IBM trained Watson for Oncology on hypothetical patients, then sold it worldwide

IBM built Watson for Oncology to be Memorial Sloan Kettering in a box

In 2012, IBM and Memorial Sloan Kettering Cancer Center announced Watson for Oncology: software that would weigh a cancer patient's chart against the medical literature and rank treatments for the doctor. Dr. Mark Kris, chosen to lead its training, called the plan "this comprehensive, evidence-based approach."¹ A pilot on lung, breast, and prostate cancer was set for later that year, with a wider release due in 2013.¹

By 2017, hospitals in Thailand, Mongolia, China, South Korea, Taiwan, and India had all licensed Watson for Oncology, and Jupiter Medical Center in Florida built television ads around it.² None of that had waited for outside proof: at the time, there was "no published research supporting its impact on patient care."²

The expansion had outrun the foundation under it. Internal IBM presentations from the summer of 2017 said Watson's medical judgment rested on hypothetical cases invented by a couple dozen doctors at one hospital, not on how real patients elsewhere had fared.³ Company specialists and paying customers had already found "multiple examples of unsafe and incorrect treatment recommendations," STAT News reported a year later, after reviewing slide decks IBM's own deputy chief health officer had shown to the division's management.⁴

Watson learned cancer care from two dozen doctors' hypothetical patients

A system like Watson for Oncology learns from cases, each paired with the treatment a human expert chose. What it learns is only as good as how well those cases represent the patients it will later meet. Memorial Sloan Kettering and IBM started, in 2014, by training the system on real patient records.⁵ They soon switched to hypothetical patients, cases the doctors invented themselves rather than pulled from real charts, because standards of care were changing faster than historical records could keep up with.⁵

Dr. Andrew Seidman, the Memorial Sloan Kettering oncologist who trained the system, did not treat this as a flaw to hide: "We are not at all hesitant about inserting our bias... it's a very unapologetic bias."⁶ IBM's own account did not always agree. Deborah DiSanzo, then general manager of IBM Watson Health, told the same reporters that "the bias is taken out by the sheer amount of data we have," a claim IBM later said meant Watson generally, not Oncology specifically.⁶ One defended the bias as intentional. The other denied it was there.

By July 2017, an internal presentation put a number on how thin the foundation was. Across eight cancer types, the training set ranged from 635 hypothetical cases for lung cancer down to 106 for ovarian cancer.³ Memorial Sloan Kettering doctors and IBM engineers compiled the cases with no statistical check on whether the mix represented how cancer actually presents.³

A couple dozen specialists at one hospital cannot describe how cancer is treated everywhere, and the gap shows in how often outside oncologists agreed with Watson's advice. Four measurements below, none directly comparable since they cover different cancers and panels, still slide the same direction: down, the farther a study sits from Memorial Sloan Kettering's own patients and specialists.⁷

Watson was sold to hospitals years before anyone tested whether it worked

STAT found hospitals trading on Watson's name recognition regardless of clinical value: Manipal Hospital in Bangalore, India, charged patients about $150 for a "Watson opinion" consult.² Pilar Ossorio, a bioethicist at the University of Wisconsin, called the sequence backwards: "you should have to prove that there's safety and efficacy before you can just go do this."⁶

Internal presentations show IBM's own deputy chief health officer, Dr. Andrew Norden, weighing that risk after the product had already been sold. One slide said running "a rigorous study now is a very high-risk endeavor that could be embarrassing at best and have serious adverse business consequences at worst."³ Another noted the concordance studies IBM had published so far were "designed in a way that makes negative findings unlikely."³ The Manipal study shows why: its 73-to-93-percent jump came only after the tumor board reviewed the disagreements a second time.⁹ Almost every published concordance study was funded by a paying customer, co-authored by an IBM employee, or both.⁶

The clearest documented case: Watson recommended a chemotherapy combination including bevacizumab for a hypothetical 65-year-old man with severe bleeding, a drug carrying a black-box warning against use in bleeding patients.³ Dr. Eric Topol of the Scripps Translational Science Institute called it worrying: "Any time you have an algorithm that makes a recommendation that's dangerous, that's extremely worrisome."⁴ Memorial Sloan Kettering said the case never reached a patient. It was "part of IBM's system testing," the hospital told STAT, and "no technology can replace a doctor."⁵

MD Anderson Cancer Center in Texas spent roughly $62 million on its Watson project, the Oncology Expert Advisor. A University of Texas System audit found the spending had skirted procurement rules: almost $40 million went to IBM and over $21 million to the consultancy PwC, largely without board approval.¹⁰ The system was never piloted outside MD Anderson; IBM ended its support in September 2016, and MD Anderson's president resigned soon after.¹⁰¹¹

In January 2022, IBM sold the Watson Health data-and-analytics business that Watson for Oncology had been part of to the investment firm Francisco Partners.¹² IBM did not disclose a price; reporting at the time put it near $1 billion, a fraction of the more than $3.6 billion IBM had spent building the business.¹³

IBM says it was testing. The record says it was already for sale.

IBM and its critics do not disagree about what happened. They disagree about what a company owes a hospital before selling it a medical opinion.

Both can be true at once: a bad recommendation may never have reached a patient, and the system may still have been marketed as more validated than the record shows. What would resolve the dispute is not another concordance study: those measure only whether Watson agrees with a human panel, not whether its advice helped a patient live longer or better. It would take an independent, prospective trial: Watson-guided care against standard care at multiple hospitals outside Memorial Sloan Kettering, tracking survival and cost rather than agreement. No such trial was ever run.⁶⁷⁸

The same gap is open in the next clinical AI pilot

Nothing about this failure required Watson to be especially bad software. Any system trained on one hospital's patterns and judged against that hospital's own panel will look strong at home and weaker elsewhere. The independent meta-analysis found exactly that pattern beyond IBM's own case: concordance was consistently lower for advanced, metastatic disease than for earlier stages.⁷ The authors traced part of the international spread to population differences in drug tolerance that Memorial Sloan Kettering's training cases never captured.⁷

IBM's own portfolio shows the alternative. Watson for Genomics matched a tumor's genetic profile against curated literature rather than issuing a treatment ranking, and it is the one Watson Health product reporters have since called a success.⁸ The fix was narrowing what the system decided, and proving that narrower claim before selling it.

Clinical decision-support and diagnostic models are still built the way Watson for Oncology was: trained on one institution's patients, checked against that institution's own panel. Many still reach patients before an independent trial shows whether the advice helps anyone. A system built this narrowly cannot announce its own limits, so the hospital buying it must ask what the seller did not answer first: proven where, on whom, and against what.