Regulating every form of AI as though it were the same technology does more harm than good. Should some AI be spared from the current backlash?
Anthropic CEO Dario Amodei recently called for a slowdown of AI, citing a host of security risks. He’s the latest skeptic to join an expanding group of self-appointed regulators who see doom in the future of AI.
Generalizations about AI often fail to distinguish between general-use LLMs and valuable industry-specific AI tools. Healthcare, for instance, uses many AI tools, but even the catch-all term “medical AI” can describe many different technologies with different risk profiles.
Holding all medical AI tools back, however, could cause real patient harm.
Catastrophizing frontier medical AI could make patients distrust useful technology. The fear and confusion sparked by pessimistic regulators could scare patients, clinicians, hospitals, and investors away from adopting tools that could deeply improve healthcare outcomes.
Yet the political debate continues to shrink all AI into one “AI” category.
As one ER doctor explained, “The algorithm that helps a radiologist flag a pulmonary nodule? We call that AI. The ambient tool that drafts a note while I talk to a patient, so I am not charting at midnight? AI. The system that flags a sepsis trajectory earlier than a busy resident? AI. The frontier models now raising alarms about autonomous cyberattacks and engineered pathogens? Also AI.”
As of June 2026, there are 1,614 FDA-authorized medical devices that use AI. Scan.com uses AI to search referrals for imaging centers with available appointments. Aletta was authorized to use AI in robotic phlebotomy. Queen of Hearts, an ECG model designed to recognize acute heart attacks that conventional criteria miss, snagged De Novo authorization this month.
On September 8, Google’s AlphaGenome announced its free-to-use academic research AI tool that shows how changes in DNA may affect gene regulation. To date, it’s the most comprehensive log of how genetic mutations interact with molecular biology. It’s not yet possible to test all 9 billion possible mutations in the human genome in the lab without predictive AI.
On September 10, OpenAI published a case study showing frontier general-purpose AI already being used in antibiotic discovery. Anthropic itself, just days after declaring an AI slowdown, set up biology labs to ramp up an AI drug discovery program and announced a partnership with OpenEvidence to address physician shortages in 100 countries.
None of these things are a reason to slow medical AI. What’s more, an NBER paper shows that broader AI use could lead to roughly $200-360 billion every year in healthcare savings.
Until now, most medical AI has handled paperwork and administrative tasks where the regulatory burden is lowest: handling billing, writing notes, alongside FDA-cleared tools that help doctors read scans. HHS is allowing for Medicare pilots that don't need FDA pre-market approval, which has drawn criticism.
But AI could change patient outcomes for the better: a quarter of hospitalized Medicare patients already experienced harm during a single month, and reviewers judged 43 percent of it preventable.
At what point would recommendations that hold back AI cause harm to patients if they delay patient diagnosis, treatment, understanding of their condition, or a full picture of their health future?
Patients and doctors do like and use AI, despite the recent news backlash. A March AMA survey found that 76 percent of physicians think AI offers advantages in patient care — which is up from 65 percent a few years ago.
We should be free to choose AI tools for information gathering, guidance, better understanding, or second opinions. And yes, in the future, for medical assessments — clinical decision support software that could be used alongside clinicians if desired.
An AI slowdown could cause patient harm in the long term: insights missed, diagnoses delayed, and tools kept from those who want them.
That’s not to say privacy and security concerns aren’t real. They certainly are.
Patients deserve to know when AI shapes a diagnosis or treatment, what it does, what data it uses, and who answers for it, and we should be able to refuse it if we want. But overregulating AI in medicine and science could harm patients who already experience medical error without the use of AI, suffering through fragmented or unattainable care.
AI tools are not interchangeable. Treating them that way does all of us a disservice. The answer is simple: innovate more. Don’t slow down AI.