Garry Kasparov played two matches against IBM's Deep Blue. He won the first, in Philadelphia in February 1996, by 4–2. He lost the rematch, in New York in May 1997, by 2.5–3.5 — the first time a reigning world champion lost a classical match to a machine under tournament conditions.
The single most historic game, though, came a year earlier. Deep Blue won Game 1 of the 1996 match, and that was the first time a computer had beaten a reigning world champion in a classical game at tournament time controls. Kasparov then worked out how to beat it and won three of the remaining five games.
This page covers both matches, the anti-computer method Kasparov used to win in 1996 and why it stopped working, the 1997 controversy in the terms it was actually argued, and what genuinely changed in computer chess afterwards.
The two matches at a glance
| | Philadelphia 1996 | New York 1997 | | --- | --- | --- | | Dates | February 1996 | May 3–11, 1997 | | Venue | Philadelphia Convention Center | Equitable Center, New York City | | Final score | Kasparov 4 — Deep Blue 2 | Deep Blue 3.5 — Kasparov 2.5 | | Machine wins | Game 1 | Games 2 and 6 | | Kasparov wins | Games 2, 5 and 6 | Game 1 | | Draws | Games 3 and 4 | Games 3, 4 and 5 |
Before Deep Blue: the machines that led up to it
Deep Blue did not appear from nowhere. It grew out of a line of purpose-built chess hardware developed at Carnegie Mellon University — ChipTest, then Deep Thought — by a team including Feng-hsiung Hsu and Murray Campbell, who later continued the project at IBM. Deep Thought was already strong enough to beat grandmasters in the late 1980s.
In 1989 Kasparov played a two-game exhibition against Deep Thought in New York and won both games without much difficulty. That result set the tone for the next several years: machines were dangerous tactically, but a world champion could still steer the game onto ground where tactics did not decide anything.
The IBM machine that arrived in Philadelphia in 1996 was a different animal — a parallel system with custom chess chips, searching on the order of a hundred million positions per second. And on the very first day, it won.
Philadelphia 1996: the game that made history
Deep Blue took Game 1. It is the game that belongs in every history of computing: a reigning world champion, at classical time control, losing to a machine over the board. Whatever happened afterwards, that record stands.
What happened afterwards is that Kasparov solved the problem in real time. Over the remaining five games he won three and drew two, taking the match 4–2. He did it by changing what kind of game he was willing to play.
The anti-computer playbook
The machines of the mid-1990s were superb at one thing: enumerating forcing sequences. Give a 1996-era engine a position with hanging pieces, open lines and checks and it would out-calculate anyone alive. Give it a position where nothing is hanging, no line is open, and the correct plan takes twenty moves to bear fruit, and it had almost nothing to hold on to.
So Kasparov played for locked positions. The classic anti-computer structure is a blocked centre, where the pawns interlock and neither side can force contact:
There are no captures available here and no checks in sight. A search-based engine looking ten plies ahead sees dozens of quiet moves that all evaluate to roughly the same number, so it shuffles. A human knows immediately that White wants to play on the queenside with c4–c5 and Black wants ...f7–f5 on the kingside, and starts arranging pieces for that fight. See King's Indian Defense for how both plans actually run.
The other half of the playbook was leaving the opening book early. Deep Blue's book was prepared by grandmasters, so playing a main line meant walking into preparation. A flank opening sidesteps it within two or three moves:
Positions like this are also, by design, positions where Kasparov's own opening preparation was worth less than usual. That trade-off mattered enormously in 1997. If you want the ideas behind this setup, see the Réti Opening.
The horizon problem
The technical name for the machine's weakness in these positions is the horizon effect. A fixed-depth search evaluates the position at the end of its search, not the end of the game. If something bad is going to happen at depth 20 and the engine only sees to depth 12, it can spend moves postponing the bad news beyond its own horizon and score the result as fine. The mirror image is worse: a plan whose payoff arrives at move 40 is invisible, so the engine never starts it.
This is why anti-computer chess worked, and why it is genuinely instructive even now. It is the same reason prophylaxis and long-term pawn structure judgment beat move-by-move calculation in slow positions between humans too.
New York 1997: what IBM changed
IBM went back to work. The 1997 machine — the team nicknamed it "Deeper Blue" — was roughly twice as fast, a parallel IBM RS/6000 SP system carrying hundreds of custom chess chips and searching on the order of 200 million positions per second.
Speed was not the important change. The important change was chess knowledge. Grandmaster Joel Benjamin worked with the team, and other grandmasters were consulted, both to build the opening book and to find and fix the specific positional misjudgements Kasparov had exploited in Philadelphia. The evaluation function was rewritten with many hand-tuned terms — king safety, pawn structure, piece activity — so that the machine's assessment of a quiet position was less naive than a year before.
The result was a machine that still could not plan, but could no longer be counted on to drift.
The 1997 rematch, game by game
Game 1: Kasparov wins
Kasparov took the first game and looked comfortable doing it. The match seemed to be following the 1996 script: lose or draw early, learn the machine's habits, then grind it down.
One curiosity from this game outlived it. Late in a lost position Deep Blue produced a move that made little apparent sense, and Kasparov took it as evidence of depth he could not fathom. Members of the IBM team said afterwards that the move came from a fallback in the software — when the search failed to return in time, the program picked a move more or less at random. A bug, in other words, that Kasparov read as genius. Whether that reading affected his play in Game 2 is unknowable, but he has said the machine's apparent inscrutability got under his skin.
Game 2: the move that broke the match
Game 2 is the one people argue about. Deep Blue had White and steered into a Closed Ruy Lopez — precisely the slow, blocked, manoeuvring structure Kasparov had been counting on the machine to misplay:
The machine handled it well. The moment usually singled out comes around White's 36th and 37th moves, where Deep Blue declined an available pawn and instead played quietly to shut down Black's counterplay. An engine of that era was expected to be materialistic; passing up material to keep a positional bind looked, to Kasparov, like a human hand on the board.
