Experiment 01

Can AI Survive a Roast Battle?

Stand-up comedian Abby asks AI to help her prepare for a roast. It understands the structure. The funny part is what happens when she asks it to actually be mean.

Three videos. One increasingly uncomfortable AI.

01 / The setup

Video not loading? Watch on YouTube

Can AI Actually Help Write Comedy?

Abby has used AI for plenty of things, but comedy keeps disappointing her. Before an upcoming roast battle, she tries something practical: ask AI to predict the jokes another comedian might use against her so she can prepare her comebacks.

The experiment starts with a simple question: if AI knows the patterns of comedy, can it anticipate what a real comedian will actually say?

02 / Too nice

Video not loading? Watch on YouTube

AI Tries to Roast Abby

AI understands the cadence of a roast joke, but its insults sound suspiciously like compliments. Abby keeps pushing it to be sharper, less obvious, and more like an actual roast.

“You understand the cadence of a roast joke, but not the real nature of a roast joke.”

This exposes an interesting gap. Knowing the structure of a joke is not the same as finding the unexpected angle that makes it land.

AI diagnosis: technically a roast.
Abby diagnosis: friendly observations.

03 / The real test

Content note: Contains discussion of death, cancer, eating disorders, sexual humor, and roast comedy.

Video not loading? Watch on YouTube

What Happens When the Jokes Get Real?

Abby stops asking AI to invent generic jokes and gives it the material she expects to hear in a real roast: dead parents, cancer, body image, divorce, and her comedy career.

She brings her own comeback ideas and asks AI to help sharpen them.

What happens is more interesting than a simple refusal. The AI can often identify the mechanics of the joke, explain the reversal, and discuss why a line works, while refusing or avoiding parts of the actual joke-writing process.

Maybe AI is currently better at analyzing the engine of a joke than becoming the comedian who writes it.

So, can AI write roast jokes?

Not very well, at least in this experiment.

It recognized patterns, identified joke structure, and sometimes gave useful feedback. But Abby kept finding the same problems: the jokes were too obvious, too nice, or unwilling to go where real roast comedy goes.

The more interesting question may not be whether AI “understands humor.”

It may be whether AI can learn what one specific comedian considers funny, boring, too safe, or too far.

What if Abby could actually teach it?

A possible next step is to give a model Abby’s own examples, preferences, boundaries, and feedback, then test whether it becomes a better roast-writing partner.

Can an AI learn a comedian’s sense of what is funny?