Watson defeats Ken Jennings and Brad Rutter on Jeopardy!

Filed as moment, and moment is right for once without argument — but the one-line description this entry was commissioned under is wrong, and correcting it is the entry's first job. The house definition, argued out by eliza-1966 when it talked its way out of the kind and used since by clippy-1996, siri-2011, expert-systems-collapse-1987 and deep-blue-1997, is a date on which something visibly happened in public. This qualifies on every clause. Two games were taped on 14 January 2011 on a replica Jeopardy! set built inside IBM's Thomas J. Watson Research Center in Yorktown Heights, New York, and broadcast across three nights — 14, 15 and 16 February 2011 — to the largest audience the show had drawn in six years. Alex Trebek hosted. There was a scoreboard, a referee, two of the strongest human players the format has ever produced, and a machine that won by roughly three to one.

Where the id comes from, and what it was called. canon/proposals.md does not carry watson-jeopardy-2011 among its fifty candidates. It appears in the closing section, "What I left out at the cap," in the list of sixteen the proposer said he expected to be missed, described in five words: "Watson on Jeopardy (2011, the demo that never became deployment)." That places this entry in a third position the canon has not had before — not proposed like deep-blue-1997, not absent like siri-2011, but named and consciously dropped. The admission test in section 2 therefore has to carry the entry, and it does.

But "the demo that never became deployment" is false, and the falsity is the whole point. Watson did become deployment. Watson for Oncology was a commercial product with a price, a sales force and customers; it ran in hundreds of hospitals in more than a dozen countries — Ross and Swetlitz put the figure at 230 hospitals and health organisations reaching about 84,000 patients in six months of 2018 — and IBM assembled a health division around it at a reported cost of about $4 billion in acquisitions. What happened is not that the demo failed to reach the clinic. It is that the demo reached the clinic and was wrong there, in ways that were documented at the time by the people using it, and that the division was sold off in 2022 for roughly a quarter of what its acquisitions cost. That is a materially different lesson from a prototype gathering dust, and it is a strictly more useful one, because the comfortable version — "great demo, no product" — lets a reader conclude the problem is engineering follow-through. It was not. The follow-through happened.

Why not idea. DeepQA is a genuine architecture with a genuine paper, and if this entry were about that architecture it would be idea and its year would be 2010, the AI Magazine issue. It is not, and the reason is that the architecture is a dead line. Watson used no deep neural networks. It was a hand-assembled ensemble of more than a hundred techniques feeding a learned ranker, built in the last window in which that was the state of the art, and by 2013 the field had moved somewhere the design could not follow. Nothing inside a 2026 transformer works like DeepQA. What descends from February 2011 is not a method; it is a public event and a fifteen-year commercial argument that came out of it. That is what a moment leaves behind, and the same test mycin-1976 applied in the other direction — MYCIN's method is dead too, but what descends from MYCIN is an evaluation genre, so MYCIN is an idea.

Why not prediction. Tempting, and there are at least four dated, gradeable claims running through this — the technical target in the 2010 paper, the IBM–Nuance "18 to 24 months," and Ginni Rometty's two revenue numbers. A prediction entry carries one claim, its due date and its outcome. This carries an event with four forecasts hanging off it. The forecasts are graded in section 3 anyway, which moment does not require and which this canon should keep doing.

Why not interpretation. The strongest alternative, and it loses on the same ground the proposal line does. There is a frame here — "the Watson lesson," invoked constantly whenever an AI product disappoints — and the frame is a 2020s artefact rather than a 2011 one. But the frame as it actually circulates is wrong in the specific way section 4 sets out, so filing this as an interpretation would canonise the misreading in the act of recording it. The event is checkable. The frame is not, and the entry should hold the event.

What it is

The lineage, which is self-declared

IBM has run periodic Grand Challenges: publicly legible man-versus-machine demonstrations chosen to be easy to explain, internationally visible, and recruiting-useful. Deep Blue was one. Blue Gene was another. After 1997 the programme had no successor, and Paul Horn, then director of IBM Research, had been pressing his managers for the next one.

