MEC Practice Academy 11.19
Opinion|Articles|November 18, 2026

Why a slowdown on AI could cost patients their lives

Author(s)Mika Newton
Listen
0:00 / 0:00

Key Takeaways

  • Diagnostic error and poor interoperability impose measurable population-level harm; delaying clinically validated AI effectively preserves that baseline of preventable death and disability.
  • Sepsis early-warning systems using continuous vitals/labs surveillance have demonstrated mortality reductions when alerts are rapidly confirmed and acted upon by clinicians.
SHOW MORE

We know artificial intelligence is already helping physicians treat patients. Why wait because of speculation about what might happen?

The loudest voices in the artificial intelligence (AI) debate are asking medicine to wait, and patients would have to pay the price.

In recent weeks, some of the most powerful executives in technology have urged the world to slow AI down. They warned that the systems they build could one day escape human control. The president called the warnings a hoax. Both camps skipped the question that should settle this: What happens to patients while we wait?

I run a company that assembles a patient's medical history from every place they have received care. Every day, I see the information a doctor needs stuck in an electronic health record system at another facility, and that system will not talk to the one in front of the physician who needs it. In 2023, Johns Hopkins researchers estimated that 795,000 Americans die or are permanently disabled by diagnostic error every year. That is the baseline a slowdown would protect. The case for pausing AI rests on harms that might arrive someday, and it asks us to accept harms we can count today.

Sepsis shows the stakes most clearly. It starts out looking like an ordinary illness, and it turns deadly when treatment starts late. Johns Hopkins researchers deployed an AI early warning system across five hospitals, reading vital signs and lab results for 590,736 patients around the clock. Among flagged patients whose alert a clinician confirmed within three hours, sepsis mortality ran 18.7% lower, according to results in Nature Medicine. The software replaced no one. It gave the care team back the hours that decide who survives.

Cancer screening tells the same story. Researchers in Sweden randomly assigned 105,934 women in the country’s mammography screening program to AI-supported reading or standard practice, in which two radiologists review every scan. The AI group had 29% more cancers detected with no increase in false alarms, according to The Lancet Digital Health. Most of the added detections were small, early-stage invasive tumors, the ones doctors most want to find before they spread. Screen-reading workload also fell 44% in a specialty with chronic staffing shortages.

These results came from the kind of evidence medicine demands before it changes practice: a large deployment published in a peer-reviewed journal and a randomized trial. The slowdown argument has no comparable evidence behind it. It offers a scenario. I take long-term risk seriously, and researchers should keep studying it. A thought experiment should not outrank a randomized trial when patients are on the line.

A slowdown also has no way to tell the difference between the tools. The Food and Drug Administration has authorized more than 1,500 AI-enabled medical devices. Most of them do dull, vital work, like spotting strokes on CT scans or flagging dangerous drug interactions, under regulatory review and clinician supervision. None of them resembles the runaway intelligence in the headlines. If the people calling for a slowdown mean only frontier models, they should say so plainly and write the carve-out for medicine themselves. If they do not, they are asking hospitals to shelve a stroke detector over fears about a technology it does not resemble.

The cost of that caution would not fall on the people asking for it. It would fall on a patient in a community hospital whose sepsis alert never fires. It would fall on a woman whose early tumor waits another screening cycle for a tired radiologist to catch it. A policy that sounds careful in a boardroom can be reckless at a bedside.

The skeptics inside medicine have a better argument, and it deserves a straight answer. Nurses marched in cities across the country last year against monitoring tools they said contradicted their clinical judgment and flooded them with faulty alerts. I have watched hospitals scale back or rebuild AI deployments that produced noise instead of insight, and they were right to do it. Pew Research Center polling released in August found that 72% of adults want providers to disclose any use of AI in their care, and nearly half do not know whether it already has been used. Those findings also answer the people calling this a hoax: The risks are real, and they show up in ordinary tools, not science fiction. In health care, trust is the only currency that matters. You earn it with evidence, disclosure and answers that trace back to the source.

That is the standard medicine should hold every AI tool to: one application at a time, measured against the system patients actually live in. Tools that fail get pulled. Tools that pass get used. A blanket slowdown skips that test and freezes a status quo that already loses hundreds of thousands of people to preventable error. The people arguing about the future of AI should explain to the patients in front of us why their care has to wait.

Mika Newton is a health care technology innovator and CEO of xCures, where he has led the company's transformation into a SaaS platform delivering AI-powered precision medicine and real-time clinical decision support.


Related to this article

Vincent Cantone: ©WellSpan Health
With projected shortfalls of 86,000 to more than 187,000 physicians within a decade, burnout, an aging workforce and capped residency slots are squeezing supply. Here are the training, payment, prior authorization and workplace changes that could help close the gap.