,

Can An Algorithm Greenlight A Movie? - The AI Insights of Film Development

,

Intro

Like many industries, the entertainment industry is exploring how to use artificial intelligence (AI) to make business decisions. What happens to original storytelling when Hollywood uses AI to predict a film’s success? In this episode of Tech and the Arts,  AMT Lab Contributor Minky Kang investigates whether AI can be used to determine if a movie should be produced, or “greenlit.”

Show Notes

Evaluating Creative Work with Artificial Intelligence: Evidence from Constrained Innovation Tasks

Ghost in the (Hollywood) Machine: Emergent Applications of Artificial Intelligence in the Film Industry

Kim, Hyungjin Lukas. “The Algorithm Paradox: Why Data-Driven Film Production Systematically Misses Blockbusters.” Applied Economics Letters, published online April 27, 2026.The Algorithm Paradox: Why Data-Driven Film Production Systematically Misses Blockbusters

Lehrer, Steven F and Tian Xie. “Box Office Buzz: Does Social Media Data Steal the Show from Model Uncertainty When Forecasting for Hollywood?” Review of Economics and Statistics 99, no. 5 (2017): 749–755.
Box Office Buzz: Does Social Media Data Steal the Show from Model Uncertainty When Forecasting for Hollywood?

Lehrer, Steven F and Tian Xie. “The Bigger Picture: Combining Econometrics with Analytics Improves Forecasts of Movie Success.” Management Science 68, no. 1 (2022): 189–210.The Bigger Picture: Combining Econometrics with Analytics Improves Forecasts of Movie Success

Leukel, Joerg, Zejun Liu and Vijayan Sugumaran. “Potential Overinterpretation of Results in the Abstracts of Machine Learning Studies for Movie Box Office Revenue Prediction: A Systematic Review.” Journal of Revenue and Pricing Management 25 (2026): 368–377.Potential Overinterpretation of Results in the Abstracts of Machine Learning Studies for Movie Box Office Revenue Prediction: A Systematic Review

Liu, Yan and Tian Xie. “Machine Learning versus Econometrics: Prediction of Box Office.” Applied Economics Letters 26, no. 2 (2019): 124–130.Machine Learning versus Econometrics: Prediction of Box Office

Moon, Sangkil, Nima Jalali and Reo Song. “Green-lighting Scripts in the Movie Pre-production Stage: An Application of Consumption Experience Carryover Theory.” Journal of Business Research 140 (2022): 332–345.
Green-lighting Scripts in the Movie Pre-production Stage: An Application of Consumption Experience Carryover Theory.

Navar-Gill, Annemarie. “The Golden Ratio of Algorithms to Artists? Streaming Services and the Platformization of Creativity in American Television Production.” Social Media + Society 6, no. 3 (2020): 1–11.
The Golden Ratio of Algorithms to Artists? Streaming Services and the Platformization of Creativity in American Television Production

Paul, Chiranjib and Prabir Kumar Das. “Predicting Movie Revenue before Committing Significant Investments.” Journal of Media Economics 34, no. 2 (2022): 63–90.Predicting Movie Revenue before Committing Significant Investments.

Reilly, Dan. “No, A.I. Isn’t Deciding Which Movies to Green-Light.” Fortune, February 25, 2020.
No, A.I. Isn’t Deciding Which Movies to Green-Light.

Rose, Steve. “It’s a War between Technology and a Donkey’—How AI Is Shaking Up Hollywood.” The Guardian, January 16, 2020
It’s a War between Technology and a Donkey’—How AI Is Shaking Up Hollywood

Zalipour, Arezou. “AI at the Greenlight: Negotiating Creative Agency in Early-Stage Film Development.” Media Practice and Education, published online March 2, 2026.AI at the Greenlight: Negotiating Creative Agency in Early-Stage Film Development

Transcript

Minky Kang

The moment a revenue prediction starts functioning as a green light score, something subtle has happened. We've moved from commercial prediction to creative evaluation, and those are not the same thing.

Ife Martin

Welcome to another episode of Tech and the Arts. The podcast series of the Arts Management and Technology Laboratory, also known as AMT Lab.

The goal of our podcast series is to exchange ideas, discuss emerging technologies, and uncover ideas that arts managers and arts geeks like us need to know. My name is Ife Martin. I'm the podcast producer and technology manager for AMT Lab.

In this episode, Minky Kang takes us into the world of TV and film, exploring how algorithms affect the content we see and make.

Minky Kang

Lately, I've noticed something when I watch movies and series on streaming platforms, especially Netflix. A lot of them are fine. Nothing is particularly wrong with them, but so much of it feels strangely familiar, a little too polished and safe, and then a week later, I can barely remember them.

I've worked in entertainment production myself, so maybe that's why this bothers me. Because some of the stories I remember most vividly are not the ones that felt perfectly polished. They're the strange ones, the risky ones, the movies where, when you first hear the premise, you can imagine an executive sitting in a conference room saying, "You wanna make what?"

