lol - Some AIs' web sites pop an interstitial if you click on a link they provide you, like "You're leaving Claude to visit this external link" which is funny because it implies you need to be warned, i.e. you're crossing a boundary from a trusted context to an untrustworthy one, watch out.

Of course it's exactly the opposite. I had to specifically prod the AI to give me a link at all, so I could verify the dubious claims of the cognitive bias reinforcement engine. Sheesh.

I'm sure I've seen this many times before, but I'm fuzzy on where --- the main impression is this was already a howler when early corporate intranets were doing it in the early 2000s.

I want to say that, then as now, they almost always had the trust gradient backwards.