Then came the sting in the tail. Kasparov resigned Game 2 in a position that later analysis showed was drawable — Black had a perpetual check available. He resigned a drawn game to a computer and did not find out until after the match had ended.
Games 3, 4 and 5: three draws under pressure
Rattled, Kasparov leaned harder on the anti-book approach, opening with offbeat first moves to get the machine thinking for itself as early as possible. It half worked. Deep Blue was out of book quickly, but Kasparov was also playing positions he knew less well than his own repertoire, against an opponent that never tired and never doubted itself. All three games were drawn, and the match went to a decider at 2.5–2.5.
Game 6: nineteen moves
Playing Black in the last game, Kasparov chose the Caro-Kann Defense and then walked into one of the best-known traps in the whole opening.
1. e4 c6 2. d4 d5 3. Nc3 dxe4 4. Nxe4 Nd7 5. Ng5 Ngf6 6. Bd3 e6 7. N1f3 h6?7...h6 is a known mistake, and the refutation was old theory long before 1997. Deep Blue played it instantly:
8. Nxe6! Qe7 9. O-O fxe6 10. Bg6+ Kd8 11. Bf4The piece is not the point; the black king is. After the forced sequence Black has an extra knight and a wrecked position: the king is stuck on d8, the rooks are not connected, the pawns on e6 and c6 are weak, and White develops with threats every move.
Kasparov resigned a few moves later, after nineteen in total. Deep Blue won the match 3.5–2.5.
The controversy, stated fairly
The accusation was never that Deep Blue was a fraud. It was narrower: that during Game 2, a human — most plausibly one of the grandmasters attached to the project — had influenced a move choice, which the match rules did not permit during play.
What is documented:
- Kasparov asked to see the machine's logs during the match and was not given them at the time. IBM published logs afterwards.
- IBM declined a further rematch and retired Deep Blue shortly after the event.
- No evidence of intervention has ever been produced, and the moves in question have since been reproduced in spirit by ordinary engines running on ordinary hardware.
The last point is the one that has aged most decisively. A quiet positional move that looked impossibly human in 1997 is unremarkable for a modern engine on a phone. Kasparov himself has publicly stepped back from the cheating claim in later years, framing the loss instead in terms of preparation, the match conditions, and his own state of mind. What he has continued to criticise is the secrecy — a scientific result that cannot be inspected is not much of a scientific result.
What actually changed for computer chess
Deep Blue was a landmark, but its architecture turned out to be a dead end. Three things happened next, and none of them involved custom silicon.
Software overtook hardware. Deep Blue's advantage came partly from purpose-built chips. Within a decade, commercial programs on ordinary personal computers were stronger than Deep Blue had been, because the search itself got much smarter. Better pruning means an engine that examines fewer positions per second can play far better chess. In 2006 Vladimir Kramnik, then world champion, lost a match to the program Deep Fritz running on a normal multi-core PC — and it barely made headlines. The trajectory of that software line is covered in Stockfish engine evolution.
Hand-tuned evaluation gave way to learned evaluation. Deep Blue's positional judgement was written by people, term by term. Two decades later, AlphaZero showed that a network trained by self-play could produce judgements no one had programmed, and neural evaluation is now standard in the strongest engines. Deep Blue's "knowledge" was borrowed from grandmasters; a modern engine's is its own.
Anti-computer chess died. Every weakness Kasparov aimed at in 1996 has been closed. Closed positions no longer confuse engines, long plans no longer sit beyond the horizon, and dragging an engine out of book just gets you a worse position sooner. The idea survives only as a piece of chess history, not as a practical method.
What did not happen is the part people predicted loudest: chess was not solved and did not die. Engines became the standard tool for study, opening preparation moved onto laptops, and human play has measurably improved because everyone now trains against a stronger sparring partner than any human coach. The story of that shift runs through the whole modern era — see the world championship history for the human side of it.
Frequently asked questions
Who won Kasparov vs Deep Blue? Both, once each. Kasparov won the 1996 match in Philadelphia 4–2. Deep Blue won the 1997 rematch in New York 3.5–2.5.
What year did a computer first beat a world champion?
- Deep Blue won Game 1 of the Philadelphia match, the first classical-time-control win by a computer over a reigning world champion. Kasparov still won that match overall.
How many games did Deep Blue win in 1997? Two — Games 2 and 6. Kasparov won Game 1, and Games 3, 4 and 5 were drawn.
Did IBM cheat? No evidence of cheating has ever been produced. Kasparov suspected human intervention in Game 2 and objected to IBM withholding logs during the match; IBM denied intervention, declined a rematch, and retired the machine. The moves that looked suspicious in 1997 are routine for engines today, and Kasparov has since backed away from the accusation.
What was Kasparov's anti-computer strategy? Lock the centre, avoid tactics, leave the opening book early, and play for plans that take longer to unfold than the machine's search depth. It won him the 1996 match and stopped working almost immediately afterwards.
Was Deep Blue stronger than a modern engine? No — not close. A free engine running on a phone today is far stronger than Deep Blue, and it gets there by searching far more selectively rather than by brute speed.
Try the positions yourself
Every board above opens in the free engine on LocalChess with one click. The two worth playing first are the King's Indian blocked centre — try to find a plan when there is nothing to capture — and the Caro-Kann position after 7...h6, where you get to play 8.Nxe6 yourself and see how quickly the attack runs.
If you want a slower, more instructive test than a modern engine at full strength, read how the LocalChess engine works and then set up a game at a level that makes you think. For more historic games to replay, see famous chess games and Fischer vs Spassky 1972.