IBM Watson Defeats Jeopardy Champions

IBM’s Watson competed against Ken Jennings (winner of a 74-game consecutive win streak in 2004) and Brad Rutter (highest all-time Jeopardy prize earnings) in a three-game match taped in January 2011 and broadcast February 14–16, 2011. Watson won with $77,147 in combined scores versus Rutter’s $21,600 and Jennings’ $24,000. The $1 million first prize was donated to charity; IBM had spent approximately four years and $900 million on the project. Watson’s hardware comprised 90 IBM POWER7 servers with 2,880 processor cores and 16 TB of RAM, occupying 10 racks. The system had no internet access during the match; all 200 million pages of content (Wikipedia, Britannica, literary works, newspapers, and databases) were stored locally. Watson could analyze 200 million pages per second.

Watson’s DeepQA architecture decomposed each Jeopardy clue through several parallel pipelines. A question analysis stage parsed the clue’s grammar and identified the answer type required (a person, a place, a date, a film title). Hundreds of candidate answers were generated by different retrieval algorithms — keyword search, named entity linking, analogical reasoning, passage retrieval from the local corpus. Each candidate was then scored by hundreds of independent evidence-scoring algorithms examining different aspects: how many sources mentioned the candidate in proximity to the clue’s key terms, whether the candidate’s type matched the expected answer type, whether indirect references (cultural, literary, wordplay) could be resolved. A final machine-learning model combined these evidence scores into a single confidence estimate. Watson buzzed in only when confidence exceeded a threshold tuned through competition simulations, since buzzing with a wrong answer incurred a penalty.

Watson made one famous error: in Final Jeopardy with the category “U.S. CITIES,” the clue referenced a city’s airport, deep-dish pizza association, and a railway station. Watson answered “Toronto??” — a Canadian city — with low confidence (reflected by the double question mark it displayed). Post-match analysis indicated Watson’s city disambiguation algorithm had insufficient geographic constraints for the “U.S. CITIES” category label. Jennings wrote on his answer card: “I, for one, welcome our new computer overlords.” IBM subsequently applied Watson technology to healthcare (Watson Health, Watson for Oncology partnership with MD Anderson Cancer Center 2013) and enterprise applications, though both programs were eventually scaled back by the early 2020s as the approach of hand-crafted evidence scoring proved difficult to adapt to the breadth of real-world clinical and business questions.