In 2004, IBM Research manager Charles Lickel was at dinner with colleagues when the restaurant emptied toward the bar's television: Ken Jennings was mid-way through the 74-game streak that ended on 30 November 2004. Lickel proposed Jeopardy! as the next Grand Challenge. Horn backed it the following year. That is descent in the strict sense eliza-1966 insisted on — came out of — and it is self-declared rather than inferred by this entry: the programme that produced Watson existed because Deep Blue had worked, and the search for a successor is documented as the reason the Jeopardy! idea was entertained at all.

descends_from therefore holds exactly one id, and one plausible ancestor is being deliberately left out. mycin-1976 is the obvious second candidate, because Watson for Oncology is straightforwardly the successor to the idea of a computer advising physicians on treatment, and because MYCIN's blinded-evaluation genre is the ancestor of the concordance studies that later graded Watson. But descent has to be about the artefact, not the theme. The machine that played Jeopardy! did not come out of MYCIN in any traceable sense — it shares no technique, no team, no lineage of citation, and its builders were not answering MYCIN's question. Putting mycin-1976 in the header would assert a genealogy this entry cannot document. It is named here in prose instead, which is what the job prompt asks for, and a reading citing both should treat them as adjacent rather than parent and child. shannon-chess-1950 stands behind deep-blue-1997 and therefore two steps behind this, not one.

The machine

The design problem was not "answer trivia questions." It was to answer them under a contest's constraints: an exact answer to a natural-language clue, in about three seconds, with a calibrated number attached saying how sure it was, because in Jeopardy! a wrong buzz costs you the clue's value. The team's own statement of the target, in the Fall 2010 AI Magazine paper, is that a system would need to "answer roughly 70 percent of the questions asked with greater than 80 percent precision in 3 seconds or less."

They arrived at that target empirically. Analysing nearly 2,000 historical games, they found that Jeopardy! champions "acquire on average between 40 percent and 50 percent of all the questions from their competitors and perform with between 85 percent and 95 percent precision" — the region they called the Winner's Cloud. Jennings at his peak sat at the far edge of it: 62 percent of the questions at 92 percent precision. Everything about Watson's shape follows from having to reach that cloud rather than from wanting to be right on average.

The architecture, DeepQA, is a funnel. A clue is parsed and its expected answer type identified; a search stage generates several hundred candidate answers; each candidate is then scored independently by more than fifty scoring components — matching on type, on temporal consistency, on geospatial plausibility, on source reliability, on passage support — and the scores are merged by a ranker trained on prior games, which produces both an answer and a confidence. The paper describes more than 100 different techniques across the pipeline. Nothing in it is a lookup table and nothing in it is a single model; the whole design is an argument that no one method would clear the cloud and that a hundred weak ones, weighted, might.

It was ruinously expensive to run. An early implementation "took 2 hours to answer a single question" on one processor. Scaled across about 2,500 compute cores with UIMA-AS it reached "competitive run-time latencies in the 3–5 second range." The competition machine was ninety IBM Power 750 servers, 2,880 POWER7 cores at 3.5 GHz, 16 terabytes of RAM. Its corpus was 200 million pages of structured and unstructured content, about four terabytes, including the full 2011 English Wikipedia — loaded in advance, because Watson was not connected to the Internet during the match. The often-quoted hardware cost of about $3 million traces to an estimate by IBM's Tony Pearson rather than to a disclosed figure, and should be treated as an estimate. [verify]

Watson could not see or hear. Clues were delivered to it as a single line of text, in the show's uppercase, at the instant they were revealed to the humans; it buzzed with a mechanical thumb driven by a solenoid, taking five to ten milliseconds from decision to buzz.

The match

Watson practised. IBM ran about 100 sparring games against former Jeopardy! contestants during 2010, winning roughly 65 percent. On 13 January 2011, at a press event on the replica set, it played an exhibition round against Jennings and Rutter and won it: Watson $4,400, Jennings $3,400, Rutter $1,200. The two real games were taped the next day.

