Google Search is moving further from lists of links toward direct AI-generated responses. That shift is especially sensitive when the question involves health, because a bad answer can move quickly from embarrassing to harmful.
Recent examples of Google AI Overviews show the problem clearly. The system has reportedly produced strange, incorrect, or poorly sourced answers, including responses on kidney stones, stem cell therapy, and pregnancy-related questions.
Why AI Overviews Change the Search Experience
At Google I/O, Google announced a major rollout of AI Overviews. The idea is simple: instead of only sending people to websites, Search can summarize information and present a direct answer at the top of the results.
That makes the answer feel more authoritative. A link invites the reader to evaluate a source. An AI Overview can look like Google itself has resolved the question.
The source article says Google appears to combine information from several places to generate these overviews. That process can create mistakes, especially when sentences are removed from their original context or when weak sources are treated as useful evidence.
The risk is not limited to obscure trivia. Google said in 2019 that about 7% of searches were health-related, including symptoms, medications, insurance, and similar topics. That means even a small failure rate can matter when users are asking about their bodies, treatments, or urgent decisions.
The Medical Examples Are the Real Warning Sign
One analysis described in the source found that out of 25 health terms, 13 had an AI result. The same article says the rate was lower for money-related questions.
The reported health examples are troubling because they mix correct framing with dangerous or unsupported claims. In one kidney stone answer, Google’s AI said drinking fluids such as water, soda, and juice can help. Then it added: "You should aim to drink at least 2 quarts (2 liters) of urine every 24 hours."
The source notes that this is dangerous because the high salt in urine can dehydrate someone and disrupt electrolytes. Google removed that kidney stone answer after it spread widely, but the same advice reportedly still appeared for a similar query: "how to pass kidney stones fast".
Another example involved stem cell therapy for knee problems. Biology professor Paul Knoepfler said the AI Overview "is like an ad for unproven stem cell clinics." The source says Google cited dubious clinics as a main source, and Knoepfler noted that "there is no good evidence that stem cells help knees."
A third example concerned pregnancy. Google’s AI correctly said pregnant women should not engage in sumo wrestling, but then added that sumo was safer than shooting guns while pregnant. The issue is not only that the answer sounds odd. It shows how a system can produce a comparison that may not help a person make a real health decision.
Bad Sources Can Become Polished Answers
The problem is easier to understand through non-medical examples. A Reddit user jokingly suggested mixing cheese with glue to stop it from sliding off pizza 11 years ago. Google’s AI Overview repeated the idea, while specifying that the glue should be non-toxic.
Google removed that specific overview for "cheese not sticking to pizza." But the source says the sticky cheese suggestion appeared again when the query was changed slightly. That matters because users do not all ask the same question in the same words.
Another example asked how many rocks someone should eat. The AI reportedly answered "at least one small rock a day" rather than making clear that eating rocks is a bad idea. The source says Google’s AI appears to have relied on a satirical article from The Onion.
These examples show a core weakness of AI search: the system can transform low-quality, joking, or satirical material into a clean answer format. When that happens inside a search product used by billions of people, the presentation can make weak information look settled.
Google Says It Is Taking Action
According to the source, a Google spokesperson told The Verge that Google is "taking swift action" to remove inappropriate AI overviews and is using them as examples to improve the overall system.
That response addresses individual failures, but it does not fully answer the bigger question. If AI Overviews are generated from web content, Google must decide which topics are safe for direct answers, which sources are trustworthy, and how to handle cases where the system is confident but wrong.
The source also says Google had not confirmed or denied the search results at that point. It had indicated only that the examples were unusual search queries. The article adds that it is unknown if and how Google fact-checks its AI answers.
That uncertainty matters most in health search. A user reading about pizza glue may recognize absurdity. A user searching for kidney stone advice, stem cell therapy, or pregnancy-related guidance may be looking for help under stress.
The Accountability Question Is Getting Harder
AI Overviews also raise a publishing problem. Traditionally, publishers and authors are responsible for what they publish. Google has presented itself as a platform rather than a media company.
Direct AI answers blur that position. If Google’s system takes web content, changes it, and presents it as an answer, the result is no longer just a snippet pointing to a source. The source article argues that this has legal and business implications for the content ecosystem.
Reddit is part of that debate. The source says Google’s use of Reddit in AI answers is intentional and notes that the companies recently signed a $60 million dollar deal giving Google more access to Reddit data for AI usage.
The concern is that crowdsourced posts can feed AI answers while publishers receive less benefit. The source also says Google has long argued with publishers over free snippets in search results, but snippets at least still send traffic to websites. Generative AI summaries go further by using more of the underlying content while potentially reducing the need to click.
For health information, the immediate issue is accuracy. For the web, the larger issue is responsibility. Google AI Overviews may make Search faster, but the reported examples show that speed is not enough when the answer concerns someone’s health.