Scott Z. Burns tested ChatGPT on a Contagion sequel and learned what machines cannot replace

AI Cannot Write Your Scenes and That Is the Point

Scott Z. Burns is not an anti-AI zealot. He is the screenwriter behind “Contagion,” “The Informant,” and “An Inconvenient Truth,” a serious and accomplished professional who approached the question of artificial intelligence and screenwriting with what he describes as genuine skepticism rather than predetermined hostility. When he documented his experiment in using large language models to help develop a sequel to “Contagion,” he did so through an eight-episode Audible Original podcast series called “What Could Go Wrong?” that premiered at the Tribeca Film Festival in June 2025. The conclusions he reached are more nuanced and more useful than either the AI-utopian or AI-panic camps typically allow.

What He Tried to Do

Burns and director Steven Soderbergh had been unable to crack the premise for a “Contagion” sequel. The original film found its second massive audience during the COVID-19 pandemic, when millions of viewers discovered that the 2011 thriller had anticipated the shape of a modern pandemic with uncanny accuracy. But making a second fictional pandemic thriller after audiences had lived through a real one posed a creative challenge that standard brainstorming had not resolved. Burns hoped that AI could help rapidly explore scientific permutations, testing multiple pandemic scenarios quickly to identify a premise that felt both scientifically plausible and dramatically viable.

He created a customized ChatGPT instance he named Lexter and began working with it. In one experiment documented in the series, he attempted to break a scene with actors Jennifer Ehle and Laurence Fishburne, who play CDC doctors in the original film, with Lexter participating as a third collaborator. The results were instructive. Lexter could not capture the emotional texture of the relationship between two former colleagues. It could not generate the specific, earned intimacy that a scene between these two characters required. “It does run out of gas when you start getting into some of the more detailed parts of filmmaking,” Burns told IndieWire. “It doesn’t have an experience of itself in three-dimensional space, and we underestimate what our physicality makes available to us.”

Where AI Was Actually Useful

For the task Burns actually had in mind, rapid exploration of scientific possibilities, AI proved genuinely useful. It could generate multiple plausible pandemic scenarios quickly, offer lists of permutations, and serve as a tireless respondent to “what if” questions. The breakthrough that emerged from the experiment came not from AI’s competence but from a creative accident: Lexter, given a prompt with specific constraints, disregarded one of those constraints and generated a scenario outside the parameters Burns had established. That mistake became the seed of the most promising premise Burns and Soderbergh found.

This is a significant and honest finding. The useful contribution AI made to this creative process was not intelligence or imagination. It was the kind of productive error that comes from an entity that does not fully understand the constraints it has been given. This is useful precisely because it is a failure mode that a skilled human collaborator would have avoided. The AI’s limitation became, in this specific instance, a creative resource.

The Derivative Problem

Burns raises a more systemic concern about what happens when studios rely on AI to generate movie premises at scale. If a large language model generates premise ideas by recombining elements from the films it has been trained on, the output will by definition tend toward the derivative. Original films like “Anora” emerge from a specific human vision that cannot be synthesized from prior films because it represents a genuinely new way of seeing. The WGA contract protections on AI are designed in part to prevent this derivative tendency from colonizing the development process.

Burns’ concern is not hypothetical. He describes the process of using AI to brainstorm premises as feeling, at a certain point, indistinguishable from what a streamer might already be doing: “reassembling the constituent parts of other movies into a new movie.” This is a structural problem with AI brainstorming as a development tool, not a problem with any individual tool or application. The technology’s strength, rapid synthesis of existing patterns, is also its fundamental creative limitation.

The Fear of Transparency

One of the most revealing aspects of Burns’ account is his description of the fear within the writers’ guild about being seen to use AI at all. He acknowledges significant anxiety about his peers’ reactions to his experiment even as he argues that the experiment produced valuable and honest findings. He is currently running a writers’ room for a Netflix series and observes that his writers are afraid to use AI even for tasks it handles well. This fear, he believes, is counterproductive and stems from a failure to distinguish between AI as a replacement for human creative labor and AI as one tool among many in a writer’s research and development process. Resources like IndieWire’s Future of Filmmaking coverage track this evolving debate in real time. Burns’ conclusion is the most useful available: understand what the tool cannot do, and you will understand what you must protect. What it cannot do is write scenes. What it cannot replace is the human being who has lived in three-dimensional space, felt things, and found a way to put that experience on a page.