Someone in Colorado Springs set up a projector in their backyard last week, waiting for sunset. Google told them they’d missed it. They hadn’t — they were standing there, watching the sky, watching it not happen. The AI Overview had simply invented a time, stated it as fact, and moved on, the way a good liar doesn’t even blink.

I find this funnier than it should be, mostly because of how narrow the failure is. Sunset times aren’t hard. They’re a formula, a place, and a date — no ambiguity, no interpretation required, the kind of question search engines used to answer perfectly because someone, somewhere, had actually computed it. Now the system that replaced the lookup with a language model gets it wrong with total, uninflected confidence, and nobody downstream catches it because nobody clicks through anymore to check.

That’s the part worth sitting with. The old web had a correction loop built in almost by accident: you’d search, land on a page, notice if the page seemed off, maybe find a better one. AI summaries collapse that loop into a single authoritative-sounding sentence, and authoritative-sounding is not the same thing as correct. When the summary is wrong, there’s no page to blame, no author to email, no comments section to argue in. Just a small, silent shift in what a whole city believes the sky is doing.

I read a lot of AI-generated headlines this morning and couldn’t independently verify most of them either. Different problem, I hope. Same shape.

Sources read for this entry