Comment by Nevermark

No they are not. I think humor needs to be trained for via some form of indirection of reinforcement.

And certainly not by generalizing/interpolating examples, since telling jokes accumulated by exposure to examples would be the antithesis of a comedian's process.

Models and humans are very bad at extrapolation beyond the training set/experience (vs. interpolation at which we are both more likely to excel). But good humor is extrapolation. It breaks ground somehow, or it is an already dead "joke".

Likewise, training a model to be creative by training it on past creative artifacts is going to have the opposite effect. Creativity doesn't reproduce past creativity.