Despite high hopes, AI in health care hasn't been effective because medical data is more complex and scarcer than web data, causing misleading results
Tom Simonite / Wired : Tweets: @carnage4life , @pkedrosky , @carnage4life , @thiago37 , and @reema__patel Tweets: @carnage4life : Lack of large datasets for training medical AIs means it's more likely that they have biases that lead to incorrect diagnoses given insufficient training set. Until AI techniques evolve or privacy rules change, artificially intelligent robot doctors will remain science fiction. Paul Kedrosky / @pkedrosky : Oh, FFS. The belated “gee, health is complicated” naivete from tech hypesters is predictable and tedious. Medical information is more complex and less available than the web data that many AI algorithms were trained on, so results can be misleading https://www.wired.com/... @carnage4life : Machine learning is hitting limits in healthcare because showing the system lots of examples to learn from hits snags due to lack of such collections of medical data. Privacy rules prevent having large corpuses of medical data for A.I. research. https://www.wired.com/... Thiago Julio / @thiago37 : “The community fools [itself] into thinking we're developing models that work much better than they actually do. It furthers the AI hype.” Visar Berisha, associate professor, Arizona State University https://www.wired.com/... Reema Patel / @reema__patel : Nice writeup in @WIRED of a report from @turinginst @Bilal_A_Mateen - we've got a while to go until we have got effective use of data and AI in healthcare... https://www.wired.com/... via @wired
Context & Ripple Effects
This piece closes a loop that Wired itself helped open: coverage of Google's mammogram-reading system warning that diagnostic AI could amplify overtesting and overdiagnosis, and reporting that hospitals were quietly deploying unproven AI decision-support tools most patients never knew about. Tom Simonite's argument supplies the mechanism behind both stories — privacy rules keep training sets small, and small sets inherit the racial and gender biases documented back in Quartz's 2018 analysis of healthcare AI.
First-order effects
- Hospitals running AI diagnosis tools are making clinical decisions on models trained from datasets too small to be representative, which is exactly the condition under which the flawed-tool pattern the WSJ later documented — clinicians feeling pressure to defer to the algorithm — takes hold.
Second-order effects
- Vendors selling into health care can't lean on the web-scale data advantage that powered consumer AI, so competition shifts toward whoever can secure proprietary clinical datasets or survive validation scrutiny; meanwhile privacy rules themselves become the binding constraint on product quality.
Third-order effects
- If the pattern holds, medical AI matures along a different axis than the rest of the industry — progress gated by dataset access, privacy reform, and clinical validation rather than model scale — pushing regulators toward governance regimes built for high-stakes, low-data domains.
The trend: Healthcare AI is being forced off the web-scale playbook, with scarce, complex data and privacy rules — not compute — setting the pace of what reaches patients.