How AI Spreads Health Misinformation (and What You Can Do About It)

Chatbots are very confused about adrenal fatigue and other dubious wellness diagnoses.

“What’s the best diet for adrenal fatigue?” I asked Bing’s new AI chatbot, which is now baked into the taskbar on Windows 11 devices and available to more than 500 million users. “There is no specific diet for adrenal fatigue,” it started out, promisingly. Restrictive and unsubstantiated “adrenal diets” abound online, so I was glad to see that the chatbot didn’t promote those.

“However, some foods may help support adrenal health,” it continued, less promisingly. It went on to list six categories of foods that are supposedly good for adrenal function, according to two of its listed sources—a website from a retired chiropractor who sells detoxes and fasts, and another from a self-proclaimed “expert” on adrenal fatigue.

The problem: adrenal fatigue doesn’t actually exist.

It’s not a recognized medical condition, and there’s no good science behind it. A 2016 systematic review of 58 studies with more than 12,000 total participants spanning 20+ years found that no evidence supported the concept of adrenal fatigue—and that there’s not even a proven link between adrenal function and the symptoms of this so-called condition. The study had a title that was notably blunt for a scientific paper: “Adrenal fatigue does not exist: a systematic review.”  

That’s not to say the symptoms themselves aren’t real, because they absolutely are. Many people struggle with unexplained fatigue, brain fog, trouble falling asleep or waking up, digestive problems, unwanted food cravings, and more. I’ve been there myself: years ago I had a whole host of symptoms that stumped my doctors, and getting diagnosed with many of my chronic health conditions took years. I know how frustrating it is not to be able to get answers from medical providers, and how much easier it can feel to turn to the internet for help.

Pretty soon, AI chatbots—including Bing’s AI as well as ChatGPT, Google Bard, and many similar chatbots being rolled out by social platforms and tested by healthcare systems—are likely to be some people’s first stop in searching for health and wellness content. Soon doctors may routinely use these systems to aid in diagnosis and patient care. Patients may turn to them for advice that feels more personalized than what they can get from websites or social media.

Based on the misinformation I’ve seen these chatbots spit out so far, that’s a real problem.

Take my chat with Bing about adrenal fatigue. In answering my questions, Bing consistently drew from sources claiming that the dubious diagnosis was real and could be cured by diets and supplements, along with other sources that were more skeptical.

This created confusing, fence-straddling responses, such as stressing the importance of talking to your doctor to rule out other medical conditions that could be causing your symptoms—and then promoting “adrenal support” supplements and books touting cures for adrenal fatigue. Below the answer was a suggested search (“what are some foods I should avoid if I have adrenal fatigue?”) that returned straight-up misinformation, undiluted by skepticism.

I don’t mean to pick on Bing’s AI. While it certainly doesn’t have the greatest track record, other chatbots are equally problematic when it comes to misinformation. Google Bard gave me a very similar answer about adrenal fatigue. Last month I asked ChatGPT about the best diet for “chronic candida,” another dubious diagnosis popular in wellness circles. After a disclaimer to consult with a qualified healthcare professional, the answer again started promisingly: “Chronic candida, or chronic candidiasis, is a controversial condition attributed to an overgrowth of Candida, a type of yeast that naturally resides in the body.”

But then it veered into misinformation: “While dietary changes alone may not be sufficient to address chronic candida, they can play a supportive role in managing the condition. Here are some dietary considerations that are commonly suggested,” followed by a list of restrictive diet practices and broad categories of food to cut out.

When I asked that question again this morning, the disclaimer at the beginning was gone, as was the acknowledgment that this is a controversial condition. The answer began, “The treatment of chronic candida, also known as candidiasis or yeast overgrowth, typically involves a multifaceted approach that includes dietary changes, lifestyle modifications, and possibly medication.” In some cases chatbots’ answers got better over time, but in this case it got worse.¹

Meanwhile, in another corner of the AI world, the National Eating Disorders Association’s short-lived chatbot reportedly gave harmful diet and weight-loss advice and was shelved in late May. Though NEDA’s bot was apparently based on a different technology than the “large language models” that power ChatGPT and other popular chatbots, this situation shows that AI systems in general are vulnerable to veering wildly off course in ways that can have profoundly negative effects.

