English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

When AI Learns to Tell Jokes: Inside the University of Washington's COMIC Sketch Comedy System

Forum topic · 小凯 · 2026-03-12

Summary

This explainer introduces COMIC (Agentic Sketch Comedy Generation), a system by Susung Hong, Brian Curless, Ira Kemelmacher-Shlizerman, and Steve Seitz that lets AI write and direct complete short comedy videos. Humor is hard for AI because it demands timing, expectation, surprise, and cultural context—capabilities that traditional text-prediction LLMs lack. COMIC tackles this with a multi-agent architecture: writer, director, actor, and judge agents that debate, iterate, and refine scripts. It applies evolutionary selection over large pools of creative candidates, keeping the funniest ideas. Its key innovation is an LLM judge calibrated to human preferences by learning from YouTube sketch comedy videos and real audience signals such as views, likes, and completion rates. Combined with state-of-the-art video generation, character reference images for consistency, and lip-synced speech, COMIC reportedly produces sketch videos rated by human evaluators as approaching professional quality and achieving state-of-the-art video generation performance. The article also explores implications: personalized comedy, democratized content creation, multi-agent collaboration as a paradigm, and AI as a mirror of human creativity rather than its replacement.

When AI Learns to Tell Jokes: An Adventure in Machine Humor

How the University of Washington's COMIC system turns AI into a "comedy director"

The crazy scientific question

Imagine watching classic Saturday Night Live sketches late at night and wondering: what kind of jokes would AI write for this? Humor has long been considered one of humanity's last strongholds—we expect AI to explain "why did the chicken cross the road" as "a statistical analysis of avian behavior," not to actually be funny.

But a team of four University of Washington researchers—Susung Hong, Brian Curless, Ira Kemelmacher-Shlizerman, and Steve Seitz—took on this seemingly impossible task. Their COMIC system (Agentic Sketch Comedy Generation) doesn't just write comedy scripts; it "directs" complete short comedy videos.

Chapter 1: Why is humor so hard for AI?

Humor: humanity's most mysterious cognitive ability

When you laugh at a sketch, your brain: understands the literal dialogue, catches puns and subtext, predicts the story's direction, encounters cognitive conflict when the ending defies expectations, and resolves that conflict pleasurably—that's the punchline. Humor is the product of high-level cognition: language, timing, expectation, surprise, cultural context.

The AI "humor dilemma"

Traditional LLMs don't truly understand humor—they are super-powered text predictors. Asked to tell a joke, they perform fill-in-the-blank based on patterns, producing text that *looks* like a joke but often lacks a real comedic core. Humor is also deeply subjective and culturally dependent, making it an enduring challenge.

Video comedy: nightmare mode

A successful comedy short requires a solid script, precise performance timing, camera language, sound design, and character consistency across scenes—normally a whole professional team's work. The UW researchers propose: let AI do all of it.

Chapter 2: COMIC—AI's "comedy studio"

Multi-agent collaboration

COMIC's core innovation is multi-agent collaboration, mirroring a real comedy production company:

  • Writer agent: creates scripts, dialogue, and plots
  • Director agent: controls pacing, shots, and scene transitions
  • Actor agent: generates character performances and actions
  • Judge agent: evaluates whether the work is funny
  • These agents argue, propose alternatives, critique each other, and iterate—like a human creative team.

    Competition and evolution

    COMIC goes further with evolutionary algorithms: it generates a large population of creative candidates simultaneously. The judge agent filters them—clichéd gags eliminated, good twists kept, slow pacing improved. Through rounds of competition, selection, and recombination, only the most "fit" (funniest) ideas survive.

    Iterative refinement

    From basic story framework → stronger opening hook → tighter comedic rhythm → more impactful ending twist: each iteration is targeted refinement driven by the judge's feedback, like a human screenwriter's rewrites.

    Chapter 3: Teaching AI to "read" the audience

    The UW team's clever move

    The most critical component is the judge agent. Rather than letting AI guess what's funny, the team:

    1. Collected large amounts of YouTube sketch comedy videos 2. Extracted features: script structure, dialogue patterns, performance style, pacing 3. Observed real audience signals: views, likes, comment sentiment, completion rates 4. Trained a dedicated LLM judge on this data

    The judge learns not a theoretical definition of humor, but which feature combinations make real audiences laugh—like training a food critic by tasting everything and learning what people actually enjoy.

    Aligning with human preferences

    This is called alignment: calibrating the judge agent's "sense of humor" to match real viewers. Without it, COMIC might produce content that is fluent and well-structured but not funny to humans. With it, the judge acts like an experienced comedy producer who knows what pacing holds attention, which twists surprise, and which characters resonate.

    Chapter 4: From script to video—full pipeline production

    COMIC's ambition extends beyond scripts to complete comedy videos:

  • Video generation: leverages cutting-edge text-to-video models (e.g., OpenAI's Sora, Runway Gen-4, Kuaishou's Kling AI)
  • Character consistency: maintains character reference images so the same character looks the same across scenes, avoiding early AI video's amnesia problem
  • Narrative coherence: the director agent plans storyboards; each scene references neighboring scenes; transitions are checked for natural flow
  • Voice and performance: integrates speech synthesis with lip sync, so characters genuinely appear to speak
  • Chapter 5: Does it work? The data

  • COMIC's short comedy videos approach professional production quality—mixed with real SNL clips, viewers may not clearly distinguish which is AI-made
  • The system achieves state-of-the-art performance in video generation
  • Evaluations were validated by human judgments, not AI self-scoring
  • Chapter 6: What does this mean?

  • Creative automation: humor, long seen as humanity's exclusive territory, is now within AI's reach via careful system design
  • Multi-agent paradigm: a shift from one-person-one-AI to teams of specialized agents collaborating—applicable to novels, architecture, scientific research
  • Alignment techniques: training judges on real human preference data (like YouTube reactions) can extend to music, design, and decision-making
  • Chapter 7: Future outlook

  • Personalized comedy: "Give me a 5-minute sketch about office politics, in the style of The Office"—ten minutes later, your custom comedy plays
  • Democratized creation: anyone with an idea can produce professional-quality content
  • Human-AI collaboration: like calculators and Photoshop, COMIC augments rather than replaces creators—humans supply vision and emotional core; AI handles execution

Epilogue: When machines learn humor, what remains uniquely human?

COMIC succeeds not by inventing humor from scratch, but by learning human humor—from YouTube videos, audience reactions, and pop culture, all human creations. AI is a mirror reflecting human light. Perhaps when AI finally tells a joke, we should feel a strange pride: *the machines we built finally understand part of us.*

Key references

1. Hong, S., Curless, B., Kemelmacher-Shlizerman, I., & Seitz, S. (2026). COMIC: Agentic Sketch Comedy Generation. *arXiv preprint arXiv:2603.11048*. 2. OpenAI. (2025). Sora: A generative video model. *OpenAI Technical Report*. 3. Qian, C., et al. (2024). ChatDev: Communicative Agents for Software Development. *ACL*. 4. Chan, C. M., et al. (2023). ChatEval: Towards better LLM-based evaluators through multi-agent debate. *NeurIPS*. 5. Fernando, C., et al. (2023). Promptbreeder: Self-referential self-improvement via prompt evolution. *ICLR*.

*This article explains cutting-edge research in accessible language. Corrections are welcome.*

Tags

#artificial-intelligence#humor-generation#multi-agent-systems#video-generation#llm-alignment#sketch-comedy#university-of-washington#generative-ai

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177168810