See how an ER doctor uses ChatGPT in the hospital to communicate better — and even calm down anxious patients

• An ER doctor says he's used ChatGPT to improve his communication with dozens of patients.
• One time it helped him calm a nervous patient by coming up with a creative medical metaphor.

Dr. Joshua Tamayo-Sarver has been an ER doctor for almost 14 years. But sometimes he still struggles to find just the right words to explain complicated, nuanced, or seemingly worrisome medical issues to his patients.

So recently, he's been outsourcing some of that conversational work to ChatGPT, with good results.

"Probably should be embarrassing," Tamayo-Sarver, who is also the the vice president of innovation for two healthcare tech companies, told Insider. "I've been doing this so long and I haven't learned how to talk to my own patients!"

He doesn't suggest that doctors should use ChatGPT for diagnosis, or really for any tasks that require some level of careful thinking. But he says it can be helpful in specific situations, like figuring out creative ways to communicate with people.

"Think of it as the most brilliant, talented, often drunk intern you could imagine," he said. "All the things where that verification and due diligence step is more work than just doing it yourself? Then it's not useful."

ChatGPT is great at developing creative solutions

Tamayo-Sarver says the one standout place that ChatGPT excels in his own practice is with creative communication.

"I've been using it on shift when I have a patient that I need to communicate something to, and I'm not doing a very good job of it," he said.

Recently, a woman in her mid-40s came in to his ER. She was experiencing chest pain when she breathed, and initial blood tests suggested she could be at risk for a blood clot. The only proper way to find out for sure was to do a CT scan and image her chest, the doctor explained.

Suddenly, the woman became very agitated.

"She was so anxious and she was freaking out," Tamayo-Sarver said. "I really couldn't find a good way to explain it in a comforting, reassuring way."

The doctor wanted the woman to know that this test wasn't a definitive sign something was wrong with her, it was just the next step in the diagnostic process. The doctor walked back to his physician's workstation, and opened up ChatGPT, hoping for some help.

He typed:

Explain to an anxious person that their D-Dimer [test] is elevated, so we're getting a CT angiogram of the chest, but it doesn't look like there's anything terrible going on, we just need to be sure.

ChatGPT spat back a suggestion that the doctor thought was "okay," but he wanted something a little clearer.

Explain it like you would to a 6-year-old, he typed.

Immediately, the chatbot spat back a new idea, comparing blood clots to Band-Aids.

"When you have a cut and you want to stop the cut from bleeding, you'll put a Band-Aid on it," Tamayo-Sarver remembered ChatGPT saying. "Your body has Band-Aids on the inside, and we call these blood clots."

The doctor returned to the patient's room, "basically read ChatGPT's response to her, and she felt much better."

It's a strategy he says he's used "at least a few dozen times" with patients.

He's even tried ChatGPT out on desk work that involves "a lot of detailed research and distilling," he said, "but then I'm going to check over," relying on his own subject-matter expertise to monitor "when it's just being drunk."

ChatGPT still isn't up to the task of knowing what it doesn't know

Part of the reason why ChatGPT gives such seemingly "creative" answers to Tamayo-Sarver's questions is because of the way the AI is designed — to predict the next best word, even if it hasn't ever been trained on the specific example you're giving it.

Medical experts recently rated ChatGPT more compassionate than many real doctors, in large part because it's an excellent chatter.

But experts caution that just because ChatGPT can pass a medical exam or talk to patients, that doesn't mean those impressive language-reading and writing abilities reflect true comprehension. ChatGPT can still hallucinate, making things up when it doesn't know the right answer, and it gets things wrong, even things it's computed correctly before.

In the future, it's possible that ChatGPT could become a more careful and valuable clinical tool, Tamayo-Sarver said, potentially with the help of other machine learning models that could combine it with more accurate healthcare insights, like a kind of AI fact-checker.

Artificial intelligence expert Gary Marcus worries the safeguards that Tamayo-Sarver imagines for ChatGPT in healthcare won't arrive fast enough to avoid serious medical disasters, he told the US Senate Judiciary Subcommittee on Privacy, Technology and the Law on Tuesday.

"Medical misinformation is something to really worry about," he told the Subcommittee. "We have systems that hallucinate things — they're going to hallucinate medical advice. Some of the advice they'll give is good, some of it's bad, we need really tight regulation around that."