The use of artificial intelligence in health care has skyrocketed in the past few years, from taking visit notes to summarizing medical research to reading imaging and helping make diagnoses.
This year more than 80% of U.S. physicians reported using AI in their work, more than twice the percentage who did in 2023, according to a survey by the American Medical Association. And more than 70% of respondents said they think AI can help improve health care, particularly by making the system more efficient and helping to reduce health care worker burnout.
But with the benefits of AI use come risks. Some early research suggests that using AI can hurt a person’s critical thinking, creativity, and development and retention of skills.
“There’s really just no denying how helpful AI can be,” says Fares Alahdab, MD, an associate professor and director of graduate studies in health informatics at the University of Missouri [Mizzou] School of Medicine. “[But] there’s a downside that creeps in over time, very silently.”
In a 2025 experiment, researchers at MIT measured people’s brain waves while assigning them to write an essay, either with or without AI assistance. The researchers found that, when using the AI, people’s brain activity was reduced.
Another study, which analyzed 370,000 college-admission essays, observed that, after the advent of generative AI chatbots, the range of original ideas shrunk, suggesting that AI use could also stunt creativity.
“A lot of people, when they think of creativity, they think about art,” says Adam Green, PhD, a neuroscientist who led the study and is director of the Georgetown University Laboratory for Relational Cognition. “But creative intelligence and creative problem-solving also lead to innovation, especially in the areas of science, medicine, and technology. Being able to solve problems in new ways requires us to think for ourselves.”
Leaders in medical education find themselves in a particularly challenging position, as many are trying to balance the use of AI as a tool while ensuring their students develop the cognitive skills that are essential to practicing medicine.
“There’s a sense of developing a skill by practice. If AI completely eliminates that challenge, [learners] don’t end up developing the skill in the first place,” Alahdab says. “The whole field is trying to figure out the dimensions of the risk and what, realistically, can be done without throwing the baby out with the bathwater.”
The role of critical thinking in medicine
Many medical students and a reported 65% of doctors have turned to OpenEvidence, an AI application that is trained on only peer-reviewed medical journals, as a resource for helping to diagnose patients. The model’s parent company, OpenAI, announced last year that OpenEvidence was the first AI to score 100% on the United States Medical Licensing Examination, and a 2025 Mayo Clinic study found that it was generally accurate as a supplement to primary care. However, a pilot study published in December 2025 showed that OpenEvidence answered only about a third of questions related to complex and subspecialty clinical cases correctly.
Some leaders in medical education warn that medical students, who are often overwhelmed and short on time, are particularly at risk of over-relying on AI and not taking the time to question its output.
“There’s a shift I’m seeing over the past two years, that when students have questions during an application exercise or during small-group discussion, they immediately go to Google AI summary, ChatGPT, or OpenEvidence,” says Youngjin Cho, PhD, a professor of medical education in biochemistry and immunology at the University of Georgia School of Medicine. “OpenEvidence is giving a pretty good summary, with references, and can be efficient at times, but I do worry that it cuts [learners’] essential skills related to reasoning on their own, knowing the right resources to find information, and judging the information.”
Alahdab has noticed similar trends among his students too.
“Before they know it, they’ve started to off-load their cognitive skills,” he says. “They’re starting to over-trust [AI].”
Medical educators are developing methods to combat this temptation among their students.
Alahdab served on a Mizzou School of Medicine committee tasked with reworking the curriculum so it integrated AI while preserving critical thinking and scholarly inquiry. For example, the students might be assigned to assess output from an AI model and to determine which parts are accurate and which are hallucinated.
Rohini Ganjoo, PhD, an associate professor of biomedical laboratory sciences at the George Washington University School of Medicine and Health Sciences who studies AI, sometimes directs her students to solve problems in class with their laptops closed.
Cho has developed guidance that says students should be instructed to use their brain first and check AI second, be taught how to critique AI output, and be directed to reflect on how using AI was helpful in the learning process.
“Historically, medicine has rewarded whoever has the most information. We look toward a physician as a repository of knowledge, but that reality is changing,” Ganjoo says. In the future, “it’ll no longer be about who knows the most facts. It’ll be about who will be able to challenge that output generated by AI and have that mental ability to override it, if needed.”
The risk of de-skilling
But medical students and new doctors are not the only ones at risk of surrendering their cognitive skills to AI. Another concern for AI use in health care, which a recent industry survey found worries nearly 75% of clinicians, is the risk of de-skilling — the dulling or loss of a previously learned skill — among experienced doctors.
