Research suggests doctors might quickly become dependent on AI

Artificial intelligence is beginning to help doctors screen patients for several routine diseases. But a new study raises concerns about whether doctors might become too reliant on AI.

The study looking at gastroenterologists in Poland found that they appeared to be about 20% worse at spotting polyps and other abnormalities during colonoscopies on their own, after they’d grown accustomed to using an AI-assisted system.

The findings, published in the journal Lancet Gastroenterology and Hepatology, suggest that even after a short period of using AI, experts may become overly dependent on AI to do certain aspects of their jobs.

“We were quite surprised,” says Marcin Romańczyk, an M.D.-Ph.D. gastroenterologist at H-T Medical Center in Tychy, Poland, who led the study.

But not everyone is convinced that the paper proves doctors are losing critical skills because of AI.

“I think three months seems like a very short period to lose a skill that you took 26 years to build up,” says Johan Hulleman, a researcher at Manchester University in England who has studied human reliance on artificial intelligence.

Hulleman believes statistical variations in the patient data might be part of the explanation for why the numbers appear to drop. Factors such as the average age of the patients used in different sections of the study might explain the variation, he says.

AI in medicine

Artificial intelligence is becoming increasingly common for a number of routine medical scans. The next time you get a scan for your eyes, breast cancer, or colon disease, there’s a decent chance that AI might be analyzing the images.

“AI is spreading everywhere,” Romańczyk says. At the same time, many doctors are playing catch-up, because learning how to use the technology wasn’t part of their training.

“We’ve been taught from books and from our teachers,” he says. “No one told us how to use AI.”

A few years ago, four clinics in Poland tried out an AI system to detect polyps and other abnormalities during colonoscopies. The AI works in real time, analyzing video from a camera inside the colon.

If it spots something, it will highlight the area for the clinician to see.

“In this particular one there’s a green box, showing where the polyp might be,” he says.

In the box

The clinics were collecting data on whether the AI system worked. It turns out it did, but when Romańczyk and his colleagues reanalyzed the data, they found something else: After the system was introduced, doctors became significantly worse at detecting possible polyps when the AI was switched off.

According to their analysis, after doctors got AI, detection rates of possible polyps fell from 28.4% to 22.4% when their new AI system was switched off. In other words, the doctors seemed to become quickly dependent on AI systems catching the polyps. Romańczyk says he’s not quite sure why it’s happening, but he has some theories.

“We are subconsciously waiting for the green box to come out to show us the region where the polyp is and we’re not paying so much attention,” he says.

There are other examples that support that idea: A similar study has shown that nonexperts do a worse job scanning mammograms if they know they can get an AI system to help them with the push of a button.

Johan Hulleman, who helped lead that mammogram study, describes it as a “safety-net effect.” He says these latest results could be interesting, but he’s skeptical. The study of colonoscopies took place over three months, and the doctors participating had decades of experience. He thinks statistical variations due to a number of factors, such as age of the patients, might be behind the apparent drop.

Additionally, he says, “We don’t know how many polyps there really were, so we don’t know the ground truth.” By that he means it’s unclear how many of the possible polyps that doctors supposedly missed were actually medically important.

The study’s author, Romańczyk, does believe the drop is real — though he admits that studying AI in a clinical setting like this can be tricky. There are a lot of variables the researchers couldn’t control.

He’s not against using AI. He actually thinks the little green boxes help him to do better colonoscopies. But he thinks that there should be more studies like these examining how AI might be changing the way doctors work in the real world.

“Because look what’s happening,” he says. “We have AI systems that are available, but we don’t have the data.”