HandsOnAzureRU

When Does AI Crack the First Joke?

A model can be quick to pick up a joke. A hint about something we discussed a little earlier is often enough. How often has one volunteered a funny line in the middle of a serious conversation, though? I have a harder time remembering those occasions.

There’s a wrinkle in the question. In an ordinary chat the user gets the first turn. Within its reply the model can still crack a joke without being asked. That’s already a choice of timing, even within narrow limits. Sending you a funny thought an hour after the conversation ended is a different technical problem.

I’d try a short exchange with an opening where a joke might land. Show some readers the whole conversation and others just the last line. Did the line amuse them on its own, or did it call back to the start of the conversation? To test initiative, the agent also needs the option to keep quiet. Otherwise even a very funny line is the product of a mandatory exercise.

Poems and drawings raise a similar question. You can judge the finished piece, read it to someone else, and come back to it a week later. It tells you much less about why its author sat down to work that particular day. Sometimes a wonderful poem turns out to have been commissioned. That doesn’t make it a worse poem.