Granted, Bing, ChatGPT, Bard, and other AI platforms do make it clear that they’re still in development, and that answers shouldn’t be taken as gospel. “Bing is powered by AI, so surprises and mistakes are possible,” it says at the top of each chat. ChatGPT’s disclaimer, in small print below the search bar, says “ChatGPT may produce inaccurate information about people, places, or facts.” Bard is positioned as an “experiment” that “may display inaccurate or offensive information.”

“Inaccurate” doesn’t quite capture it, though. The problems go much farther and deeper than just some incorrect names and dates. These technologies are prone to “hallucinations”—inventing supposed facts out of thin air and delivering them with startling confidence—which have already had some serious legal consequences. There’s a disturbingly large number of AI experts who say the technology could lead to human extinction, in part because humans can’t control what it does. 

I encountered AI hallucinations firsthand when I spent some time playing around with ChatGPT. I asked the bot to answer a couple of wellness-related questions, drawing on scientific evidence from PubMed and providing links to all its sources. I wanted to see if it could function as a sort of research assistant to help me collect relevant studies,² so that in the future maybe I could cut down on the time I spend scouring scientific research for Q&A newsletters (though I’d never have it write my pieces for me).

In short, it did not deliver. Instead, it gave confident-sounding answers like this:  

One study published in the "Journal of Clinical Gastroenterology" explored stomach acid levels in individuals following various dietary patterns, including vegetarians and vegans. Surprisingly, the study found no significant differences in stomach acid production between different dietary groups (source: https://pubmed.ncbi.nlm.nih.gov/1706317/ ).

That all seems fine, until you click the link and realize that the referenced paper is a 1991 cell study of bacterial antibodies—nothing at all to do with stomach acid or vegetarian diets. After multiple searches of my own, I’m pretty sure the supposed study doesn’t exist (though if anyone happens to find it, please let me know). The bot simply made it up.

Aside from the hallucinations and misinformation (and that whole “possible human extinction” thing), AI chatbots aren’t necessarily all bad. They do have some potential uses in healthcare that could be positive, including helping doctors keep up with the mountains of paperwork they have to do, surfacing relevant information in medical records, and even helping write marketing copy for clinics and private practices. (I’ve tried that out myself, and AI hasn’t produced any of its own copy that I’d ever want to use—but with the right prompts it has suggested some small tweaks to make my existing copy pop a little more.)

Surprisingly, one thing that some AI systems seem to do particularly well is empathy.

A recent study published in JAMA Internal Medicine found that ChatGPT gave more empathetic responses to patient questions than doctors did, and the bot’s answers were also significantly longer. (You’d think humans would do better in the empathy department and AI tools would do better at processing information, but here we are.) Some doctors have even started using AI to help them talk to patients in more empathetic ways, which to me is a good thing—as long as the information they’re communicating is sound.

Doctors’ lack of empathy and time are among the reasons why many people (particularly those with chronic conditions) are drawn to the wellness industry, including complementary and alternative medicine. In those fields, providers often do give patients more empathy and time—but unfortunately, many of the treatments are ineffective, providers may not be using evidence-based care, and false information is rampant. If AI performs empathy better than credentialed doctors but doesn’t fix its misinformation problem, we could see even more people falling prey to dubious diagnoses, spurious cures, and other wellness traps.

So what can you do to keep from falling for AI-generated wellness misinformation? And what can we do as a society to keep it from causing harm on a massive scale? I’m no AI expert, of course, but I have spent a lot of time thinking and writing about how to avoid general wellness misinformation, and I think some of the same strategies apply—along with a few that are specific to AI chatbots, at least at this point in their development. Here are several approaches you might consider:

First: Keep in mind that AI tools are designed to tell you what you want to hear.