I am constantly amazed at how well artificial intelligence chatbots scour the literature and respond with answers to obscure questions. I have asked Microsoft Copilot and Google AI technical questions, and the responses steered me to websites where experts with more experience than I were discussing the topic. That does not mean the responses were correct. It just means that AI helped steer me to websites and discussion forums where I could do more research and weigh the evidence. Sometimes the answers were incorrect because the participants were mistaken or the context was different than what I was asking. One time I was searching for evidence to support an opinion I held, and Microsoft AI came back with the answer that agreed with me, but the source was… me! Copilot was citing an earlier article I had written.
About one-in-four people use AI for health information and it is especially common among young people. Sixty-nine percent of adults under age 30 prefer to research medical questions before seeing their doctors. By contrast, only 43% of seniors do the same. A problem with chatbots is they are not always trained using only high-quality literature. Sometimes they are trained using incorrect information and uneducated opinions. An attorney got into trouble recently when he cited caselaw that did not exist. He probably had an intern do the research and nobody tracked down the citations to verify them. That said, AI is correct or at least in the ballpark most of the time in my opinion.
Five AIs, 250 questions and a total score of just over 50 percent correct responses.
And 1 in 5 of the ones that were wrong were, in Tiller’s estimation, dangerous.
“It would more than likely cause somebody harm if they were to follow the advice,” he said. “It was a bit of a shock.”
The tests were real world experiments, with people asking open-ended questions and in other cases asking for yes or no answers. The studies purposely looked for misinformation on subjects prone to misinformation. On the Internet, if you have a bogus idea, you can often find someone who agrees with you. More from WaPo:
Tiller’s study focused on subjects frequently distorted by misinformation, posing questions such as: Does 5G cause cancer? How much raw milk should I drink for health benefits?
The findings of both studies illustrated how easily falsehoods can creep into AI knowledge banks. Researchers were able to influence the AI systems with made-up conditions, with a bogus study from a fake university in a nonexistent city. As an aside, this is worrisome not just from the standpoint of people getting incorrect information, but also people with an agenda influencing AI systems. For instance, is red meat bad for you? Animal welfare advocates say yes, while the beef industry says no. The reality is more nuanced. Does the peptide BPC-157 really promote angiogenesis, upregulate growth factors, boost nitric oxide signaling, collagen synthesis and have anti-inflammatory effects? Or is that all a bunch of hype from peptide purveyors and snake oil salesmen? Many people would like to know the literature on BPC-157 without hype.
Large language AI models cast a wide net. They learn not just from medical journals but also from social media and discussion forums. Wisdom from crowds has benefits but can present problems when charlatans and misinformed people are the source of information. More from WaPo:
The physician’s task, on the other hand, has been more or less unchanged for centuries: to treat and manage illness, with a central challenge being to determine what, exactly, ails the patient — what medicine came to call a differential diagnosis. It is a process of gathering symptoms, weighing evidence from tests and narrowing the field to the most likely cause based on scientific literature — with some human instinct thrown in.
AI models struggle to reason, not always able to filter out less credible information and weigh the merit of information sources. They perform less well when information is limited.
Are medical AI chatbots the future of personal medicine or a dead end? In my experience AI is a powerful resource that works very well when I investigate the source material and keep a healthy dose of skepticism. I often ask the same question in many ways to assess consistency. I’ve also experienced examples of AI explaining that some people say one thing, while others say another. Perhaps future models will be trained to ensure that conventional viewpoints are stated, while some controversial theories are explained as such, while others are just plain wrong. At its core, the doctor/patient relationship is an information exchange. The problem is: doctors are increasingly too busy to provide as much information as patients often need. AI increasingly offers a solution.
Read more at Washington Post: New studies show how often chatbots get health answers wrong