A commentary in a top medical journal co-authored by Vinod Khosla says AI outperforms doctors on certain tasks, sparking debate on how far startups should push the technology into medicine.

The JAMA article rejects the notion that artificial intelligence should only be used as a support to physicians. Instead, it argues that generative AI rivals or outperforms doctors on five cognitive tasks, such as prescribing treatments in line with clinical guidelines and managing chronic diseases.

Further, keeping physicians in the loop to correct errors is likely to worsen, not help, AI performance, because doctors can introduce errors or make poor decisions on when to trust AI, according to the article, which is based on a review of medical AI studies published since Jan. 1, 2024.

Superior, autonomous AI will likely be ready for real-world cognitive medical tasks in some or maybe many workflows by 2030, the article projects. In addition to Khosla, founder of Khosla Ventures, a top investor in AI, technology and healthcare, its authors include his son, Neal Khosla, co-founder and chief executive of Curai Health, a Khosla Ventures-backed startup that uses AI and human doctors to provide primary care.

"There are many domains where AI is going to be working autonomously. Not tomorrow, but soon," said Dr. Ezekiel Emanuel, an oncologist, bioethicist and professor at the University of Pennsylvania, and a co-author of the article, along with research fellow Abe Baker-Butler. "A lot of people recognize we're right."

The article, published Aug. 17, is spurring debate and criticism from physicians and entrepreneurs who say it doesn't fully account for a doctor's many responsibilities, such as managing patients over time and building their trust.

"These tools can have great value; AI certainly plays a role," said Dr. John Whyte, CEO of the American Medical Association. "But it doesn't replace a physician and it's not going to."

Healthcare-AI startups began with tools to streamline administrative tasks, but some are now driving the technology into patient care.

Typically, startups' clinical AI is meant to support, not replace, a physician's judgment, but a few are now pursuing autonomous applications. Startup Doctronic, for one, is piloting technology in Utah that enables AI to refill prescriptions for certain drugs.

The JAMA article will push the field to contemplate appropriate use of AI in medical care, said Dr. Sunny Kumar, a partner with Informed Ventures.

While some startups with autonomous AI systems have promising technology, the regulation of these tools is unclear and they often face resistance from medical societies and doctors, he added.

One criticism of the article has been that most of the studies it reviewed relied on simulations rather than actual, real-world instances of AI use in medical care. The authors acknowledged this limitation, noting that lack of clarity on issues such as regulation and liability limits efforts to study autonomous AI.

Technology alone could take off and land a plane.

"But who would board such an aircraft?" said Dr. Barry Stein, vice president and chief clinical innovation officer of Hartford HealthCare. "The real challenge isn't choosing AI alone over physicians supported by AI. We need both to achieve the safest outcomes."

Even if AI answered medical questions accurately, patients might be unwilling to accept a diagnosis from a robot. A human doctor might still be necessary to deliver medical news, said Alex Kotlar, founder and CEO of startup Bystro AI, a platform that interprets genetic data using conversational AI.

"Even if the information is communicated precisely, it might not be communicated effectively," he said.