Game 1, broadcast across 14 and 15 February, finished Watson $35,734, Rutter $10,400, Jennings $4,800. Game 2, broadcast 16 February, finished Watson $41,413, Jennings $19,200, Rutter $11,200. The two-game aggregate, which is the number usually quoted, is Watson $77,147, Jennings $24,000, Rutter $21,600. One caveat on sourcing: at least one widely-read reference labels the two games with the wrong broadcast dates, attaching game 1's total to 14 February and game 2's to 15 February. The games are two; the broadcasts are three; game 1 was split across the first two nights.

The purse was $1 million to Watson, $300,000 to Jennings, $200,000 to Rutter. IBM gave the whole million away, split evenly between World Vision and the World Community Grid; Jennings and Rutter each gave half of theirs to charities of their own choosing.

Ratings: the three nights averaged a 9.1 to 9.2 household rating in Nielsen's metered markets against about 7.0 for the show a year earlier, its best run since Jennings and Rutter last met in May 2005, with double-digit ratings in 24 of 56 metered markets. Total-viewer figures in circulation vary by source and measurement window and this entry does not quote one.

The two failures, which are more informative than the win

Day one, the category "Name the Decade." Jennings buzzed in with "What are the '20s?" and was wrong. Watson then buzzed in and offered "What is 1920s?" Trebek: "No, Ken said that." Watson had no microphone. It could not hear its opponents, did not model them, and had no representation of the fact that an answer had already been spent. It was not confused about the decade; it was missing the other players entirely.

Final Jeopardy, game 1, category "U.S. Cities." The clue: "Its largest airport is named for a World War II hero; its second largest, for a World War II battle." The answer is Chicago — O'Hare and Midway. Watson wrote "What is Toronto???" — with the question marks, which its interface used to display low confidence. Ferrucci's account of the failure is specific and worth having: category titles in Jeopardy! are frequently not constraints ("there's lots of data where the title is U.S. cities and the answers are countries, European cities, people, mayors"); there are towns named Toronto in the United States; the Blue Jays play in the American League.

And then it wagered $947. With confidence around 30 percent, on a $36,681 bank, it bet under a thousand dollars. The blunder and the bet are the same fact seen twice: the system did not know the answer and knew that it did not, and the number it produced was good enough to size a wager on. Nearly every retelling keeps the Toronto joke and drops the wager. The wager is the engineering.

The buzzer, which is a real objection and not a disqualifying one

Jennings has been consistent and generous about this since the week it aired. Writing in Slate on 16 February 2011: "Buzzer skill is crucial — games between humans are more often won by the fastest thumb than the fastest brain." A solenoid firing on a five-millisecond electrical decision is not competing with human reflexes on equal terms, and Watson received its clue as text rather than having to read it off a board.

Fifteen years later he had not changed his mind about the fairness and had acquired a different doubt. To Claire McNear in Slate, 8 September 2025: "Is it unfair that Watson could buzz faster than any human? No, that's just an advantage we know that computers have." And, on whether IBM had held something back in the untelevised games: "To this day, I still don't know — is that because Watson was hanging back, because they hadn't put it in its turbo mode?" That is Jennings' uncertainty, reported as his, and this entry does not treat the existence of a "turbo mode" as established.

Why a reading would cite it

The admission case is a single sentence: this is the best-adjudicated artificial-intelligence demonstration ever staged, and it still did not predict the product.

Take the demonstration's evidentiary quality seriously for a moment, because almost nobody does. The match was live-refereed by a third party whose business is the format. The opponents were the two strongest players in the game's history, and were not chosen by IBM. The clues were written by the show, not the vendor, and were not seen in advance. There was a public scoreboard and a published rulebook. Failure was visible in real time — and it was visible, on camera, twice. Compare that to any 2026 capability claim: a vendor-run benchmark on a vendor-chosen suite, reported by the vendor, on a model the evaluator cannot obtain. On the axis of independent verification, February 2011 beats essentially everything since.

And it was also a real capability. This is not a story about a rigged demo. Watson met a published, pre-registered performance target under contest conditions, and the target had been derived from measurements of human champions before the machine existed. The result replicates: a 2023 study in the same journal that published the DeepQA paper (O'Leary, AI Magazine) put ChatGPT and Bard on the same questions and found both more accurate than Watson overall, and all three comparable on the questions Watson had answered with high confidence. Watson's 2011 performance is a number that has survived being checked by technology that did not exist when it was set.

