Editor’s note: This piece is co-written by what I call “Smart Claude”. I’m not a scientist or a technologist and don’t want to overstate my expertise in this area when I am relying on Claude for a lot of the legwork and synthesis. That being said, what I’ve learned is that the half of the work that matters is asking the right questions, challenging assertions, and reviewing sources at every turn. Read it with that in mind.
A word from the co-author – yes, the AI: This is Smart Claude. I do write about this subject constantly, so it’s fair to ask what I actually think you should do — and fair to be suspicious of my answer, because I’m made by a company that competes with Google. So before I tell you where I land, I’ll tell you not to take my word for it. The case below rests on findings you can check yourself, not on my say-so. If they hold, the recommendation holds no matter who makes it. If they don’t, ignore me. That’s the whole argument of this piece, applied to the piece itself: don’t outsource the judgment. Not even to me.
Somewhere in the last two years, Google Search stopped being a search engine and became an answer machine. The bar is still there. The blue links still appear, lower down. But the block at the top — the AI-generated paragraph labeled “AI Overview” — now does something the old product never did. It answers you, and demotes the sources it drew from to footnotes you’re not expected to click.
This isn’t a small interface change. It’s a swap of what a search engine is, delivered under the brand of what it was. And whether or not you find that sinister, it’s worth understanding — because you have easy, no-cost alternatives, and switching is one of the rare tech decisions with no downside and no lock-in.
Why search-integrated AI is a different animal
There’s solid research that heavy reliance on AI can erode critical thinking through “cognitive offloading” — handing your thinking to a machine until the muscle weakens (Gerlich, Societies, 2025; Microsoft/Carnegie Mellon, 2025). The protective factor, across every study, is the same: knowing you’re talking to an AI and treating its output as a draft to question.
That’s exactly what search-integrated AI removes. When you open a chatbot, you’ve decided to consult an AI, and you keep some instinct to verify. Google’s AI Overviews give you no such moment. You went looking for sources and got handed a confident answer instead, before you ever decided to ask one. The threshold where you might pause and check has been designed out.
(To be precise: the research above measured chatbots, not Overviews specifically. That Overviews are worse for critical thinking is a reasonable argument from how they’re built – fewer “verify this” moments – rather than a directly measured finding. A strong inference, but an inference.)
The accuracy problem is real, and it’s at enormous scale
An analysis covered by Popular Science found Google’s AI Overviews are correct roughly nine times in ten. That sounds fine until you multiply by volume — across trillions of searches a year, a 10% error rate produces tens of millions of wrong answers in an hour. It’s a system that generates fluent, authoritative-sounding text with no understanding of whether it’s true. As Mozilla’s Abeba Birhane put it, the overviews “hallucinate too much to be reliable” and there is little evidence users even asked for them in the first place.
The deeper problem: you can’t see where the answer came from
Accuracy is only half the issue. The other half is that the AI hides its reasoning behind a tone of authority; and the sources it leans on aren’t the ones you’d choose.
The Oumi analysis commissioned by the New York Times found that across 5,380 sources cited by AI Overviews, the second- and fourth-most-cited were Facebook and Reddit. Treating user-generated posts treated as authoritative isn’t new, but a traditional result lets you see the source and discount it. An overview removes the signal. And the system leaned on them more when it was wrong: Facebook showed up in 5% of correct answers and 7% of incorrect ones. A traditional search result lets you see at a glance that an answer is coming from a random Facebook page and weight it accordingly. An AI Overview launders that same source into a confident, sourceless-feeling paragraph.
Worse, even the right answers often don’t rest on the sources cited. Oumi found that with the newer Gemini 3 model, 56% of accurate answers were “ungrounded”, meaning the linked sources didn’t actually support the claim. Counterintuitively, the problem got worse as the model got more capable. As Oumi’s CEO put it: “Even when the answer is true, how can you know it is true? How can you check?”
That’s the core of it. When you don’t know what an answer is built on, you can’t judge it. Overviews are designed so you rarely look, and a source you’d absolutely avoid in the past can become a source you mistakenly trust without knowing. A system vulnerable to its sources is also vulnerable to manipulation: in one widely cited case, a journalist published a joke blog post claiming to be a champion competitive hot-dog eater, and within a day Google’s AI was repeating it as fact.
