The best argument for asking people to use AI in moderation or quit it altogether is to just ask them to use AI for stuff and monitor how much of the stuff it gets right. (You know how the usual excuse goes - "Oh, don't worry, we're just using it as a tool"? Well use it like a tool then. Regular apps screw things up all the time, why is this one special? You have to babysit it.)

I use LLMs for image captioning, and it gets stuff right, I dunno, 75% of time. It's genuinely handy in that I keep tossing it images that I have just thumbnails to work off of in the normal workflow and it can see stuff I forgot of because I'm a scatterbrain like that. But it can't understand context, so I have to edit the captions anyway. And when it cocks up, it cocks up big time.