017 book Deep Thinking and discussed in a Chess.com analysis of the 2024 dramatised miniseries about the match and its psychological dimension, has gradually softened from the immediate post-match accusations of human interference (he initially alleged that an IBM grandmaster had been adjusting Deep Blue’s play in real time, an allegation IBM denied and that has never been substantiated) to a more reflective acknowledgment that the loss represented a real technological milestone. “My loss to Deep Blue,” Kasparov wrote in 2017, “was also a victory for humans, its creators and everyone who benefits from our technological leaps.” Deep Blue itself never played another game. IBM declined Kasparov’s request for a third match, retired the machine, and reassigned the development team to other projects. The 480 custom chess chips that had constituted Deep Blue’s specialised hardware were dismantled. Half of the physical machine now sits in the Smithsonian’s National Museum of American History in Washington. The other half is in the Computer History Museum in Mountain View, California. Both halves are silent. Modern chess engines, running on ordinary personal computers without specialised hardware, now play at approximately 3,500 ELO — roughly 650 rating points above any human player who has ever lived, a gap so large that the strongest human grandmasters today have no realistic prospect of ever winning a single game against current top engines under classical time controls. The crossing-over that happened in a Manhattan office tower on the afternoon of 11 May 1997 was, in retrospect, only the first of many. It was, however, the cleanest.
