Why did the AI agent cross the road

Workshopping, an iterative process where creators share ideas, test what works, and refine what doesn't through collective feedback, is at the heart of any writers' group. This collaborative dynamic inspired George Mason University PhD students Shiwei Hong to explore whether artificial intelligence (AI) could benefit from a similar approach.

Hong developed a novel multi-agent system that simulates a collaborative comedy writing environment in which AI agents generate, critique, and refine humorous content together. Her research, recently accepted to the prestigious 64th Annual Meeting of the Association for Computational Linguistics, investigates how structured interaction among AI agents can improve creative output.

Humor remains challenging for artificial intelligence. "Generating a joke or something that is very funny is actually a very tricky question, because it requires a lot of how humans interpret the generated piece," said Hong, who is working on a degree in computer science.

She created a "sandbox" environment to mirror real-world creative collaboration, populated by 35 AI agents assigned distinct roles and personalities powered by GPT-4.

"Five agents acted as performers. The rest acted as critics or audience members," said Hong.

Designing 35 unique agent personas required imagination as well as technical expertise. Drawing on her interest in writing fiction, Hong created characters with varied social and cultural backgrounds, adding depth to the simulated interactions.

In one experimental environment, the performer agents generated two scripts in sequence without feedback; in another, they engaged the audience members in critique and discussion between the rounds of generation. The difference was significant.

Human evaluators compared outputs across multiple criteria, such as humor, coherence, and social tone. In more than 70 percent of cases, evaluators preferred scripts produced in the second round of the discussion-based environment.

The improvement appears to stem from collective learning. As agents interact, they build on shared insights, identifying patterns in audience preferences and refining their outputs accordingly. Technically, this process is supported by a shared conversational memory stored in a vector database, allowing agents to retrieve relevant past exchanges and incorporate them into new content. The result is an iterative, emergent form of creativity that begins to resemble human collaboration.

Looking ahead, Hong and her advisor, Assistant Professor Zhicong Lu, aim to expand the system by incorporating human participants and richer modes of interaction. Future directions include integrating voice, gestures, and immersive technologies such as virtual and augmented reality to better capture the nuances of live performance.

"In real comedy, we have faces, gestures, and real voices," she said. "We hope to bring more of that into the system."

By demonstrating that workshopping can significantly enhance machine-generated humor, Hong's work suggests the future of artificial intelligence may lie not only in more advanced individual models, but also in communities of agents learning, creating, and perhaps even laughing together.

This week in AI

CIO Charmaine Madison honored with Change Champion Award

Charmaine Madison, vice president for information technology and chief information officer, has been honored by Cloudforce with a Change Champion Award for leading transformative change at George Mason in the realm of AI, inspiring teams, and challenging the status quo to adopt new ways of working.

Learn how Madison champions the adoption of generative AI across the university.

Explore Research at George Mason

Related Stories