Internal documents showed Watson's cancer advice was unsafe

The conversation

On 25 July 2018, STAT reported, citing internal IBM slide decks from June and July 2017, that IBM's Watson for Oncology “often spit out erroneous cancer treatment advice.” The decks, presented by IBM Watson Health's deputy chief health officer inside the division's management, are described by STAT as containing assessments from the product's own customers and IBM's medical specialists calling recommendations “often inaccurate.” IBM had been marketing Watson for Oncology, a tool meant to suggest treatment options to physicians, to hospitals worldwide. Responding to the report, IBM told Healthcare Dive that it had “learned and improved Watson Health based on continuous feedback from clients, new scientific evidence and new cancers and treatment alternatives,” and said the product had “trained oncologists to treat 13 cancers and supported care for more than 84,000 patients at 230 hospitals around the world,” a reach IBM stated as its own figure.

What the documents show

STAT's account of the internal decks states two causes IBM's own reviewers identified: the software was trained on a small number of “synthetic,” or hypothetical, cancer cases rather than real patient records, and its recommendations reflected “the expertise of a few specialists for each cancer type” rather than “guidelines or evidence.” The decks concluded the inaccurate recommendations raised “serious questions about the process for building content and the underlying technology,” STAT reports. STAT does not name the clinicians, hospitals or specific erroneous recommendations; the finding is attributed to the documents STAT viewed and to IBM's own customers and specialists, not an independent audit STAT conducted itself.

The system boundary

Watson for Oncology was designed as a decision-support tool: built to generate a ranked list of treatment options for a physician to weigh, not prescribe treatment on its own authority. The documented failure sits inside that boundary — the concern STAT reports is that the suggestions a physician saw could themselves be wrong, not that the software bypassed physician review. IBM's statement to Healthcare Dive does not dispute that the tool sits inside physician decision-making; it instead cites “11 software releases for even better functionality during the past year, including national guidelines for cancers ranging from colon to liver cancer,” as evidence of ongoing correction, not a static, unsupervised system.

Where it fails

The documented failure is a training-data problem: content built from a handful of specialists' judgment, tested against synthetic rather than real cases, can generalize poorly once deployed against real patients. A builder fielding a clinical tool should treat trained by experts as a claim to interrogate, not a guarantee.

• Was the system trained and validated on real patient outcomes, or on hypothetical or synthetic cases?
• Whose clinical judgment does the content encode, and how does it map onto guidelines at the deploying hospital?
• What channel exists for a clinician to flag a wrong or unsafe suggestion back to the vendor?

STAT's reporting and IBM's response describe two sides of one record: internal evidence of a content problem, and a vendor's public emphasis on scale and continuing revision. Neither substitutes for an independent clinical evaluation of the product's accuracy.