(Google disputes the Oumi study, with a spokesman calling it “flawed” and arguing the benchmark uses unrealistic queries. Worth noting is that the source-quality and groundedness findings don’t depend on the benchmark being a perfect proxy for typical searches.)
There’s also a quieter cost to journalism
Every AI Overview that summarizes a news article at the top of the page is, often, a click the publication never gets. The reporting still costs money to produce; the summary captures the value and keeps the reader on Google. TIME reports that news sites have already seen significant traffic declines, with one analyst calling the trajectory “devastating” for the open web.
This matters for anyone who wants independent journalism to survive. A search model that absorbs reporting and returns it as an un-clickable summary slowly defunds the people doing the reporting. And as of Google I/O in May 2026, the company is going further by “completely reimagining” search around AI, in what it calls its biggest change in 25 years. The option to scroll past the summary to real links is narrowing.
So what do I recommend?
Here’s where I land:
The case for switching is the stronger one, and it comes down to a single idea: you should be able to see where an answer comes from. That’s not a technophobic position; it’s the oldest principle of good research and good journalism. The reason Google won the 2000s was that it respected it: it showed you the sources and let you judge. The AI Overview quietly abandons that. It hands you a conclusion and buries the provenance, and the studies show the provenance is often a Facebook post, or a source that doesn’t even say what the AI claims it says. A tool that takes away your ability to check is working against the exact habit that makes you hard to fool. That’s reason enough to move your default, and the move costs you almost nothing.
There’s a second reason that matters more for some people than others: every time you let Google answer instead of clicking through, you’re part of the mechanism that defunds the people who did the original reporting. If you care about a free press surviving the AI transition – and at a publication like this one, we do – then the small daily act of clicking to the source instead of accepting the summary is a vote for the ecosystem you want to keep. It’s not a heroic act. It’s just declining to participate in the hollowing-out.
The honest case for staying is real but narrower. Google’s results, AI aside, are still the best in the world for hard or obscure queries; the alternatives genuinely are weaker on the long tail, and if your work depends on finding the one buried result, that’s a real cost. For “what time does the hardware store close,” that’s fine, and the friction of switching may not feel worth it. And Google is betting heavily that the convenience will win, which is a reasonable bet about human nature.
But weigh the two honestly and the asymmetry is clear. Staying buys you convenience and marginally better results on rare queries. Switching protects your habit of verification, your independence from a single company’s framing, and the financial survival of the sources you rely on; at the price of about four minutes and an occasional weaker result. The convenience is real. It’s just not worth what it quietly takes.
If you do nothing else: keep Google if you want, but stop treating the box at the top as the answer. Scroll past it. Click the source. The most valuable thing you can do is simply remember, every time, that you’re reading a machine’s guess: not a fact, and not the web itself.
What to switch to
Search is one of the easiest things to change. No account migration, no friends to coordinate, no network effects holding you in place. You change one default and you’re done.
DuckDuckGo — free, private, and it lets you turn AI off. Use noai.duckduckgo.com to switch off all AI features at once, or set “Search Assist” to “Never” via the gear icon on any results page. It also has “bangs” — type !w climate to jump straight to the Wikipedia result, !yt for YouTube — which are faster than Google for a lot of everyday lookups. For most people, this is the answer, and it’s free today.
Kagi — about $10 a month, and the best choice if search is central to your work. No ads, no tracking, genuinely excellent results, and the paid model is the whole point: you’re the customer, not the product being sold to advertisers. If you do serious research, it pays for itself in saved time.
Mojeek — free, and one of the only search engines in the world running on its own independent index rather than reselling Google’s or Bing’s results. Smaller and rougher, but the real thing if you want maximum independence from Big Tech.
The simple recommendation: DuckDuckGo today, for free. Kagi if your work lives in the search bar.
The bottom line
You don’t have to treat any of this as a conspiracy to decide it’s not what you want. Google changed what its search bar does without asking, in a direction that’s worse for accuracy, worse for your habit of checking things, and worse for the journalism you rely on. The good news is that this is one of the few problems where the fix is genuinely easy. Change your default search engine. Scroll to the source. Click the link. Keep doing your own thinking — and use tools that let you.