COMIC: Agentic Sketch Comedy Generation

If you ask an AI to write a joke, you often get the digital equivalent of a cringeworthy “dad joke.” But what's surprising is that today, there is an AI system capable of creating comedy sketches that can almost keep up with the pros on Saturday Night Live. Its name is COMIC, and it operates like a real writers' room: there are AI agents who play the roles of screenwriter, director, critic, and even stage manager, each with a different personality, and all competing with one another. The idea is this: to generate true creativity—especially in a field as elusive and subjective as comedy—it is not enough to ask a model to guess the “right joke” based on fixed objectives. You need to put multiple artificial intelligences in a continuous competition, where ideas are refined through challenges, revisions, and opinions that often clash with one another. In essence, the quality of a joke is not determined by how closely it approximates a universal ideal, but by how well it fares against its rivals in a series of duels judged by critics with different tastes. Susung Hong, Brian Curless, Ira Kemelmacher-Shlizerman, and Steve Seitz led the development of COMIC, and the key detail is the use of critics: not just AI agents, but “critics” tuned to what people actually like, based on an analysis of nearly 5,000 YouTube comedy videos, from Key & Peele to Foil Arms & Hog. A quote from the paper itself sums it all up: “Rather than imposing a ground-truth quality ceiling, COMIC embraces relativism, where a script’s fitness is defined not by its distance from an ideal but by its relative performance against current competitors.” In short, for AI, comedy is not a peak to climb, but a continuous race, somewhat akin to the Red Queen theory in evolutionary biology: to stay in the game, you have to keep changing. COMIC’s structure is akin to an idea island: scripts are created on separate “islands,” each governed by a committee of critics with different tastes—some prefer slapstick, some like dry humor, and some favor the absurd. The scripts compete against each other in one-on-one tournaments: the loser is rewritten based on the critic's advice and then put back into play. This process, repeated over multiple generations, continually raises the bar and allows for the creation of diverse sketches, not just a bland version of “what works best on average.” Here’s a real-world example: using this method, COMIC generated videos starting from little more than a photo and a snippet of a character’s voice, building coherent stories and recognizable characters scene by scene, and maintaining continuity between one shot and the next—something that standard video models, such as Sora or Veo, still struggle to do. The results in the eyes of human viewers? In a blind test, COMIC videos received comedy and engagement scores comparable to mid-range professional productions on a scale of 1 to 7. Traditional models, on the other hand, while more visually realistic, fall short in terms of storytelling and the desire to “see more.” Here's a fun fact you can share at dinner: the system is so efficient that it can produce quality sketches in a single day with a $5 GPU budget—while a typical episode of a show costs orders of magnitude more. What’s more, the selection of critics is not decided manually: the AI generates hundreds of different critic “personas” and then keeps them only if they can truly distinguish between what people like and what doesn’t work, measuring their ability to predict which videos will generate the most engagement on YouTube. The most striking aspect is that quality is not imposed from above, but rather emerges from the ongoing dialogue between scripts and critics, who keep each other informed, just like a room full of human writers exchanging jokes, correcting each other, and challenging one another. One final thought: This methodology, with the competition between agents and the selection of critics aligned with real audience tastes, could be a way to teach AIs other forms of creativity as well, not just comedy. After all, humor is just the most challenging laboratory. The key takeaway is this: for AIs, comedy is not measured by a formula, but is achieved through constant challenges and revisions, just like humans do. If you think this idea could change the way you think about artificial creativity, you can click I'm In on Lara Notes – it's not a like; it's your way of saying: This perspective is now yours. And if, in a few days’ time, you find yourself telling someone that there is an AI that learns to be funnier by challenging other AIs, like in an SNL writers’ room, on Lara Notes you can tag the person who was with you: it’s called Shared Offline. This work comes from arxiv.org, and you’ve just saved about 23 minutes compared to reading the full article.