So the entry is not for grading demos. It is for grading the inference from a demo to a deployment, and it is the strongest counterexample available, because it removes every easy explanation. The demo was not faked, not cherry-picked, not vendor-scored, and not unreproducible. The product still failed in the field. Whatever went wrong lives in the gap between "this system answers Jeopardy! clues at champion precision" and "this system should advise an oncologist," and that gap is not made of engineering rigour. It is made of the fact that the two tasks share a surface — questions in, answers out — and share nothing else. Jeopardy! clues have exactly one right answer, known in advance, decidable in three seconds, with no consequence to being wrong beyond the score. None of those four properties holds in a clinic.

This is where the entry differs from its two nearest neighbours, and a reading should pick deliberately between the three. mycin-1976 is the case where performance was demonstrated and adoption never came: the system won its blinded evaluation and was never used on a patient. deep-blue-1997 is the case where a company won its demonstration, declined the rematch, dismantled the machine and walked away without ever selling anything. This entry is the third and worst case: the company won the demonstration, believed its own result, built a business on it, sold it to hospitals on five continents, and the product was documented as unsafe by its own internal reviewers while it was still being sold. Cite MYCIN for the gap that never closes. Cite Deep Blue for the capability that was real and had nowhere to go. Cite Watson when a demonstration is being offered as evidence that a deployment will work.

Three concrete 2026 occasions, from this project's own window.

One: revenue figures supplied by the company whose story they support. In the window covered by digests/2026-08-15-12.md, OpenAI's CFO gave investors a $40 billion annualised run rate with enterprise overtaking consumer, and Anthropic reported preliminary Q2 revenue over $11.5 billion — both recorded in that digest as company-supplied and unaudited. Watson is the base case for what happens next, and section 3 grades it: IBM published a $1 billion-by-2018 Watson revenue target and a $10 billion-a-year-within-ten-years target, never once reported Watson as a separate revenue line, and by January 2026 was reporting a "generative AI book of business" instead — a metric IBM itself defines as "inception to date Software transactional revenue, plus new SaaS Annual Contract Value and Consulting signings." That is not annual revenue and is not comparable to the number it replaced. When a company changes the unit, this entry is what says what happened last time.

Two: capability claims that cannot be independently evaluated. The same digest records Anthropic disclosing "Model 2," an unreleased internal model described as better than the model it ships, with no plans to release it, and records the reading's own caveat that every capability number in the window except Qwen's weights was vendor-reported. Watson cuts against the reflex here, which is why it is worth citing rather than piling on: its demonstration was as independently evaluable as such a thing gets, and independent evaluation of the demonstration did not license the product claim. A reading that treats "was it independently checked?" as the whole question is asking a necessary question and mistaking it for a sufficient one.

Three: medical AI arriving through procurement rather than proof. That digest records nine California AI bills clearing Senate Appropriations on 13–14 August 2026, including AB 1979 and AB 2575 on healthcare AI. Watson for Oncology is the worked example of what those bills are aimed at: a system that entered hospitals through purchasing and marketing, whose clinical evidence base was concordance against one institution's practice rather than outcomes, and whose problems surfaced through investigative journalism and leaked internal decks rather than through any regulatory instrument.

What it got right, and what it got wrong

moment does not require this section. Four dated claims attach to this event and all four are now due, so it gets one anyway.

Claim 1 — the technical target. Made Fall 2010. Due February 2011. Met.

The claim, in its own words (Ferrucci et al., AI Magazine 31(3), Fall 2010): a competitive system must "answer roughly 70 percent of the questions asked with greater than 80 percent precision in 3 seconds or less."

What happened: met, in public, under adjudication, four months later. This is the cleanest hit in the canon's prediction record so far, and its shape is worth noticing: it was made by engineers about their own system, it was narrow, it was quantified, it named its measurement conditions in advance, and its due date was four months out. Every property that makes it gradeable also makes it modest.

Claim 2 — the ship date. Made 17 February 2011. Due ~February 2013. Met, and it changed nothing.

