IBM Sold a Cancer Doctor Trained on Imaginary Patients. The Bill Came to $4 Billion.

In a 2017 internal slide deck, an IBM doctor laid out a problem the company was not saying out loud: Watson for Oncology had produced "multiple examples of unsafe and incorrect treatment recommendations."⁶ The system meant to help oncologists choose cancer therapies was, in some cases, suggesting the wrong ones. And it had a reason. Watson hadn't learned from a stream of real patients walking through a real clinic. It had been trained on a small number of synthetic, hypothetical cancer cases — patients who never existed, diseases no one ever had.⁶ IBM had told the world the opposite.

The official story is that AI wasn't ready for medicine, and IBM was simply early. That gets it backwards. The technology being early is forgivable. What IBM did was industrialize a marketing narrative before it had validated the science underneath it — selling a product trained on imaginary patients as if it ran on real clinical data, then spending billions to acquire the data it could never get the product to use.

That is the whole diagnosis. Watson Health wasn't a bet that lost. It was a story that got funded, scaled, and sold to hospitals before anyone proved the story was true.

What IBM bought, and what it could never wire together

IBM launched Watson Health in April 2015 and went shopping. It acquired Phytel, then Explorys, then Merge Healthcare for $1 billion, then Truven Health Analytics for $2.6 billion — more than $4 billion in health data assets in roughly a year.³⁴ On paper this was a thesis: own the world's clinical and claims data, point a learning machine at it, and out-diagnose the doctors. The logic was sound. The execution assumed that data you own is data your product can use, and in healthcare that assumption is fatal.

Medical data does not sit in a clean lake. It lives inside electronic health record systems that don't speak to each other, in formats that resist machine reading, behind privacy rules that throttle every transfer. Owning the assets did not solve the integration; it just put a price tag on the gap. IBM had spent $4 billion buying the cargo and never built the road to move it.

MD Anderson: $62 million to discover the obvious

The flagship customer was MD Anderson Cancer Center, which built its Oncology Expert Advisor on Watson. It ended in 2016, after a University of Texas audit found the center had spent $62 million on the project.⁵ Read the audit's complaints and you see the gap made physical: the system could not sync with MD Anderson's Epic electronic health record, it ran on outdated data, and it drowned in delays, cost overruns, and procurement problems.⁵ The most advanced medical AI in the world could not read the patient records sitting in the building next door.

Here is the detail that turns a project failure into an indictment. IBM declared internally, in September 2016, that the Oncology Expert Advisor was "not ready for human investigational or clinical use."⁵ That is not a hedge; it is a moratorium. The science was visibly not there. And yet the public posture stayed bullish — the collaboration was something to promote, not to pause. The internal verdict and the external pitch had stopped pointing in the same direction. That is the mechanism of the fall, and it is not a technology problem at all.

Why hype-first AI destroys more value than it builds

Most AI projects fail quietly: the model underperforms, the pilot ends, the team moves on, and little is lost. Watson Health failed loudly and expensively because IBM inverted the normal order of operations. The right sequence is validate, then product, then market. IBM ran it market, then product, then — somewhere down the list — validate. When you lead with the narrative, every downstream cost gets locked in before the science gets a vote. You sign $4 billion in acquisitions to feed a claim. You sell hospitals a clinical tool on the strength of a press release. And when the validation finally arrives and says no, you can't quietly walk it back, because you've already told the world it works.

Wasn't IBM just early — and didn't it sell at a profit?

The fair objection is that pioneers always look foolish in hindsight, and that medical AI is genuinely hard — the EHR mess and the data-integration wall would have hurt anyone. True. But being early doesn't require lying about your training data. The synthetic-patient problem wasn't a limit of the era; it was a choice to ship and promote before the real data was wired in.⁶ Early is a defense for the difficulty. It is no defense for the gap between the internal moratorium and the external pitch.

There's a second, more tempting objection: IBM sold the assets and even booked a gain, so where's the failure? In January 2022 it agreed to sell the healthcare data and analytics assets — Health Insights, MarketScan, Clinical Development, Micromedex, imaging software and more — to Francisco Partners.¹ Its Q2 2022 SEC filing disclosed $1.065 billion in cash and a pre-tax gain of about $230 million.² A gain, technically. But IBM spent more than $4 billion just acquiring the data pieces, before a dollar of the AI development on top.³ Selling the inventory for a fraction of what you paid, and calling the accounting entry a 'gain,' is not a win. The buyer kept the data business and rebranded it Merative; IBM kept the Watson brand and walked away from medicine.⁸ The thing that didn't survive was the original promise.

IBM didn't get beaten by the future. It built a cancer doctor out of patients who never existed, spent $4 billion buying data it couldn't feed to it, told hospitals it was real while telling itself it wasn't ready, and finally sold the wreckage for a quarter of the cost — and recorded it as a gain.²³⁵⁶ The lesson isn't that AI can't do medicine. It's that a story is the easiest thing in the world to scale and the most expensive thing to walk back. Sell the proof, not the promise — because in the end, the bill always finds the gap between the two.