Google AI Overview keeps hallucinating basic facts

Search for "how many rocks should I eat per day" and the overview cheerfully recommends at least one small rock daily — citing a satirical Onion article as its source. That went viral, but the deeper problem isn't funny edge cases. It's the quiet confident errors on everyday queries that don't make headlines.

Last week I searched "Python 3.12 dict union operator syntax" and the overview claimed |= mutates the left operand in-place while | returns a new dict. Half right. |= does mutate, but | also returns a new dict — the overview implied otherwise. A junior dev copying that snippet would introduce a subtle bug. Same session: "Linux kernel version in Ubuntu 24.04 LTS" — overview said 6.8. Actual shipping kernel: 6.8 for the HWE stack, 6.5 for GA. Missed the split entirely.

The pattern: it stitches together snippets from high-ranking pages without checking internal consistency. No reasoning step, just probabilistic glue. When sources contradict, it averages them into nonsense. When a source is satirical, outdated, or flat wrong, it treats it as gospel if the page ranks well.

Turning it off helps. Chrome: Settings → Search engine → Manage search engines → add a new one with https://www.google.com/search?q=%s&udm=14 as the URL (the udm=14 param forces classic web results). Firefox: same via about:preferences#search. Safari users need an extension like "StopTheMadness" or a custom search engine workaround. Mobile app has no toggle — switch to browser.

Alternative: Kagi (paid, no ads, lets you block domains, ranks by user signals), or DuckDuckGo's HTML lite for speed. Perplexity works for synthesis tasks if you verify citations — its inline links make that faster than Google's opaque overview.

The overview isn't "AI search." It's a summarizer bolted onto a ranking algorithm that optimizes for engagement, not truth. Treat it like a confident intern who occasionally lies: useful for orientation, never for production decisions.