Zum Inhalt springen

The Move That Wasn't Intelligence: Deep Blue's Accidental Genius

Zusammenfassung

In Game 1 of the 1997 rematch between IBM’s Deep Blue and Garry Kasparov, Deep Blue made a move (its 44th) that Kasparov could not explain and attributed to “superior intelligence.” It was the result of a software bug: when the program failed to select a move from its search, it fell back to a fail-safe that produced an essentially arbitrary move. Kasparov won Game 1 — but, by his own account and the later analysis of statistician Nate Silver, the inexplicable move unsettled him so badly that he never recovered his composure, played too cautiously, and went on to lose the match. The world’s best chess player read a failure mode as a sign of genius.

The 1997 Match

Deep Blue was IBM’s chess-playing supercomputer: a custom RS/6000 SP parallel computer with 480 dedicated VLSI chess chips, capable of evaluating 200 million positions per second. It played Garry Kasparov — widely considered the greatest chess player in history — in a six-game match in New York in May 1997. Deep Blue won the match 3.5–2.5, making it the first computer to defeat a reigning world champion in a classical chess match under standard time controls.

The match was a sequel to a 1996 match in which Kasparov had won 4–2. IBM had significantly improved Deep Blue between the matches, and the 1997 version was a more capable player. The IBM project and the history of computer chess before it connect to the broader story of AI development covered in The Rise of Artificial Intelligence.

Game 1 and the Anomalous Move

Game 1 was played on May 3, 1997, and was won by Kasparov in 45 moves. The bug occurred in this game, at Deep Blue’s move 44.

The move did not fit the standard patterns of Deep Blue’s play. To Kasparov it appeared not as a blunder but as a quiet, purposeful choice — the kind that seems to require deep positional understanding rather than the brute-force calculation of forced sequences. He attributed it to “superior intelligence” and became convinced that Deep Blue was playing in a qualitatively different way than previous chess computers.

Kasparov won Game 1 anyway, but the move rattled him. He lost the pivotal Game 2 — resigning in a position later shown to be a draw — and afterward accused IBM of human intervention. The two strong moves that defined Game 2 (36. axb5 and 37. Be4) were genuine engine choices, not the bug; but in Kasparov’s mind they confirmed the impression formed in Game 1. The psychological impact carried through the rest of the match: he never recovered his best form.

The Bug

IBM researcher Murray Campbell later explained that the move 44 anomaly was a fail-safe: when the program could not complete the selection of a move from its search, it fell back to playing an essentially random legal move rather than stalling. The move Kasparov had read as evidence of strategic depth was the output of that fallback, not of any plan. Campbell’s account of the bug was reported by statistician Nate Silver in The Signal and the Noise (2012) — years after the match.

Kasparov and his team requested access to Deep Blue’s move logs after the match to verify that the computer had not been assisted by human operators between games — a question of match integrity centered on Game 2. IBM declined the request. The refusal intensified Kasparov’s conviction that something unusual had occurred. The benign explanation — a bug producing an unexpected but harmless move in Game 1 — was not publicly confirmed until long after.

What the Incident Reveals

The Deep Blue bug incident is frequently cited in discussions of AI interpretability and the danger of anthropomorphizing machine behavior. Kasparov, the most qualified human observer possible for chess, interpreted a program error as evidence of machine intelligence. The move looked intelligent because it was unexpected, not because it was the product of an intelligent process.

The pattern — humans attributing understanding to AI systems that are exhibiting sophisticated-looking but mechanically simple behavior — predates Deep Blue and continues with modern language models. The difference in 1997 was that the stakes were a world championship and the observer was uniquely well-positioned to evaluate the behavior. Even under those conditions, the attribution was wrong.

📚 Sources