When I've tried to use it myself for verifiable things (i.e. code basically) I've had the 'confidently wrong' experience. When I've seen colleagues visibly use it (i.e. mentioned in commit messages) that confidence has rubbed off on them. Given that, why would I believe it when asking something medical, legal, historical, or otherwise outside of my domain or that I can't somehow verify?
You should be able to verify facts outside of your domain of expertise. Every time a new study comes around saying “actually, everyone should be drinking 3 glasses of wine a night”, do you take a trip to the liquor store? LLMs are far from the only thing you rely on that will confidently lie to you.
But with ChatGPT we're not saying disregard what it says, we're saying only disregard some of what it says and don't disregard what it says in future. Which becomes a lot of work to check everything it says every time.