AI in health: progress and potholes
Concrete regulatory wins—like an AI tool cleared for gestational dating—sit alongside big caution signs about validation and privacy, suggesting AI in healthcare is maturing unevenly. Critics note that many AI health claims still lack hard outcomes or approved drug discoveries, and California lawsuits over offsite processing of recorded doctor visits show privacy risks that can erode patient trust. (diagnosticimaging.com) (thenextweb.com) (arstechnica.com)
Most medical artificial intelligence is not a robot doctor. It is pattern-matching software that looks at an image, a waveform, or a transcript the way spellcheck scans a sentence for likely mistakes. (diagnosticimaging.com) That narrow kind of tool is where the clearest progress is showing up. On March 30, 2026, Butterfly Network said the United States Food and Drug Administration cleared its gestational age tool, which estimates how far along a pregnancy is from ultrasound sweeps in less than two minutes. (diagnosticimaging.com) Gestational age is the pregnancy clock doctors use to time scans, steroids, and delivery decisions. Butterfly says its tool works between 16 and 37 weeks and was trained on more than 21 million images, with results comparable to sonographer biometry measurements. (diagnosticimaging.com) The trick is that this system does not need a specialist to freeze the perfect frame and measure tiny structures by hand. Butterfly calls it a “blind sweep” workflow: a user follows a guided three-step scan, and the software turns that moving video into an age estimate. (diagnosticimaging.com) That is the version of medical artificial intelligence that regulators can actually evaluate: one task, one output, one clinical setting. It is a very different promise from the broad claim that artificial intelligence will soon discover blockbuster drugs or replace clinical judgment across a hospital. (thenextweb.com) The gap between those two stories is huge right now. The Next Web reported on April 11, 2026 that no drug discovered by artificial intelligence has yet won approval, even though companies can now screen millions of molecules a day with machine-learning systems. (thenextweb.com) Generative artificial intelligence chatbots are even messier, because they produce language that sounds finished even when the evidence is thin. The same report says 40 million people use ChatGPT for health advice daily, while critics argue many headline-grabbing health claims still lack hard outcome data in real patients. (thenextweb.com) Then there is the privacy problem, which is less about accuracy than about who gets to hear the room. Ars Technica reported on April 11, 2026 that several Californians sued Sutter Health and MemorialCare over allegations that an artificial intelligence transcription tool recorded doctor visits without patient consent. (arstechnica.com) Those cases focus on ambient scribes, which are note-taking systems that listen during appointments and turn conversation into draft charts. If a pregnancy scan tool is like a calculator for one measurement, an ambient scribe is more like a microphone placed in one of the most private conversations people have. (arstechnica.com) So the field is not moving in one clean line. In April 2026, the strongest evidence sits with tightly scoped tools that do one medical job under regulatory review, while the biggest trust risks sit with systems that touch sensitive conversations or make sweeping claims before outcomes are proven. (diagnosticimaging.com) (thenextweb.com) (arstechnica.com)