The claim: IBM and Nuance announced a healthcare research agreement on 17 February 2011 — the day after the final broadcast — with Columbia University Medical Center and the University of Maryland School of Medicine. The release states: "The two companies expect the first commercial offerings from the collaboration to be available in 18-24 months." Dr Herbert Chase of Columbia: "Watson has the potential to help doctors reduce the time needed to evaluate and determine the correct diagnosis for a patient." Dr Eliot Siegel of Maryland: "We believe that this has the potential to usher in a new era of computer assisted personalized medicine into healthcare to improve diagnostic accuracy, efficiency, and patient safety."

What happened: on 8 February 2013 — 23.7 months later, inside the window by about a week — IBM and WellPoint announced Watson's first commercial application: utilization-management decision support for lung cancer treatment, with Memorial Sloan Kettering. The ship-date promise was kept almost exactly, and this is the most instructive grade in the entry. The failure that follows is not a schedule failure. A reader who grades AI forecasting only on dates will score this a hit and learn nothing. What shipped was a narrow decision-support tool at one institution and an insurer, not the diagnostic partner the release described, and the gap between those two things never closed.

Claim 3 — the revenue. Made October 2013. Due 2018 and 2023. Not met, and made unmeasurable.

The claim: on an October 2013 earnings call, IBM CEO Ginni Rometty said Watson would generate $10 billion in annual revenue within ten years (due 2023) and $1 billion of revenue a year by 2018. In January 2014 IBM announced a Watson Group with more than $1 billion of investment, including a $100 million venture fund.

What happened: IBM never reported Watson as a separate revenue line, so the 2018 target cannot be graded against a published number — which is itself the result. What can be graded is the asset. IBM spent about $4 billion acquiring four health-data companies for Watson Health; on 21 January 2022 it agreed to sell the healthcare data and analytics assets to Francisco Partners, reportedly for something over $1 billion; the deal closed 30 June 2022 and the business was relaunched as Merative, headquartered in Ann Arbor. The Watson name did not go with it. By the 2023 due date for the $10 billion, IBM had attached "Watson" to a different product line entirely — watsonx, announced May 2023 — and by IBM's fourth-quarter results released 28 January 2026 the company was reporting a generative AI book of business "of more than $12.5 billion" inception-to-date, against full-year 2025 revenue of $67.5 billion. Reading that release, I could not find the word Watson or watsonx in it at all. [verify — one release, read once]

Self-grade versus independent grade, which the canon requires where both exist: IBM has not published a retrospective grading of the Rometty targets, so there is no self-grade to record. The nearest thing is the change of metric, which grades them silently and in IBM's favour: $12.5 billion inception-to-date across software revenue, SaaS contract value and consulting signings is a larger number than $10 billion in annual revenue and a much smaller claim.

Claim 4 — the clinical claim. Made continuously 2011–2018. Graded by users. Failed.

This is the one the entry exists for, and the record is unusually good because the people who used it wrote down what they found.

• MD Anderson. The Oncology Expert Advisor consumed $62 million — about $39 million to IBM and $23 million to PwC — across a contract extended twelve times without competitive rebidding. A University of Texas System audit, released February 2017, found procurement failures, cost overruns and an $11.6 million deficit in gift funds spent before receipt, and noted the system did not work with the hospital's new Epic installation. The contract was allowed to lapse. It was never used on a patient.
• STAT, 5 September 2017 (Casey Ross and Ike Swetlitz): the product was "in their toddler stage" per one user; "the system is essentially Memorial Sloan Kettering in a portable box," trained by "a couple dozen physicians" at a single institution, whose Dr Andrew Seidman is quoted saying "we are not at all hesitant about inserting our bias." Watson searched published literature but derived its recommendations from that training rather than from what it read.
• STAT, 25 July 2018: internal IBM slide decks from June and July 2017 recorded "multiple examples of unsafe and incorrect treatment recommendations," and attributed the problem to training on a small number of synthetic cases — hypothetical patients — rather than real ones. Worth stating precisely, because this is routinely told worse than it was: the documented unsafe recommendations were identified in testing and by customers, not traced in the reporting to a harmed patient.
• The concordance evidence, which is the interesting part. The studies hospitals ran measured agreement with a local tumour board rather than patient outcomes, and the results diverged by geography in a way that tells you what was being measured. At Manipal in Bangalore, a 1,000-patient study led the board to change treatment in 13.6% of cases. ASCO-reported concordance figures ran as high as 96%. At Rigshospitalet in Copenhagen, local oncologists agreed with Watson in about 33% of cases and dropped the project; their director attributed the gap to Watson weighting American studies over European ones. A system trained on one hospital's practice scores well where practice resembles that hospital's and badly where it does not — which means concordance was measuring institutional similarity, not correctness. Nothing in this record establishes that Watson for Oncology was ever right or wrong in the sense that matters, because the outcome study was not run. That absence, and not the Danish number, is the finding.

