Comment by shadowgovt

Disney's had a notable amount of success with that formula. Turtle Talk with Crush arguably saved The Living Seas pavilion space at EPCOT, and it's executed via a digital puppet operated by a behind-the-scenes cast member doing their best Crush voice.

I sincerely doubt that what makes that experience magical can be replicated with AI in my lifetime. Too much contextual knowledge, too much detail in the nuances of human-human interaction, and too much je ne sais quoi in the timing of getting humor right. I've seen Turtle Talk deal with a particularly excited young person leaning on Crush's "tank" by having Crush look at him and go "Hey little minnow... One of the big humans behind you is gonna come scoop you up. I've seen it happen, lots of times!" You can program that interaction in, but the domain-space of having an interaction for every possible "improv moment" might be outside the bounds of what the next several generations of learning models are capable of.

... or I'm wrong, in which case I look forward to enjoying robo-Seinfeld in the retirement home.