As for the accuracy of the chocolate answer, Dr Dariush Mozaffarian, a cardiologist and professor of medicine at Tufts University, said that it had some facts mostly correct and that it summarised research into chocolate’s health benefits. But it does not distinguish between strong evidence provided by randomised trials and weaker evidence from observational studies, he said, or provide any caveats on the evidence.
It’s true that chocolate contains antioxidants, Dr Mozaffarian said. But the claim that chocolate consumption could help prevent memory loss? That hasn’t been clearly proved, and “needs a lot of caveats,” he said. Having such claims listed next to one another gives the impression that some are better established than they really are.
The answers can also change as the AI itself evolves, even when the science behind a given answer hasn’t changed.
A Google spokesperson said in a statement that the company worked to show disclaimers on responses where they were needed, including notes that the information that shouldn’t be treated as medical advice.
It’s not clear how, exactly, AI Overviews evaluate the strength of evidence, or whether it takes into account contradictory research findings, like those on whether coffee is good for you. “Science isn’t a bunch of static facts,” Dr Yasmin said. She and other experts also questioned whether the tool would draw on older scientific findings that have since been disproved or don’t capture the latest understanding of an issue.
“Being able to make a critical decision – to discriminate between quality of sources – that’s what humans do all the time, what clinicians do,” said Dr Danielle Bitterman, a physician-scientist in artificial intelligence at Dana-Farber Cancer Institute and Brigham and Women’s Hospital. “They are parsing the evidence.”
If we want tools like AI Overviews to play that role, she said, “we need to better understand how they would navigate across different sources and how they apply a critical lens to arrive at a summary,” she said.