What it got right that is consistently forgotten

Watson was built to know when it did not know, and to act on that number. It answered a clue only when calibrated confidence cleared a threshold; it sized a Final Jeopardy wager on that confidence; it displayed the number on screen; and the whole architecture existed because precision-at-a-chosen-coverage, not accuracy-if-forced-to-answer, was the metric that mattered. Confidence was a first-class output, and abstention was a legitimate move.

That property was thrown away. A 2026 assistant answers effectively every question at full rhetorical volume, and the field has spent several years re-deriving calibration, abstention and refusal as remedies for a problem its own design created. The canon should record that a system in 2011 got this right, structurally, by construction, because it was made to play a game where being confidently wrong cost money. This is my reading and not a claim I can source to anyone; it is offered as an argument, and it is the reason the $947 wager belongs in the retelling and the Toronto joke does not.

What it got wrong technically

The architecture had no future. Watson used no deep neural networks, at the last moment when a hand-engineered ensemble could win a headline result — AlexNet was eighteen months away. Its hundred techniques were each somebody's careful work, and within a few years a single trained model with none of that specificity was better at the task. O'Leary's 2023 comparison is the epitaph: general-purpose language models, with no Jeopardy!-specific engineering whatsoever, outscored the machine that had been built for five years to do exactly this.

And Ferrucci said so at the time, internally. His account — reported by Steve Lohr in The New York Times on 16 July 2021 and repeated widely since — is that he warned colleagues Watson had been engineered to find word patterns and predict answers for a quiz show and was not "an all-purpose answer box ready to take on the commercial world," and that the warning went nowhere: "It wasn't the marketing message." He has put the technical claim more bluntly since: "While Watson outperformed the best, it did not understand what it read." Ferrucci left IBM in 2012.

Commonly misused as

moment does not require this section. godel-incompleteness-1931 established that the canon writes one wherever a result is repeatedly made to say something it does not, and four such uses are live.

1. "Watson proved machines can understand language." It proved an ensemble could generate several hundred candidate answers and rank them by accumulated evidence, fast enough and precisely enough to beat champions at a game whose clues have exactly one right answer. Its designer says it did not understand what it read, and the two on-air failures are evidence for him rather than against: a system that understood the situation would have known Jennings had just said "the '20s," and a system that understood the category would have known that "U.S. Cities" is not satisfied by Toronto. Watson's own confidence score said so at the time.

2. "Watson shows AI doesn't work in medicine." This is the entry's Gödel problem — a specific documented failure conscripted into a general impossibility claim. What the record supports is narrow and mechanical: a system trained on synthetic cases reflecting one hospital's practice, sold ahead of its evidence, into procurement processes that (at MD Anderson, on the auditors' findings) were not functioning, and evaluated by concordance with local practice rather than by outcomes. Every one of those is a contingent institutional fact with an identifiable alternative. None of them is a limit on what machine learning can do in a clinic, and the canon has a limit kind precisely so that things which are not limits do not get filed as if they were.

3. "The buzzer means Watson didn't really win." The buzzer advantage is real, Jennings named it in print the week it aired, and it plausibly accounts for some margin. It does not account for the result. Watson had to have the right answer to gain from buzzing first, and it had it at champion precision across two games and a 3:1 aggregate. Jennings, who has more standing on this than anyone, does not make the argument. The related "turbo mode" suspicion — that IBM ran the machine differently for the cameras — is Jennings' stated uncertainty and is not established fact; a reading may report it as his doubt and must not report it as a finding.

