Learning by Building: Inside Wharton’s Generative AI Studio

At the Wharton Generative AI Labs (GAIL), experimentation is not a side effect of learning — it is the learning.

GAIL, which is part of the Wharton AI & Analytics Initiative (WAIAI) and led by faculty director Ethan Mollick, aims to explore how generative AI is reshaping the way we build, design, and think in the age of AI. Its flagship student program, the Generative AI Studio, offers a different kind of academic experience, one that replaces lectures and exams with prototyping, critique, and weekly iteration.

“It’s a creative space,” says first-year MBA student Hetvee Marviya. “Just like I dance, this is something that’s creative, but using AI.”

A Different Kind of Lab

The Studio’s core directive is simple: students are there to build. Over the course of a semester, each participant develops an original generative AI project, moving from idea to working prototype. The cohort is intentionally small, just eight to ten students.

Laura Zarrow, executive director of GAIL, and Chris Callison-Burch, professor of computer and information science at the University of Pennsylvania, round out the support structure each week, offering their professional perspective to supplement the students’. “The GAIL Studio is a place that I come back to every week where we build to explore the boundaries of AI, and then give each other constructive feedback,” Marviya explains.

Some sessions function as deep work sprints; others introduce technical concepts like prompt design, system architecture, or deployment. This emphasis on critique and iteration reflects a core philosophy of the Studio: generative AI is not just a productivity tool, but a new creative medium.

Unlike Accelerators or entrepreneurship labs, the Studio does not require students to build venture-ready products. In fact, participants are explicitly encouraged not to focus on commercialization too early. “Don’t think about this as something where you have to make it into a business,” Marviya recalls being told. “It doesn’t have to be about efficiency. It has to be about how you use this medium to explore.”

The result is a shift in mindset from execution to exploration, from convergence to cycles of divergence and refinement.

From Exploration to Application

Marviya came to Wharton with a background in engineering, investing, and AI startups, and one clear goal: to transition fully into the technology sector. “I had been trying to find opportunities where I can be more immersive, where I can build things or explore AI,” she says. When she discovered the Studio during winter break, what stood out wasn’t a specific curriculum, it was the framing.

Marviya’s project reflects that exploratory ethos.

For her Studio project, she’s building an AI-powered platform that transforms current events into personalized comic strips for children — an effort (informed by the digital consumption habits of her younger relatives) to make news more engaging, accessible, and developmentally appropriate.

“At its simplest, it’s to make current affairs digestible for kids,” she explains. Users can input a topic, specify age and narrative preferences, and receive a dynamically generated comic tailored to those parameters. Beneath that simple interface, however, lies a surprisingly complex system.

What began as a straightforward idea quickly expanded into a series of deeper questions:

• How do you ensure news content is unbiased?
• How should information be adapted for different age groups?
• What makes content both educational and engaging?

“I didn’t realize how many barriers there would be,” she says. “It sounds simple… but as we uncovered it through the Studio, I was pushed to think about bias, sources, and how information changes across age groups.”

Designing the Pipeline

To address these challenges, Marviya developed a multi-stage AI pipeline — an architecture shaped through experimentation, critique, and iteration within the Studio. The first stage focuses on fact collection. “If you input a news topic, it first just collects all facts,” she explains. “That’s a document that can be reviewed before it moves on.”

This deliberate separation of fact-gathering from storytelling helps mitigate bias and ensures a more neutral foundation before transformation. From there, the system moves through several layers:

• Selecting and validating a news topic
• Aggregating and synthesizing information
• Adapting content based on age-specific parameters
• Converting the material into a narrative
• Rendering that narrative into a comic format

Marviya used tools like ChatGPT and Claude for ideation and prompt development, then transitioned to Claude Code and Cursor to build the application architecture. For design and interface work, she leveraged tools such as Figma Make, while image generation relied on models like Gemini and Nano Banana.

“We are free to use any multimedia tools that we want,” Marviya says, highlighting the limitlessness of the freedom afforded to the Studio’s participants.

“I first broke this process down into different steps…and froze each part once it was well tested,” she says. This modular approach allowed her to stabilize outputs and reduce issues like drift or inconsistency, common challenges in generative systems.

The project also surfaced more complex questions around authorship and intellectual property. Rather than relying on stylistic imitation, Marviya began exploring ways to ground outputs in original or permissioned inputs, experimenting with artist-provided reference images and emphasizing collaboration over replication. As she puts it, “there is still value of the human aspect,” a principle that continues to shape how she evolves the system.

The result is not just a prototype, but a repeatable system for building with AI.

Beyond the Studio

While the Studio emphasizes exploration over execution, many projects naturally evolve beyond the semester, including Marviya’s.

“I’d love to explore this project further, and I am actively doing that even beyond the Studio,” she says.

She sees potential applications in both education and consumer products, from daily content subscriptions for children to broader tools for engaging with news. “There’s definitely something here for educational purposes…or even for adults who find it more engaging,” she notes.

She is also exploring collaborations with artists, carefully considering questions of authorship and intellectual property in AI-generated work. “There is still value of the human aspect,” she says.

A New Model for AI Education

As generative AI continues to reshape industries, programs like GAIL’s Studio offer a glimpse into how business education is evolving alongside it. At Wharton, students are increasingly encouraged not just to use AI, but to understand it, experiment with it, and build with it.

“Hetvee came to the Studio with an ambitious but well-defined goal, and she used her time here to develop both the technical skills and the aesthetic acumen the project demanded,” said Laura Zarrow, GAIL’s executive director. “Key to this was her sincere curiosity and her ability to learn from those around her. She used each session to identify the opportunities to expand and enhance her own work, while contributing to the success of everyone else’s project.

For Marviya, the opportunity to learn and experiment with AI in this way has been critical. “GAIL has been the most impactful part of my AI learning at Wharton,” she says.