One in ten. That's how often Google's AI Overview produces a wrong or poorly sourced answer, according to a study by Oumi, an open-source AI company. The percentage sounds manageable. At over 5 trillion searches per year, a 10% error rate translates to hundreds of thousands of wrong answers per minute.
That's Oumi's central argument: the accuracy number isn't the problem. The volume is.
Three Ways It Goes Wrong
Oumi identified three main failure patterns.
Source quality. Facebook and Reddit ranked as the second and fourth most-cited references in AI Overview responses. Inaccurate responses cited Facebook 7% of the time versus 5% for accurate ones - a modest gap that shows the system leans on user-generated content with no editorial standard and treats it with roughly the same weight as authoritative sources.
Source misrepresentation. Even when AI Overview cites legitimate sources, it sometimes summarizes them inaccurately. The source is real; the description of what it says isn't.
Deliberate manipulation. The most structurally concerning finding: someone can write a misleading blog post, drive artificial traffic to it, and cause AI Overview to pick up and amplify the false content. This isn't a bug that can be patched in an update. It's a consequence of a system that treats popularity as a proxy for accuracy. Bad actors who understand this have a reproducible method for injecting misinformation into Google's top search result.
The Position Problem
AI Overview doesn't appear somewhere on the results page - it sits at the top, above all organic links, in a visually distinct format with Google's branding. That placement creates an implicit credibility signal. Users reasonably assume that the result appearing first in its own box has been vetted more carefully than a random website ranking tenth.
The study notes this position amplifies harm: users are more likely to act on information that appears to come from an authoritative synthesis rather than a single source they could evaluate themselves.
The researchers also found that AI Overview sometimes produces different answers to the same query on repeat attempts - correct one time, wrong the next. That inconsistency makes systematic quality testing difficult and makes the product unpredictable for anyone depending on it for consistent information.
Google handles enormous volumes of queries about medical symptoms, legal questions, financial decisions, and factual research. An error rate in the hundreds of thousands per minute isn't an abstract product quality metric - it's a count of people getting wrong information from the thing designed to look like the most trustworthy result on the page.