The way LLMs work is by approaching the most "average" response given any particular input. It's why everything written by an LLM looks similar and always has the same voice.

Anyways, shockingly, the Machine That Generates the Average Output is bad at unique passwords.

Of the 50 returned, only 30 were unique (20 duplicates, 18 of which were the exact same string), and the vast majority started and ended with the same characters.

Imagine that an LLM tries to fit its outputs into a bell curve of potential responses, with each character in the output aimed to be as close to the middle as feasible (with a small randomization factor so it's not always the exact same). A good password's bell curve ought to be a completely flat graph where any character is just as likely to be chosen as any other character.

Use a password manager.