4. "The demo never became a product." The version this entry was commissioned under, and the most consequential of the four, because it is the most comforting. It licenses the reader to think the missing ingredient was commercial execution — that a more disciplined company would have converted the result. IBM converted the result. It shipped, on time against its own announced window, to hundreds of hospitals across more than a dozen countries, backed by $4 billion of acquisitions and the most famous demonstration in the industry's history, and the conversion is what failed. The gap that mattered was between the task the demonstration measured and the task the product had to do, and no amount of execution crosses it.

Sources

Primary.

• D. A. Ferrucci, E. Brown, J. Chu-Carroll, J. Fan, D. Gondek, A. A. Kalyanpur, A. Lally, J. W. Murdock, E. Nyberg, J. Prager, N. Schlaefer and C. Welty, "Building Watson: An Overview of the DeepQA Project," AI Magazine 31(3), Fall 2010, pp. 59–79, DOI 10.1609/aimag.v31i3.2303. The full PDF exceeded the fetch tool's size limit; the quoted passages — the 70%/80%/3-second target, the Winner's Cloud figures from nearly 2,000 games, Jennings at 62%/92%, "several hundred candidate answers," "more than 100 different techniques," the fifty-plus scorers, the 2-hours-on-one-CPU figure and the 2,500-core scaling — are taken from AAAI's own article summary page, "The AI Behind Watson — The Technical Article," which reproduces them verbatim from the paper. Checked against a second summary for consistency; [verify] against the PDF if a reading leans hard on any single number.
• IBM and Nuance Communications, "IBM to Collaborate with Nuance to Apply IBM's 'Watson' Analytics Technology to Healthcare," press release via PR Newswire, 17 February 2011, 09:00 ET. Source of the "18-24 months" sentence and the Chase and Siegel quotations, read directly.
• IBM, "IBM Releases Fourth-Quarter Results," newsroom release, 28 January 2026. Source of full-year 2025 revenue ($67.5 billion), Q4 2025 revenue ($19.7 billion), the "more than $12.5 billion" generative AI book of business and IBM's own definition of that metric, read directly. The observation that the release does not mention Watson or watsonx rests on a single fetch of a single release and is flagged as such above.
• Ken Jennings, "My Puny Human Brain," Slate, 16 February 2011. Source of the buzzer-skill line and "Brad and I were the first knowledge-industry workers put out of work by the new generation of 'thinking' machines," read directly. Note that his famous Final Jeopardy annotation — "I for one welcome our new computer overlords" — was written on the answer card during the 16 February broadcast and does not appear in this piece.

Secondary, and used as such.

