When Were the AI Winters, and When Did Deep Blue Beat Kasparov?

Garry Kasparov <strong>won</strong> the first Deep Blue match in Philadelphia, 10-17 February 1996, by 4-2, then lost the rematch in New York City on 3-11 May 1997 by 3.5-2.5 - the first time a reigning world champion lost a match to a computer under standard tournament time controls (40 moves in two hours). Kasparov held the PCA classical world title at the time; Anatoly Karpov held the FIDE title. The two so-called AI winters are conventionally dated to roughly <strong>1974-1980</strong> (following the 1973 Lighthill Report and parallel US funding cuts) and roughly <strong>1987-1993</strong> (following the collapse of the Lisp-machine market). Neither winter has a hard start or end date: the match dates are exact and documented, the winter dates are historians' conventions.

The two Deep Blue matches, in order

There were two matches, and conflating them is the single most common error. The first, billed as the ACM Chess Challenge, ran in Philadelphia from 10 to 17 February 1996. Deep Blue won game one on 10 February - the first time a computer beat a reigning world champion in a single game at classical time control - but Kasparov adjusted and took the match 4-2.

IBM rebuilt the machine and the rematch ran in New York City from 3 to 11 May 1997. Deep Blue won 3.5-2.5, taking the decisive sixth game on 11 May, in which Kasparov resigned after 19 moves. Kasparov publicly questioned whether human intervention had occurred during game two; IBM denied it, declined a further rematch, and dismantled the machine. No evidence of cheating has ever been produced, and the result stands.

Note what the result was not. It was not a general-intelligence milestone. Deep Blue was purpose-built chess hardware with a hand-tuned evaluation function, searching enormous numbers of positions per second. It learned nothing between games except through its human handlers.

The first AI winter: the Lighthill Report and the mid-1970s cuts

The field's first funding collapse is usually dated c. 1974-1980. The trigger most often named is the report Sir James Lighthill prepared for the UK Science Research Council, published in 1973, which argued that AI's results had not matched its promises and that its methods did not scale - the combinatorial-explosion objection. UK academic AI funding was sharply narrowed in its wake.

The US decline was parallel rather than caused by Lighthill. The 1966 ALPAC report had already gutted machine-translation funding, and DARPA cut back open-ended AI research in the mid-1970s, including declining to renew its Speech Understanding Research programme, which ran from 1971 and concluded in 1976. Minsky and Papert's Perceptrons (1969) is often blamed for stalling neural-network research in the same period; historians dispute how much of that is retrospective myth-making.

The phrase AI winter is not from this period. It was coined in discussion at the American Association for Artificial Intelligence meeting in 1984, by researchers warning that the then-booming expert-systems market would end the same way.

The second AI winter: the Lisp machine collapse

That 1984 warning proved accurate. Expert systems built on specialised Lisp hardware were commercially hot in the early 1980s; then general-purpose workstations from Sun and Apollo, and later commodity PCs, became fast enough to run Lisp adequately. The specialised-hardware market collapsed in 1987, and firms including Symbolics and Lisp Machines Inc. failed or shrank drastically.

Government money followed. DARPA's Strategic Computing Initiative was cut back sharply in the late 1980s, and Japan's Fifth Generation Computer Systems project - launched in 1982 with roughly a decade's mandate - wound up in 1992, with a follow-on programme running to 1994, without delivering the machines it had promised. The second winter is usually dated c. 1987 to the early-to-mid 1990s; some accounts extend it to about 2000. Practitioners in that period often avoided the phrase artificial intelligence entirely, preferring informatics, machine learning, or knowledge-based systems.

Colossus, vacuum tubes, and the November 2022 bookend

Before either winter, and before the term artificial intelligence existed, came the electronic hardware. Colossus, designed by Post Office engineer Tommy Flowers to help break the German Lorenz cipher, was built from vacuum tubes (thermionic valves) rather than relays - roughly 1,600 in the Mark 1 and around 2,400 in the Mark 2. Colossus Mark 1 first worked at the Post Office Research Station, Dollis Hill, on 8 December 1943, was delivered to Bletchley Park in January 1944, and was first used against real enemy traffic on 5 February 1944; the faster Mark 2 ran from 1 June 1944, days before D-Day. Colossus was programmed by switches and plugboards, not a stored program, and it was not an AI machine - but it proved that thousands of valves could be made to run reliably, which every post-war electronic computer depended on.

The field's naming event came later: the summer 1956 Dartmouth workshop, whose 1955 funding proposal coined the phrase artificial intelligence. The modern public-facing bookend is 30 November 2022, when OpenAI released ChatGPT as a free research preview - the event that ended any remaining talk of winter and began the current funding boom.

What most scholars accept, and what stays contested

• Accepted: The match dates and results are documented and uncontested: Philadelphia 10-17 February 1996, Kasparov 4-2; New York 3-11 May 1997, Deep Blue 3.5-2.5, decided in game six on 11 May. Colossus Mark 1 first worked at Dollis Hill on 8 December 1943, was delivered to Bletchley Park in January 1944, and was first used on real traffic on 5 February 1944, using roughly 1,600 vacuum tubes, with Mark 2 from 1 June 1944. The Dartmouth workshop was held in summer 1956. The Lighthill Report was published in 1973. ChatGPT was released on 30 November 2022.
• Mainstream: Historians conventionally date the first AI winter to about 1974-1980 and the second to about 1987-1993, with the second sometimes extended toward 2000. The standard causal account credits the Lighthill Report and DARPA's mid-1970s retrenchment for the first, and the 1987 collapse of the Lisp-machine market plus the 1992 wind-up of Japan's Fifth Generation project for the second. The term AI winter is generally credited to discussion at the 1984 AAAI meeting.
• Contested: The winter boundary years are conventions, not events - no funding body ever declared a winter, so any specific start or end year is an interpretive choice and sources differ by several years. How much causal weight the Lighthill Report carries is disputed: the US and UK declines had partly independent causes, and Lighthill may be more marker than cause. The claim that Minsky and Papert's Perceptrons (1969) killed neural-network research is contested by historians who note the work continued and that the book's technical claims were narrower than the legend. Kasparov's 1997 allegation of human intervention in game two remains unsubstantiated; IBM denied it and the match result is not in dispute, but the episode is still argued over.

Frequently asked questions

Explore related content on this site

• Deep Blue Defeats Kasparov
• Colossus: The Vacuum Tube Paradigm
• ChatGPT and the AI Tipping Point
• The Turing Machine

Sources and further reading

• Overview of the Deep Blue matches, including the February 1996 Philadelphia match won by Kasparov and the May 1997 rematch won by the computer.
• IBM's own account of Deep Blue, the 1996 and 1997 matches, and the machine's design.
• Technical and historical detail on Colossus: Tommy Flowers's valve design, the December 1943 Mark 1, February 1944 operation at Bletchley Park, and the Mark 2 from 1 June 1944.
• Sir James Lighthill's 1973 report "Artificial Intelligence: A General Survey" for the UK Science Research Council, the document most often cited as triggering the first AI winter.
• OpenAI's announcement of ChatGPT as a free research preview, dated 30 November 2022.
• Standard textbook treatment of the field's funding history, covering both winters, the ALPAC and Lighthill reports, and the expert-systems boom and bust.