Think about Parasite or Everything Everywhere All at Once or Get Out. In hindsight, their success can almost feel inevitable, but before they existed, they were bets, and that made me wonder: as Hollywood gets better at using data to predict what audiences will watch, are we also getting better at playing it safe? And if so, what happens to original storytelling? When I started researching this, I had a simple picture in my head. Hollywood has more audience data than ever. Machine learning gets better at predicting what works. Studios become more risk-averse, and eventually we get safer and safer movies. But I found out that reality was more complicated. In 2020 , Warner Brothers announced a deal with Cinelytic, a company that uses machine learning to forecast movie performance.

And despite some of the headlines at the time, Cinelytic was not an AI executive sitting somewhere reading scripts and stamping them green light, reject, green light. Dan Riley reported in Fortune that the system was designed more like a decision support tool. Executives could change variables, cast, release strategy, genre, whether a film was a sequel or adaptation, and see how the financial projection changed.

Cinelytic's CEO, Tobias Quaasr, put it pretty clearly. "We don't believe that an algorithm can run a script and understand whether the story is good." The point was to crunch numbers and identify patterns and then let humans decide what to do with them. Okay, so if the algorithm itself isn't making the decision, maybe the creative judgment is still safely in human hands.

However, final authority isn't the only way a system can influence a creative decision.

Arezou Zalipour explored this problem in a 2026 study of AI-assisted development on the German-New Zealand feature film, Come Together. The AI system wasn't simply replacing a producer or director. It was part of the development process, analyzing the project and providing feedback, and Zalipour identified several tensions in that process: creative flattening, formulaic feedback, and blind spots. The system could offer useful market information, but it struggled with dimensions of storytelling that were emotional, cultural, contextual, and philosophical. And some of its feedback could pull the project toward patterns that were easier for the system to recognize. That distinction felt important.

An algorithm doesn't have to say, "Do not green-light this." If it repeatedly tells the people developing a project that something is unusual, risky, or inconsistent with recognizable patterns, that information can still shape what they change and what they keep. Influence doesn't require authority. So my original concern wasn't completely unfounded.

AI could affect creative choices even when humans technically remained in charge. And if these systems learn from historical data, they will naturally reward projects that resemble what succeeded before. But then I found evidence pointing in the other direction. This is where my original theory started to get more complicated. I found a study by Sangkil Moon, Nima Jalali, and Reo Song that used screenplay text to predict revenue during pre-production. But their model did not simply assume that audiences want more of whatever they liked before. It allowed for two possibilities.

Sometimes audiences want familiarity, but sometimes they become saturated and want something different. In other words, novelty itself can contain predictive information. That matters because it means the problem is not as simple as data equals sameness. An algorithm can potentially recognize that difference has value.

Annemarie Navar-Gill found something similarly counterintuitive in her research on American television writers working with streaming platforms. Data could sometimes give platforms evidence that an audience existed for a project that traditional Hollywood intuition might have dismissed. So human instinct isn't automatically the guardian of creativity either.

Humans have biases. Humans repeat conventional wisdom, and data can sometimes challenge it.

So maybe predictive AI wasn't necessarily the enemy of originality. Maybe the more important question was what exactly these systems were predicting and how Hollywood was using this data. Imagine you're a studio executive. You have two versions of the same project. Project A: seventy dollar million budget, established director, major movie star, familiar genre. Project B: thirty-five dollar million budget, emerging director, less famous cast, unusual genre combination. Now, imagine an algorithm can tell you how changing those variables affects expected revenue.

Why wouldn't you want that information? That's essentially the problem Chiranjib Paul and Prabir Kumar Das studied in 2022. They built machine learning models using information available during development, including cast and crew, genre, and distinctive project characteristics to predict movie revenue before a major investment is committed.

Their best machine learning model reduced total box office forecast error by about ten percent compared with their traditional regression benchmark. This is not trivial, and other research points in the same direction. Yan Liu and Tian Xie found value in machine learning approaches to box office forecasting.

Steven Lehrer and Xie found further gains by combining analytics with econometric methods. So if I were responsible for investing tens of millions of dollars in a movie, I would probably wanna see that forecast too. But at this point, I started to notice a pattern. What exactly were these systems predicting? Revenue. Predicting how much money this movie is likely to make is not the same question as whether this movie exists, and that distinction started bothering me. A movie can succeed financially, but it can also succeed because it introduces a new cinematic language, because it represents people who rarely see themselves on screen, because it starts a cultural conversation, or simply because twenty years later, people still remember how it made them feel.

Revenue is attractive to a predictive system because it gives you something measurable, a number. But the moment a revenue prediction starts functioning as a green light score, something subtle has happened. We've moved from commercial prediction to creative evaluation, and those are not the same thing.

And if you're a studio executive, you might say, "Okay, that's unfortunate for art, but we're running a business." Then I found a 2026 paper by Hyungjin Lukas Kim. Kim compared 156 films released between 2017 and 2024, classified as using data-driven or intuition-based production approaches.

The average returns, surprisingly similar, but the distribution of those returns was very different. Films classified as intuition-based had about four times the variance. They failed bigger, but they also won bigger. And among the top ten percent of performers in Kim's sample, 81% were intuition-based films.

