Google uses artificial intelligence to answer your health questions. Should you trust him?

Do you have a headache or is it a sinusitis? What does a stress fracture look like? Should you be worried about pain in your chest? If you search for these questions now on Google, the answers may be written by artificial intelligence.

In May, Google rolled out a new feature called AI Overviews that uses generative AI, a type of machine learning technology that is trained on information from across the internet and produces conversational answers to some search questions within seconds.

In the weeks since the tool’s launch, users have encountered a wide range of inaccuracies and bizarre answers on a range of topics. The company later appeared to roll back the feature for some searches as it tried to reduce these errors.

When it comes to AI answers to health questions, experts said the stakes are particularly high. This technology can guide people toward healthy habits or needed medical care, but it also has the potential to provide inaccurate information. Artificial intelligence can sometimes make up facts. If their answers are shaped by websites that are not scientifically based, they may provide advice that contradicts medical advice or poses a risk to a person’s health.

The system has already been shown to produce poor answers that appear to be based on faulty sources. When asked “How many rocks should I eat,” for example, AI Overviews asked some users to eat at least one rock a day to get vitamins and minerals. (Advice copied from Onionssatirical site.)

“You can’t trust everything you read,” said Dr. Karandeep Singh, chief health AI officer at UC San Diego Health. In health, he said, your source of information is essential.

Health searches have “additional guardrails,” said Hema Buddaraju, a director of product management at Google who is helping lead work on the AI ​​overview, but declined to describe them in detail. Searches that are considered serious or explicit, or that indicate someone is in a vulnerable situation, such as self-harm, do not trigger AI summaries, she said.

Google declined to provide a detailed list of websites that support the information in AI Overviews, but said the tool works in conjunction with… Google Knowledge Graphan existing information system that extracted billions of facts from hundreds of sources.

New research responses identify some sources; For health questions, these are often websites such as the Mayo Clinic, WebMD, the World Health Organization, and the scientific research center PubMed. But it’s not an exhaustive list: the tool can also pull from Wikipedia, blog posts, Reddit, and e-commerce sites. It does not tell users which facts came from which sources.

Using a standard search result, many users will be able to immediately differentiate between a reputable medical website and a candy company. But a single block of text that combines information from multiple sources can cause confusion.

“That is if people are looking at the source,” said Dr. Seema Yasmin, director of the Health Communications Initiative at Stanford University. “I don’t know if people are looking, or if we have really educated them enough.” To look at,” she said Her own research Because of the misinformation, it made it pessimistic about the average user’s interest in looking beyond the quick answer.

As for the accuracy of the chocolate answer, Dr. Dariush Mozaffarian, a cardiologist and professor of medicine at Tufts University, said that it contains some facts that are mostly correct, and that it summarizes the research on the health benefits of chocolate. But he said he does not distinguish between strong evidence provided by randomized trials and weaker evidence from observational studies, or offer any caveats about the evidence.

Dr. Mozaffarian said: It is true that chocolate contains antioxidants. But the claim that chocolate consumption can help prevent memory loss? This has not been clearly proven, he said, and “needs a lot of warning.” Listing such claims next to each other gives the impression that some are better substantiated than they actually are.

Answers can also change as the AI ​​itself evolves, even if the science behind a particular answer doesn’t change.

A Google spokesperson said in a statement that the company has worked to display disclaimers on responses when needed, including notes that the information should not be treated as medical advice.

It’s not clear how exactly AI Overviews evaluates the strength of evidence, or whether it takes into account contradictory research findings, such as those regarding whether coffee is good for you. “Science is not a set of fixed facts,” Dr. Yasmin said. She and other experts also wondered whether the tool would rely on old scientific findings that have since been disproved, or if it didn’t capture the most recent understanding of an issue.

“Being able to make a critical decision — to differentiate the quality of sources — is what humans do all the time, and it’s what doctors do,” said Dr. Danielle Peterman, a physician and artificial intelligence scientist at Dana-Farber Cancer Institute and Brigham. And women’s hospital. “They analyze the evidence.”

If we want tools like AI Overviews to play this role, she said, “we need to better understand how to navigate through different sources and how to apply a critical lens to arrive at a summary.”

These unknowns are troubling, experts said, since the new system elevates the AI’s overview responsiveness compared to individual links to reputable medical websites such as those of the Mayo Clinic and the Cleveland Clinic. Such sites have historically risen to the top of many health searches.

A Google spokesperson said AI Overviews will match or summarize information that appears in the top search results, but is not designed to replace that content. Instead, the spokesperson said, they are designed to help people navigate the information available.

The Mayo Clinic declined to comment on the new responses. A representative from the Cleveland Clinic said people seeking health information should “search directly to known and trusted sources” and contact their health care provider if they are experiencing any symptoms.

A representative from Scripps Health, a California-based health care system cited in some AI Overview summaries, said in a statement that “citations in Google AI-generated responses can be beneficial because they make Scripps Health a trusted source of health information.” “.

However, the representative added: “We have concerns that we cannot vouch for AI-produced content in the same way that we can our own content, which is vetted by our medical professionals.”

Experts said that for medical questions, it is not only the accuracy of the answer that matters, but how it is presented to users. Take the question “Am I having a heart attack?” The AI ​​response had a useful summary of symptoms, said Dr. Richard Gumina, director of cardiovascular medicine at The Ohio State University Wexner Medical Center.

But he added that he had to read a long list of symptoms before the text message advised him to call 911. Dr. Gumina also searched for “Am I having a stroke?” To see if the tool might produce a more urgent response — which it did, asking users on the first line to call 911. He said he would immediately advise patients with symptoms of a heart attack or stroke to seek help.

Experts encouraged people searching for health information to treat new responses with caution. Essentially, they said, users should note the finer details under some AI overview answers: “This is for informational purposes only. For medical advice or diagnosis, consult a professional. Generative AI is experimental.”

Danny Bloom Contributed to reports.