Search used to hand you a list of links and let you decide what to trust. Now it hands you an answer, written in confident, finished sounding sentences, and expects you to just accept it. A dean of libraries at the University of Virginia recently proposed a simple way to think about this: sort every AI answer into one of four buckets before you decide how much to trust it.
Some answers are facts you can check against a source, like a date or a chemical symbol. Some are interpretations built from evidence where reasonable people could read the same information differently, like whether remote work helps or hurts productivity. Some are things the AI creates for you, like a draft email, where there is no single correct version. And some are advice about what to actually do, like whether to take a daily aspirin, where the right answer depends entirely on your personal situation. The problem is that Google's AI answers move between these four types in the same paragraph, in the same even tone, so it is easy to treat a piece of advice as if it were a checkable fact.
This matters more than it sounds like, because of two things happening at the same time. First, a study commissioned by The New York Times found that Google's AI Overviews get it right about nine times out of ten. That means one out of every ten AI answers contains a real error, and Google's search engine handles an enormous volume of queries every single day, so the raw number of wrong answers being served is large even if the percentage looks small.
Second, a court in Munich just ruled that Google can be held legally responsible when its AI Overview makes a false statement about a person or a business. Two publishers sued after the AI summary linked them to fraud schemes they had nothing to do with, and the court decided this was not the same as linking to someone else's bad article. Google wrote those words itself, so Google owns them. Google is appealing, but the reasoning could eventually apply to any company running an AI system that answers questions in its own voice.
Here is the part that should worry any business owner. People are not checking these answers before acting on them. Research from Pew found that when an AI summary shows up, barely anyone clicks through to see where the information actually came from. That means if an AI answer misrepresents your company, your prices, or your safety record, very few people will ever discover the mistake on their own.
The practical response is not complicated. Treat factual answers as a starting point to verify, not a finished product. Treat advice about money, health, or legal decisions with real suspicion until a qualified person confirms it. And if you run a business, search your own company's name occasionally and read what the AI says about you, because right now almost nobody else will bother to check it for you.