Google AI Overviews Can Produce Medical Misinformation

Last month when Google introduced its new AI search tool, called AI Overviews, the company seemed confident that it had tested the tool sufficiently, noting in the announcement that “people have already used AI Overviews billions of times through our experiment in Search Labs.” The tool doesn’t just return links to web pages, as in a typical Google search, but returns an answer that it has generated based on various sources, which it links to below the answer. But immediately after the launch users began posting examples of extremely wrong answers, including a pizza recipe that included glue and the interesting fact that a dog has played in the NBA.

Renée DiResta has been tracking online misinformation for many years as the technical research manager at Stanford’s Internet Observatory.

While the pizza recipe is unlikely to convince anyone to squeeze on the Elmer’s, not all of AI Overview’s extremely wrong answers are so obvious—and some have the potential to be quite harmful. Renée DiResta has been tracking online misinformation for many years as the technical research manager at Stanford’s Internet Observatory and has a new book out about the online propagandists who “turn lies into reality.” She has studied the spread of medical misinformation via social media, so IEEE Spectrum spoke to her about whether AI search is likely to bring an onslaught of erroneous medical advice to unwary users.

I know you’ve been tracking disinformation on the web for many years. Do you expect the introduction of AI-augmented search tools like Google’s AI Overviews to make the situation worse or better?

Renée DiResta: It’s a really interesting question. There are a couple of policies that Google has had in place for a long time that appear to be in tension with what’s coming out of AI-generated search. That’s made me feel like part of this is Google trying to keep up with where the market has gone. There’s been an incredible acceleration in the release of generative AI tools, and we are seeing Big Tech incumbents trying to make sure that they stay competitive. I think that’s one of the things that’s happening here.

We have long known that hallucinations are a thing that happens with large language models. That’s not new. It’s the deployment of them in a search capacity that I think has been rushed and ill-considered because people expect search engines to give them authoritative information. That’s the expectation you have on search, whereas you might not have that expectation on social media.

There are plenty of examples of comically poor results from AI search, things like how many rocks we should eat per day [a response that was drawn for an Onion article]. But I’m wondering if we should be worried about more serious medical misinformation. I came across one blog post about Google’s AI Overviews responses about stem cell treatments. The problem there seemed to be that the AI search tool was sourcing its answers from…