{"id":11714,"date":"2025-11-18T05:55:20","date_gmt":"2025-11-18T10:55:20","guid":{"rendered":"https:\/\/blogs.nejm.org\/hiv-id-observations\/?p=11714"},"modified":"2025-11-20T05:50:45","modified_gmt":"2025-11-20T10:50:45","slug":"when-ai-gets-the-medical-advice-wrong-and-right","status":"publish","type":"post","link":"https:\/\/blogs.nejm.org\/hiv-id-observations\/index.php\/when-ai-gets-the-medical-advice-wrong-and-right\/2025\/11\/18\/","title":{"rendered":"When AI Gets the Medical Advice Wrong \u2014 and Right"},"content":{"rendered":"
\"\"

National Foundation for Infectious Diseases.<\/p><\/div>\n

A journalist recently reached out to ask about the shingles vaccine. We mostly talked through the usual topics \u2014 how common shingles is, why the vaccine works so well, and side effects. Plus, the whole topic of zoster vaccination has been much in the news recently given studies associating receipt of the vaccine with a reduced risk of dementia.<\/p>\n

Before we wrapped up, I suggested that hearing from someone who\u2019d had a tough time with shingles might strengthen the piece. A colleague of mine, Lisa Junker (who is Senior Director of Publishing at the Infectious Diseases Society of America), had a serious case involving her ear several years ago.<\/p>\n

It was a classic case of herpes zoster oticus<\/em>, with severe pain, vertigo, and hearing loss. What she memorably said about it when she came back to work: “Shingles:\u00a0 Do not recommend!” She graciously agreed to speak with the reporter for her WebMD piece.<\/p>\n

The article came out<\/a>, and it’s excellent. I noticed, however, that Lisa had received some wrong counseling from the ENT who cared for her:<\/p>\n

Her ENT warned that if shingles returns, it will likely strike the same spot.<\/em><\/p><\/blockquote>\n

This misconception is surprisingly common. It\u2019s right up there with the patient who\u2019s told that waxing and waning symptoms in the same dermatome after an episode represents \u201crecurrent zoster.\u201d It almost never is \u2014 which is why this consult is one of the more common outpatient questions we ID doctors receive, and why I wrote about it several years ago<\/a>. (It\u2019s rarely covered clearly in textbooks or journal articles.)<\/p>\n

In short, herpes simplex<\/em> frequently recurs in the same location, but herpes zoster<\/em> does not. Recurrent zoster usually (but not always) recurs in a different dermatome altogether, usually several years after the first episode, if it recurs at all — one study<\/a> found that roughly 10% of people had another episode within 10 years, though this may be an overestimate<\/a> since there are lots of cases of herpes simplex (and other things) misdiagnosed as herpes zoster.<\/p>\n

That’s one of the things about this unpleasant disease. It\u2019s unpredictable. And the ENT\u2019s confident statement that it’s going to come back in the same “spot” \u2014 delivered as clinical fact \u2014 simply wasn\u2019t correct. I let the reporter know, and she promptly removed the sentence.<\/p>\n

And then curiosity got the better of me. Motivated by this New York Times<\/em> piece<\/a> that highlighted that patients used AI to get medical advice, I plugged this prompt into two AI tools, OpenEvidence and ChatGPT:<\/p>\n

When people get shingles again, is it usually in the same location?<\/em><\/p><\/blockquote>\n

I deliberately left off medical terms such as “dermatome”, “reactivation”, or “recurrent.”<\/p>\n

Here\u2019s what happened: OpenEvidence hit the ball out of the park:<\/p>\n

Recurrent shingles (herpes zoster) typically does not occur in the same location; it most often affects a different dermatome during subsequent episodes.<\/em><\/p><\/blockquote>\n

With unlimited queries limited to healthcare professionals and medical students, OpenEvidence is currently the best AI tool for medical information. I\u2019ve written about it elsewhere<\/a> as the \u201cGoogle for Doctors \u2014 and nurses and PAs and anyone with an NPI.\u201d Not only did it provide the correct answer, it offered several highly relevant citations. (Disclosure<\/em>: NEJM Group, including NEJM Journal Watch, is among the publishers providing content to OpenEvidence.)<\/p>\n

ChatGPT? Let’s stick with the same metaphor (can you tell I miss baseball?): it was more Charlie Brown than Aaron Judge, a big whiff:<\/p>\n

When someone has more than one episode of shingles, the new episode most commonly appears in the same dermatome as the prior one. This fits the biology: the varicella-zoster virus remains latent in a specific dorsal root (or cranial nerve) ganglion, and reactivates from that same reservoir.<\/em><\/p><\/blockquote>\n

Not just wrong, but wrong with a pathophysiologic justification! Brilliant. This confident but incorrect answer echoed the ENT\u2019s advice and underlies the very misconception that leads to clinical referrals asking about “recurrent” zoster.<\/p>\n

When I corrected the model, it “thought” for 35 seconds, then acknowledged the error (“You’re right to push back on that”) and gave the correct answer. It even cited my 2014 blog post! So proud.<\/p>\n

But the comparison was a useful reminder of how these systems work: highly fluent, highly confident, occasionally wrong.<\/p>\n

And so, in the span of a single afternoon, the same misconception surfaced from three different sources: a clinician, a widely read health article citing that clinician, and an AI model. What can we take from this?<\/p>\n