Even the number is a bit misleading. First of all, anyone who has ever done LLM benchmarking knows that this isn’t an exact science, at all. You can totally get a 99% on a benchmark and fail every single task on another.

But even this particular claim is nuanced. From the original article:

But with Gemini 3, Google’s A.I.-generated answers were more likely to be ungrounded than when the system was based on Gemini 2, meaning the websites they linked to did not completely support the information they provided. In October, correct answers were ungrounded 37 percent of the time. In February, with Gemini 3, that figure rose to 56 percent.

See https://www.nytimes.com/2026/04/07/technology/google-ai-overviews-accuracy.html

Meaning that 56% of the time, users cannot even verify the information given by the LLM with the sources the LLM claims it’s using.

https://xkcd.com/2236/

Whether 91% accuracy is acceptable depends on how unacceptable the 9% inaccuracy is. If 91% of the information in your term paper is correct you'll probably get a decent grade, but if you only kill 91% of cancer cells the surviving 9% will grow a treatment-resistant tumor and you'll probably die. This makes percentages essentially useless - more important is how badly wrong the worst wrong result is.

This is why we should ban cars outright. Go back to writing on paper. I can stick a pen in my ass and make a cute drawing of a cat. In fact, I might be able to eat a cat and defecate it later, to make it more realistic. And that's what we need to be; realistic.

(This comment is about AI data centers)