So far there is little empirical evidence that AI use will lead to widespread professional de-skilling, but one study from Poland in 2025 raised concerns that experienced doctors could start losing clinical skills quickly when they start relying on AI assistance.
For a study published in the Lancet in 2025, researchers followed gastroenterologists at four clinics in Poland and compared the physicians’ rates of detecting benign tumors before an AI-assisted endoscopy tool was implemented and then again after three months of using the tool. The analysis found that when the physicians reverted to not using the tool, their tumor detection rate dropped by six percentage points, from 28.4% to 22.4%.
Robert Wachter, MD, professor and chair of the Department of Medicine at the University of California, San Francisco, and author of the book A Giant Leap: How AI Is Transforming Healthcare and What That Means for Our Future, called the Lancet study sobering.
“The tool became something of a crutch,” he says.
According to Wachter, though, de-skilling isn’t always a problem.
“My own feeling is that it’s important to be clear-eyed about it,” he explains. “I no longer know how to read a map, and that’s fine. I sometimes forget my wife’s cell phone number. We have basically outsourced an old skill to a technology.”
But, since AI tools are not always readily available, Wachter adds that it’s important for physicians to stay sharp.
He believes that medicine can learn from other professional fields. Airline pilots, for instance, must pass simulator tests every six months to ensure they’re not de-skilling from using autopilot. Wachter suggests designing AI tools that periodically test or coach the user instead of just produce answers.
A separate study that looked at Danish endoscopists found that experienced physicians’ tumor detection rate increased by more than 12% with the use of AI assistance. The same study showed that inexperienced endoscopists did not significantly improve their detection rates. These findings suggest that AI can enhance an experienced physician’s skills but may not replace the development of those skills in the first place.
Arthur Lazarus, MD, MBA, an adjunct professor of psychiatry at Temple University’s Lewis Katz School of Medicine, in Philadelphia, and author of the book Practicing in the Age of AI: Essays on Medicine, Meaning, and Machines, warns that even skilled physicians who become complacent and overrely on AI could miss things or fail to question a hallucinated or incomplete AI output.
“Do the work first, consult AI second. That’s my basic recommendation to everyone,” Lazarus says.
But the threat of de-skilling should not be seen as a reason to shy away from using AI in medicine, Wachter adds.
“The status quo is immoral. We have a health care system that demonstrably does not provide the quality that people want and deserve. We’ve got burned out clinicians, and we have costs that are bankrupting individuals, companies, and governments,” he says. “If we say AI is too scary, we’re going to stick with the status quo. I think that’s not a neutral decision, that’s a negative decision.”
The case for creativity
Beyond preserving critical thinking, leaders in academic medicine also have the task of encouraging creative thinking — a quality that Green sees as distinctly human.
Creativity is essential, not only for clinicians, who must find new ways of improving their patients’ lives, but also for medical researchers, who are working on the discoveries of the future, he says.
“It’s obvious to most of us that we’ve got some problems that need solving and the kinds of ideas we’ve tried haven’t worked,” Green adds. “If we don’t continue to generate truly new ideas, there won’t be a basis for really new solutions.”
Green’s lab uses AI to identify what originality humans bring beyond what AI can provide. The lab partners with institutions, including college admissions offices looking to measure applicants’ originality in addition to any AI they use in their applications.
He says that AI can help speed up the creative process with its quick analysis, refinement, and modeling capabilities, but he worries that it has become too easy for humans to turn to AI before trying to come up with their own ideas.
“We know that these capabilities are atrophying. Measures of intelligence and critical thinking are plateauing and decreasing for the first time in the internet era,” he says. “I have a hard time believing humans will choose to do more effortful thinking in spite of the much more powerful alternatives AI presents for avoiding that effort.”
This is why he thinks it will be up to leaders and systems to incentivize creativity and critical thinking. In medical education, that means testing and rewarding students and residents for, not just performance and outcomes, but also originality that shows thinking beyond AI’s homogenized solutions, willingness to take calculated risks, and diligence in questioning the given answers.
Health systems may additionally have to wrestle with how to keep more experienced clinicians engaged, while taking advantage of the value that AI can bring to the profession.
“Physicians are not the people who have the most time on their hands, which makes it a slippery slope to revert to these tools. That in and of itself is not necessarily bad. We do want them to save time, focus on the most important aspects of the job, and have a work-life balance,” Alahdab says. “At the same time, long term, there’s a silent risk [of overreliance] that creeps in. This is the paradoxical nature of AI. It makes it a very challenging type of risk to be aware of and feel when it’s starting.”
