AI systems, like Google's system for reading mammograms, have the potential to worsen pre-existing problems like overtesting, overdiagnosis, and overtreatment
Christie Aschwanden / Wired : Tweets: @robmay , @wired , @nathancortez , @allenfrancesmd , @edwardtufte , @vprasadmdmph , @went1955 , and @cragcrest Tweets: Rob May / @robmay : Yes. AI can be wrong. What will doctor liability be when they go against the advice of the AI? https://www.wired.com/... @wired : Machine-enabled health care may bring us many benefits in the years to come, but those will be contingent on the ways in which it's used. If doctors ask the wrong questions to begin with, the technology could serve to amplify our earlier mistakes. https://www.wired.com/... Nathan Cortez / @nathancortez : There's a ton of hype with A.I. in medicine and skeptics are often called Luddites. But this article (in @WIRED!) does a great job explaining why A.I. isn't a panacea and even creates its own risks. https://www.wired.com/... Allen Frances / @allenfrancesmd : Great piece on the big risk Artificial Intelligence will amplify the mistakes of the bad medicine often practiced today. If you start with unintelligent assumptions, AI will cause even more over-diagnosis/harmful treatment than we already have. All hi-tech glitter is not gold. https://twitter.com/... Edward Tufte / @edwardtufte : A+ report on chronic long-term AI diagnostic issues: garbage in, garbage out cascades of false alarms bad medicine made worse In contrast, NYT gullible account “Using A.I. to Transform Breast Cancer Care” says AI might PREDICT TREAT CURE, and even PREVENT cancer https://twitter.com/... Vinay Prasad / @vprasadmdmph : Both for content and writing, this is simply the perfect essay. https://twitter.com/... Robert Went / @went1955 : AI Makes Bad Medicine Even Worse. A new study out from Google seems to show the promise of AI-assisted health care. Actually, it shows the threat. Mark Zuckerberg's motto, “Move fast and break things” is not great for medicine, AI-assisted or not https://www.wired.com/... Christie Aschwanden / @cragcrest : That new Google study on artificial intelligence for breast cancer screening isn't an example of how AI will revolutionize medicine. It's a cautionary tale of how AI could make bad medicine worse. https://www.wired.com/...
Context & Ripple Effects
Christie Aschwanden's piece lands right after Google published results for its mammogram-reading system, and it pushes back on the framing that better detection is automatically better medicine: an AI that finds more can simply industrialize overtesting, overdiagnosis, and overtreatment if the underlying clinical questions are wrong. The piece also surfaces a question with no settled answer — what a doctor's liability is when they override the algorithm's advice.
The subsequent record bears out the concern rather than resolving it. Hospitals went ahead and deployed decision-support tools patients didn't know about (often novel and unproven), and by 2023 some clinicians reported feeling pressure from administrators to defer to flawed diagnostic algorithms. By 2025, the same worry had spread from radiology suites to patients themselves running AI-recommended diagnoses.
First-order effects
- Doctors adopting tools like Google's mammography system face an unresolved liability gap: as Rob May flagged at the time, the rules for overriding AI advice are undefined, so early adopters bear the risk personally.
- Screening programs gain throughput but inherit the pre-existing overdiagnosis problem at scale — every false-positive cascade the technology finds is one a human program now has to manage.
Second-order effects
- Hospital administrations become the decisive actor: once procurement is done, the pressure documented in US hospitals to defer to the algorithm inverts the intended human-in-the-loop safeguard.
- Vendors of competing diagnostic AI are pushed to compete on validation evidence rather than accuracy claims alone, since unproven-tool deployments became the cautionary tale of the deployment wave.
Third-order effects
- If the pattern holds, medical AI regulation shifts from approving models to governing their use — audit rights, disclosure to patients, and explicit override protocols become the battleground rather than model performance.
- The deeper structural risk is that machine-enabled care amplifies whatever incentives already exist in fee-for-service screening, entrenching overtreatment unless payment and liability structures change alongside the technology.
The trend: Medical AI is moving from lab results to embedded clinical infrastructure faster than the governance around its use, making deployment practices — not model accuracy — the limiting factor on whether it helps or harms.