Kim calls this the algorithm paradox. In an industry with fat-tailed returns, where a tiny number of extraordinary hits can generate enormous value, optimizing for predictability may actually reduce your exposure to exceptional success. This is where uncertainty becomes interesting. In most business contexts, uncertainty sounds like something you want to eliminate. But what if in Hollywood, uncertainty is sometimes an asset? What if the same characteristics that make a project difficult to predict are also what make it capable of becoming something the market has never seen before?

Kim specifically points to films like Parasite and Everything Everywhere All at Once as the kinds of projects that would have faced algorithmic skepticism based on their observable pre-production characteristics. That does not mean an AI literally rejected those movies, but it suggests something more interesting.

A system optimized to find predictable winners may be structurally bad at recognizing unprecedented ones.

And this gives Hollywood a reason to protect creative risk that has nothing to do with being sentimental about art. Money. If the economics of film are driven partly by rare extreme hits, then protecting uncertain projects isn't necessarily charity for artists. It may be good portfolio strategy. So instead of asking, how do we eliminate uncertainty, an executive might ask, what kind of uncertainty am I looking at?

Is the model uncertain because the project is weak? Or is it uncertain because there simply aren't good historical comparables? Those are two very different things, and a green light system that treats both as the same red flag could optimize away its next breakout hit.

At this point, I had what seemed like an obvious solution. If revenue is too narrow, why not give AI more things to measure? Originality, novelty, creative value, maybe even emotional impact. And surprisingly, there is research suggesting that at least part of this is possible. A 2026 study by Valerio Fedele Addis and colleagues asked ChatGPT to evaluate creative work produced by professionals in cultural and creative industries.

The tasks were constrained, verbal, mathematical, and visual creative problems. So this was not the same as asking AI whether Parasite is a masterpiece. But the results were interesting. AI evaluation showed reliability comparable to expert human judges in that setting, and the researchers could break its evaluation down across four dimensions: fluency, flexibility, originality, and elaboration.

The relative importance of those dimensions even changes depending on the task. So computational systems do not necessarily have to measure only revenue. They can potentially give us information about multiple dimensions of creativity.

But then we run into another problem. Imagine our studio dashboard now looks like this. Predicted revenue, 82. Audience familiarity, 87. Narrative novelty, 91. Originality, 88. Commercial risk, 64. Great. We've solved it, right? Except somebody eventually asks, "So should we make the movie?"

And the temptation is to collapse everything into one number. Greenlight score, 84. But how did we get 84? Who chose the weights? How much should each criterion count? And what happens when originality and commercial predictability directly conflict? At that point, the AI is no longer simply measuring.

We have encoded a set of human values into an optimization function and then given the resulting number the appearance of objectivity. Creativity is multidimensional. It is contextual. And sometimes the work that matters most is the work that violates the criteria we thought mattered. Even the ADAS study points to this limitation.

Its strongest results come from constrained creative tasks, not open-ended artistic judgment. So I don't think the answer is to build a better all-knowing greenlight score. I think it's almost the opposite.

I started with a pretty simple suspicion. Maybe predictive AI was making Hollywood better at playing it safe and worse at producing original stories. But the deeper I went, the less convincing that explanation became. Data itself doesn't necessarily push Hollywood towards sameness. It can recognize the value of novelty.

It can challenge conventional industry assumptions and, in some cases, it may even give executives more confidence to take creative risks. So I don't think the bigger problem is that Hollywood is using AI. It's what we ask AI to measure, what we define as success, and how much authority we give those measurements in the final decision. And that led me to a different question than the one I started with.

What should AI be allowed to measure, and what should those measurements mean in a green light decision? My answer isn't to use less information. It's to be more careful about what we ask the information to do. Let AI estimate revenue, let it model financial risk, identify market patterns, and maybe let it provide carefully defined measures of novelty, originality, or other creative dimensions.

But keep those outputs visible as separate signals. Don't collapse them into one authoritative number that pretends the trade-offs have already been solved. Because imagine the model tells you this project has high financial uncertainty, but extraordinary novelty, there's no good historical comparable, and the creative team has a vision everyone in the room can't stop talking about. What is the correct weight for each of those things? There isn't one.

That's not a prediction problem anymore. That's a judgment. And that judgment should remain with the human green light team, along with the responsibility for making it. The disagreement between those signals isn't noise that needs to be optimized away. That disagreement is the decision. Because Hollywood doesn't only need movies that are unlikely to fail; it needs room for movies that nobody knows how to predict yet.

Maybe the next movie we remember for the rest of our lives looks terrible in the spreadsheet.

Ife Martin

Thank you for listening to this episode of Tech in the Arts. If you found this episode to be informative, educational, or inspirational, be sure to check out our other episodes and send this to another arts or technology aficionado in your life. If you want to know more about arts management and technology, check out our website at amt-lab.org, or you can email us at info@amt-lab.org. You can follow us on Instagram at techinthearts, or on LinkedIn at our full name, Arts Management and Technology Lab.

Our season runs from September through May, so feel free to follow us on your preferred podcast app, be it iTunes, Spotify, or YouTube.