Ai stethoscopes are lying to patients—and doctors are furious
The radiology screen glows, the algorithm spits out a verdict in 0.3 seconds, and another hospital marketing team rushes to call it a revolution. Dr. Mieses Malchuk calls it something else: a dangerous farce. The Columbia-trained internist has spent the last decade watching AI vendors sell diagnostic certainty to boards and budgets while frontline clinicians watch the same code miss nuance, context, and the odd melanoma that doesn’t look like the training set.
Her ZDNet op-ed last week detonated the usual applause tracks. Instead of the polite caveats researchers tuck into the final paragraph of their papers, Malchuk went for the jugular: presenting these pattern-matching engines as reliable clinical partners is, in her words, “an elaborate theatrical production” that risks turning bedside medicine into a cargo-cult ritual of data entry and blind trust.
Why pattern-matching is not a differential diagnosis
Malchuk’s core gripe is epistemological. A convolutional neural net can label a chest X-ray with the word “infiltrate,” but it has never watched a patient’s oxygen saturation drift while antibiotics fail, never noticed the faint clubbing that hints at a decades-old asbestos exposure, never weighed the social chaos that will unravel if the patient can’t afford the prescribed follow-up CT. Those fragments live outside the pixel grid; they live in the hallway, in the chart’s blank margins, inside the physician’s own mirror neurons. Strip that context away and you get a statistical mirage dressed up as certainty.
The numbers back her up. A 2023 JAMA meta-analysis found 62 commercially cleared AI imaging tools showed a 19 % drop in sensitivity when deployed outside the tertiary hospitals where they were trained. Translation: the miracle algorithm works—until the ambulance drives ten blocks south. Patients don’t read footnotes; they read headlines promising instant diagnosis on a smartphone. When the hallucinated answer feels authoritative, the flesh-and-blood appointment gets postponed, sometimes forever.
The business model behind the curtain
Microsoft’s Dragon Copilot, OpenAI’s GPT-4-Med, Google’s Med-PaLM—each press release uses the same verb: “augment.” What rarely appears is the subscription price, the lock-in clause, or the indemnity shift that quietly slides liability from vendor to hospital the moment the “accept” button is clicked. Malchuk argues this is not augmentation; it’s a liability laundering scheme with a UI.
Meanwhile, venture funding for “clinical AI” reached $12.6 billion in 2024, a 38 % jump year-on-year. The pitch decks open with images of smiling rural patients whose nearest specialist sits 200 miles away. They close with recurring-revenue charts that look suspiciously like Salesforce slides. No one includes the cost of the malpractice premium after the chatbot confuses a Zika rash with dengue.
The human pivot hospitals still refuse to fund
Malchuk isn’t a Luddite; she runs a small pilot using NLP to flag chart omissions for trauma patients. The difference, she says, is that her program never claims to think. It highlights missing data, then gets out of the clinician’s way. The hospital still had to fight to keep the $180 k annual budget line because “AI project” sounds sexier than “hire two more abstractors.”
Her prescription is blunt: regulate the marketing, not just the algorithm. Require every diagnostic AI to carry a black-box label equivalent to cigarette packaging: “This device cannot smell poverty, fear, or the subtle tremor of early Parkinson’s.” Force vendors to publish real-world performance curves stratified by site, insurer tier, and patient accent. And, above all, stop letting CFOs believe that buying a model is cheaper than paying humans to master the craft of noticing.
The next time a pitch deck promises to “democratize diagnosis,” Malchuk wants the buyer to ask one question: democratize for whom—the patient, the balance sheet, or the Series B round? Until then, she’ll keep teaching residents the lost art of the physical exam, one stethoscope at a time, while the server racks hum overhead, selling certainty by the millisecond.