Google has a misinformation problem it can’t seem to fix. And it’s baked right into the top of its search results.
The company’s AI Overviews feature — those AI-generated summaries that appear above traditional search links — continues to surface false, misleading, and fabricated claims to users who may never scroll past them. According to a report from Futurism, the feature has repeatedly produced answers that are flatly wrong, sometimes dangerously so, drawing on unreliable sources or misinterpreting reliable ones. The problem isn’t occasional. It’s structural.
Google launched AI Overviews broadly in May 2024, integrating generative AI directly into its flagship product: Search. The pitch was simple. Instead of clicking through multiple blue links, users would get a synthesized, authoritative-sounding answer right at the top of the page. For Google, it was a way to compete with ChatGPT and other AI assistants that had begun eating into its search dominance. For users, it was supposed to save time and deliver clarity.
What it has delivered, with uncomfortable regularity, is nonsense dressed up as fact.
The early blunders were almost comical. AI Overviews told users to put glue on pizza, suggested eating rocks for nutritional benefits, and claimed Barack Obama was the first Muslim president of the United States. Google quickly patched those specific errors and called them rare edge cases. But the underlying issue — that a probabilistic language model is being used to generate definitive-sounding answers to factual queries — hasn’t gone away. As Futurism documented, the feature continues to pull from satirical sites, Reddit threads, and low-quality web pages, presenting their content as though it were vetted information.
This matters enormously because of where AI Overviews sit in the information hierarchy. They occupy the most valuable real estate on the internet: the top of a Google search results page. Traditional organic results, the ones from actual publishers and institutions, get pushed below the fold. For the vast majority of casual searchers — people looking up a health symptom, a historical fact, a political claim — the AI Overview is the answer. They don’t dig further.
Google’s own data supports this behavioral pattern. The company has said AI Overviews increase user engagement with search and that people find them helpful. But engagement and accuracy are not the same thing. A confidently wrong answer that satisfies a user’s curiosity is arguably worse than no answer at all, because the user walks away misinformed and doesn’t know it.
The timing of these concerns couldn’t be more charged. We are in the middle of a period of intense anxiety about AI-generated misinformation, particularly around elections, public health, and climate science. Researchers at organizations like NewsGuard have been tracking how AI tools across the industry handle queries about contested or sensitive topics. Their findings are not encouraging. Large language models, including the ones powering Google’s search summaries, frequently generate responses that reflect the biases, errors, and fabrications present in their training data. The models don’t understand truth. They predict text.
Google has pushed back on the criticism, arguing that it has implemented safeguards to reduce the frequency of low-quality AI Overviews. The company says it has added restrictions for sensitive topics like health and news, and that it has improved the sourcing mechanisms that determine which web content the AI draws on. A Google spokesperson told multiple outlets earlier this year that the company takes the quality of AI Overviews “extremely seriously” and continuously updates its systems to reduce errors.
But the fixes have felt like whack-a-mole. Patch one category of bad output and another surfaces. The fundamental architecture of the system — a generative model summarizing web content it doesn’t truly comprehend — creates an irreducible error rate. And when that error rate is multiplied across billions of searches per day, even a small percentage translates into millions of encounters with false information.
Publishers and journalists have raised a separate but related concern. AI Overviews don’t just risk spreading misinformation — they also strip traffic from the original sources that produce accurate information. When Google synthesizes content from a news article or an academic paper and presents it as an AI-generated summary, users have little reason to click through to the source. This undermines the economic model that supports quality journalism and research. It’s a perverse dynamic: the better the original source material, the more useful the AI Overview, and the less likely the user is to visit the source.
Several major publishers have already taken legal or strategic action. The New York Times, for example, has been vocal about protecting its content from AI scraping. Other outlets have experimented with blocking AI crawlers entirely. But opting out of Google’s AI systems can mean opting out of Google’s search results altogether — a trade-off most publishers can’t afford to make.
The competitive pressure on Google to keep AI Overviews and expand them is immense. Microsoft’s Bing has integrated OpenAI’s technology. Perplexity AI has built an entire search product around AI-generated answers. Apple is weaving AI into its operating systems. Google can’t afford to pull back on AI in search without ceding ground to rivals. So the company is caught between two imperatives: move fast on AI integration, and don’t break the trust that makes Google the default information source for billions of people.
So far, speed is winning.
Recent reporting suggests Google is expanding AI Overviews to more query types and more countries, even as accuracy concerns persist. The company appears to be betting that iterative improvement — fixing errors as they’re reported, refining models over time — will eventually bring the feature’s reliability up to an acceptable standard. That’s a reasonable engineering philosophy. It’s a much riskier proposition when the product in question is the world’s primary gateway to information.
Critics argue that Google should, at minimum, label AI Overviews more prominently as AI-generated and potentially fallible. Right now, the summaries appear with a small “AI Overview” tag that many users likely overlook. A more visible disclaimer — something along the lines of “This answer was generated by AI and may contain errors” — could help set appropriate expectations. Google has resisted making the disclaimers more aggressive, presumably because doing so would undermine user confidence in the feature.
There’s a deeper philosophical tension here that extends well beyond Google. The entire AI industry is racing to deploy generative models in high-stakes information contexts — search, healthcare, legal research, education — before the reliability of those models justifies that deployment. The assumption is that the technology will improve fast enough to outrun the harms it causes in the interim. Maybe it will. But the interim is now, and the harms are real.
For Google specifically, the stakes are existential in a way they aren’t for smaller players. Google Search isn’t just a product. It’s infrastructure. Hundreds of millions of people treat it as a public utility, the default mechanism for answering questions about the world. When that mechanism starts producing wrong answers at scale, the consequences ripple outward — into classrooms, doctor’s offices, voting booths, dinner table arguments. The damage is diffuse and hard to measure, but it accumulates.
And there’s no easy fix. You can’t bolt accuracy onto a system that wasn’t designed to be accurate in the way humans understand accuracy. Large language models generate plausible text. Plausibility and truth overlap often enough to be useful, but they diverge often enough to be dangerous. Google knows this. Every AI researcher knows this. The question is whether the commercial incentives to deploy these systems will allow for the kind of caution the technology demands.
Right now, the answer appears to be no.
WebProNews is an iEntry Publication