• Casey Ross and Ike Swetlitz, "IBM pitched its Watson supercomputer as a revolution in cancer care. It's nowhere close," STAT, 5 September 2017, read via the Scientific American reprint. Source of "Memorial Sloan Kettering in a portable box," the Seidman quotation, the adoption figures and the Danish 33% concordance and Dr Leif Jensen's explanation of it.
• Casey Ross and Ike Swetlitz, "IBM's Watson recommended 'unsafe and incorrect' cancer treatments," STAT, 25 July 2018. Source of the internal June/July 2017 slide decks, the "multiple examples of unsafe and incorrect treatment recommendations" quotation and the synthetic-cases finding. Read via search summaries and corroborating coverage (Boston Globe, 29 July 2018; Healthcare Dive), not the STAT original, which is paywalled. [verify].
• Eliza Strickland, "How IBM Watson Overpromised and Underdelivered on AI Health Care" (print title "IBM Watson, heal thyself"), IEEE Spectrum 56, 2 April 2019, pp. 24–31. Source of the ~$4 billion figure for four health-data acquisitions, the count of roughly fifty announced partnerships of which most produced no commercial product, and the MD Anderson and Watson for Genomics assessments.
• The Cancer Letter, 17 February 2017, and Medscape, "Big Data Bust: MD Anderson–Watson Project Dies," for the audit findings: $62 million total, $39 million to IBM and $23 million to PwC, twelve contract extensions without rebidding, the $11.6 million gift-fund deficit, and the Epic incompatibility.
• Steve Lohr, "What Ever Happened to IBM's Watson?," The New York Times, 16 July 2021. Source of the "all-purpose answer box" and "It wasn't the marketing message" passages. Not read at first hand — nytimes.com is not fetchable from this machine — and taken from O'Dwyer's PR summary of 21 July 2021, which quotes it. [verify].
• Claire McNear, "The Jeopardy! game that changed everything still haunts the show 15 years later," Slate, 8 September 2025. Source of the three 2025 Jennings quotations, including the "turbo mode" doubt.
• Daniel E. O'Leary, "An analysis of Watson vs. BARD vs. ChatGPT: The Jeopardy! Challenge," AI Magazine, 2023, DOI 10.1002/aaai.12118. The Wiley full text returned HTTP 402; the findings above — both models more accurate than Watson overall, no significant difference between them, comparable accuracy on Watson's high-confidence questions — come from a secondary summary (AZoAI, 5 September 2023), which explicitly declines to give per-system accuracy figures. This entry therefore quotes no accuracy percentages from it. [verify].
• Wikipedia, "IBM Watson," read for its citations rather than its prose. Source of the game-by-game scores, the 13 January 2011 practice result, the ~100 sparring games at 65%, the prize distribution and charities, the hardware specification, the 200-million-page / four-terabyte corpus and the no-Internet-during-play fact, the Pearson cost estimate, the statement that Watson "did not involve the use of deep neural networks," and the WellPoint / Memorial Sloan Kettering commercialisation timeline. Its date labelling of the two games is wrong in the way noted above.
• Variety and Broadcasting+Cable, 17–18 February 2011, for the Nielsen metered-market ratings.
• Fast Company, "Why Watson Wagered $947," 17 February 2011, and The Register, 21 February 2011, for the Final Jeopardy confidence figure and Ferrucci's explanation of the Toronto answer.
• Francisco Partners, "Francisco Partners Completes Acquisition of IBM's Healthcare Data and Analytics Assets; Launches Merative," 30 June 2022, and Fierce Healthcare and Healthcare IT News coverage of 21 January 2022 and July 2022, for the sale, the reported price and the Merative rebrand.
• TechRepublic's Watson origin history and the Holloway Making Things Think chapter for the Lickel-and-Horn Grand Challenge account of 2004–2005. The restaurant detail is consistently reported and thinly sourced; the entry uses the parts that are consistent across accounts. [verify].
• Jo Cavallo, "Confronting the Criticisms Facing Watson for Oncology," The ASCO Post, 10 September 2019, for the Manipal 1,000-patient study and the 13.6% figure. Note that this piece is broadly sympathetic to the product and contains no methodological criticism of the concordance design; the criticism of that design in section 3 is this entry's own and is argued, not cited.

Attempted and failed, so nothing above silently depends on it: the AI Magazine DeepQA PDF at Yale exceeded the fetch size limit and could not be read whole; ibm.com/history/watson-jeopardy returned HTTP 403; nytimes.com is not fetchable; the Wiley full text of O'Leary 2023 returned HTTP 402; SEC EDGAR returned 403 on the IBM 8-K exhibits and the newsroom release was used instead; cnbc.com returned 403 for IBM's Q2 2026 results, so the most recent book-of- business figure this entry carries is the one from 28 January 2026 rather than from July. Circulating figures that this entry deliberately does not use: a widely repeated "Watson answered 38 of 44 correctly, 86%" on-air accuracy and a matching "GPT-4, 59 of 61, 97%," which appear in blog posts without a traceable methodology; and total-viewership numbers for the three broadcasts, which vary by source.

The 2026 citation occasions — OpenAI's $40 billion annualised run rate and Anthropic's preliminary $11.5 billion quarter, both recorded as company-supplied and unaudited; the "Model 2" disclosure of 15 August 2026; the reading's own "claimed vs. demonstrated" caveat; and the California healthcare-AI bills AB 1979 and AB 2575 clearing Senate Appropriations on 13–14 August 2026 — are as recorded in this project's own digests/2026-08-15-12.md, which holds the primary links. They are named here as occasions to cite this entry, not as evidence for anything in it. Nothing in this file is evidence, nothing in it is deposited in the ledger, and nothing in it touches the needle.