My big question is whether that sort of deranged AI-output look will work when it arrives. It does for Youtube videos, but they're both aimed at children and searching for maximum intensity (via both character recognition and zaniness) to hold attention.

As is, I strongly suspect we're already over-testing on frontends.

First, because frequent change has a cost in user retention, and so does instability. I get A/B tested enough that I notice it happening, or notice that instructions I give for a program don't reflect what other people on the same device are seeing. And since I use multiple Gmail accounts, I've also had the experience of different accounts falling into different branches of a significant UI test - it was a horrible and aggravating experience.

Second, because isolated tests are a terrible way of judging aggregate effects. They're not necessarily run with an eye on "irrelevant" stats, and they're almost never run against a clean, single-design baseline. (Also, they're frequently run with grotesque statistical errors like stopping tests at significance...) The risk, then, is ending up with a hodge-podge of features that produced 3% improvement on some metric but did nothing for overall value. Newsletter signups designed to improve subscriptions, which destroyed retention. Buttons locked into arms races where they improve clickthrough by cannibalizing one another's traffic. Evolved-in-isolation pages that get good responses but decrease brand recognition, and so on.

None of that means aimless AI testing won't happen; the combination of buzzword appeal and hands-off profit is probably too good to pass up. But I expect it won't be as valuable as people expect because it feeds into all the mistakes we're making already.