Study: AI Generates ‘Severe’ Errors in 22% of Medical Cases

A new study from researchers at Stanford and Harvard found that even today’s best artificial intelligence (AI) models make serious errors in a significant portion of medical cases, PPC Land recently reported. The results, released Jan. 2, found that the top AI models generated “severely harmful clinical recommendations” up to 22.2% of the time, with the top-performing AI models producing 12 to 15 errors per 100 cases and the worst-performing models making mistakes in 40 out of 100 cases.

The research evaluated 31 large language models and found that errors of omission — in which the AI models failed to recommend necessary tests or treatments — made up more than 76% of severe mistakes, the publication reported. Mistakes that included inappropriate or incorrect recommendations were less common, according to the report.

“That is a very high number of errors,” said Chris Valenzuela, Broker, Professional Liability, Burns & Wilcox, Los Angeles, California. “It is concerning when you consider that we could start seeing that type of system installed across the board at healthcare facilities.”

That is a very high number of errors. It is concerning when you consider that we could start seeing that type of system installed across the board at healthcare facilities.

The findings come as more healthcare professionals appear to be using AI tools on the job. A survey from Wolters Kluwer, released Jan. 27, indicated that 46% of U.S. nurses use AI tools at work, while 77% believe generative AI will be important to their organizations’ future productivity. About 45% of nurses said the tools could help with documentation, triaging routine patient questions, and streamlining workflows.

As AI use in healthcare continues to evolve, organizations may need to establish policies regarding the use of AI and confirm how their Healthcare Liability Insurance could respond to potential lawsuits over AI-related errors, said Matt Baxter, Director, Professional Liability, Burns & Wilcox, Brokerage Division, Atlanta, Georgia.

“From an insurance standpoint, AI is not really changing the exposure, because the liability still stands with the healthcare professional,” Baxter said. “They still have the same responsibility, whether they are using AI or not, to make sure the information is correct.”

From an insurance standpoint, AI is not really changing the exposure, because the liability still stands with the healthcare professional. They still have the same responsibility, whether they are using AI or not, to make sure the information is correct.

Clinical risks increase as AI plays larger role in patient care

According to data released in September 2025 from the U.S. Department of Health and Human Services, about 71% of hospitals reported using predictive AI in electronic health records in 2024, up from 66% in 2023, with hospitals using the tools most often to predict health trajectories or risks for inpatients. In December, Becker’s Hospital Review reported on growing concerns among clinicians about the safety and accuracy of AI-powered clinical decision support tools.

“AI is becoming widespread not only in healthcare, but across all industries,” Baxter said. “At this stage, AI is a good tool and resource, but the information needs to be checked and verified by an actual human. Especially in healthcare, everybody needs to be very careful.”

According to Valenzuela, many healthcare organizations today are primarily using AI for note-taking and documentation rather than decision-making and diagnosis. “In the past, you would have a doctor working with a scribe who would follow along and take notes. Now doctors are using AI on their phone or tablet to record conversations and summarize them,” she said. “That is where I have seen AI implemented the most, including in telehealth and virtual care.”

Baxter agreed, emphasizing the distinction between administrative and clinical uses of the technology. “There is a big difference between using AI for back-office administrative work and using AI for diagnosing,” he said.

Doctors should still be making decisions about diagnostic tests, Valenzuela said. “Having hands-on physician oversight is extremely important, especially in the initial stages of implementation,” she said.

Having hands-on physician oversight is extremely important, especially in the initial stages of implementation.

How insurance could respond to AI-related claims

AI-related errors in healthcare could lead to significant financial impacts, including legal expenses and potential settlements in the event of a lawsuit. These can be covered by Healthcare Liability Insurance, a type of Professional Liability Insurance also known as Medical Malpractice Insurance. “As far as cost, it depends on the allegations and the claim,” Valenzuela said. “Because it is healthcare, it is often a bodily injury situation where someone is not treated correctly. Anytime there is a bodily injury, it can be costly.”

These policies typically do not specifically address the use of AI with either affirmative coverage or exclusions, Baxter said. “For the most part, policies are silent on it,” he said.

Valenzuela said claims or lawsuits related to AI errors in healthcare “would probably be handled like any other error that occurred” through Healthcare Liability Insurance — though this may change over time, she said. “As more healthcare professionals use AI for a variety of reasons, there may be an additional question on policy forms about the extent to which AI is used, and potential sublimits,” she explained.

In the meantime, Baxter said, healthcare organizations should keep their insurance broker updated on the integration of any new AI tools.

“We are in a watch-and-see mode as an insurance industry as far as where AI is heading,” he said. “The whole idea is that AI will continue to get more intelligent and more sophisticated. What we are seeing from AI today will be very different a year from now. Just because policies are silent today, next year we might start seeing exclusions or wording about what it will and will not pick up when an organization is using AI.”

Strengthening risk management around AI use

If AI systems are handling sensitive patient information, informed consent, data security, and regulatory risks must also be considered, Valenzuela said. Cyber & Privacy Liability Insurance can help address breach-related expenses, including regulatory fines and investigations.

“Sometimes scribing AI involves ambient listening. Having that informed consent, and how that is communicated to patients, is very important,” she said. “If you are using AI, you have to make sure it is HIPAA compliant. If there is a leak of information, you will be held liable for that.”

According to the Wolters Kluwer report, only 22% of nurses surveyed indicated that their organizations had published policies on the use of generative AI, and the same percentage reported that their organizations required formal training before deploying AI tools in nursing workflows.

“There should definitely be a policy in place,” Valenzuela said. “In the absence of those policies, you are in a gray area. Having the policy in place, and implementing it, is key to protecting yourself.”

Close supervision will remain crucial as AI tools are further integrated into healthcare operations. “Ultimately, the best way to handle AI is oversight,” she said. “In whatever scope you are using AI, you always want to have a doctor’s touch to it.”

Organizations should thoroughly research potential AI tools and ensure that proper “checks and balances” are in place, Baxter added. “They are still just as liable as they would have been before AI for any incident that comes out of it,” he said. “Therefore, they need to be as cautious, if not more cautious, than they were before. Organizations may see this as a money-saving opportunity, but we are not there yet. We have a long way to go before that trust is built.”

Organizations may see this as a money-saving opportunity, but we are not there yet. We have a long way to go before that trust is built.