Research · Part III
IP World Models: The Power of “What if?”
Priya DeshpandeNovember 20268 min read
The entertainment industry is walking in the dark. A studio spends hundreds of millions of dollars on making and marketing a film, but only finds out whom that film is for after the release when it sees who buys tickets. By then the budget is spent, decisions on casting, positioning and release strategies cannot be taken back, and there’s no way to know whether another choice would have been more successful.
Most decisions in entertainment today still rest on signals that only hint at how an audience will react, like genre trends, the size of an existing fanbase, the strength of a script, or the track record of a director or star. What if those critical decisions on casting, positioning and release strategies could be simulated first?
Simulating choices like these is what world models are built for. They provide the “what if?” option.
A world model is an AI system that learns how an environment behaves, so it can predict what will happen when something in that environment changes. It’s how self-driving cars and robots now practice in simulation before they’re trusted in the real world.
Audiences can be modeled the same way. At any moment they have a state, made up of what people are watching, what they care about and which communities they belong to. They react to what studios do, whether that’s a trailer, a casting announcement or a release, and those reactions reshape the audience as fandoms grow, fade or spill into neighboring communities. World Models ask the “what if?” question at scale, before finding out the expensive way.
World models
World models are now being built across many fields where experimentation in the real world is either dangerous, slow, expensive, complex or unrepeatable.
A pilot can fly for years without an engine failing at takeoff, yet that is the moment their training has to prepare them for. That’s why pilots rehearse it in simulators long before it happens for real. Robots now learn the same way. NVIDIA’s Cosmos and Meta’s V-JEPA 2 let them practice tasks in simulation, thousands of times over, before they’re deployed. Google DeepMind’s Genie 3 and Runway’s GWM-1 generate interactive worlds that people and AI agents can explore.
Some systems can’t be experimented on at all. No one can rerun a hurricane to see what would have happened if it had turned east instead of west. So DeepMind’s GraphCast learns how the atmosphere behaves and, starting from today’s conditions, simulates how it will unfold over the next ten days so people in the path of a potential hurricane, for example, could take precautions in advance.
Entertainment has the same problem. A film is released once. No studio ever gets to see how it would have performed with a different lead, a different trailer or a different release date.
World models for entertainment
Chronicle is building IP world models to solve the “hurricane problem” for entertainment. Each world model is trained on a single IP and learns what gives it its soul, from the characters and tone that carry it to the fans attached to each of them and the reasons they love it. Studios, filmmakers and game designers can use it to see how their audience is likely to respond to a decision on a project before they commit time and money to make it.
An IP world model makes it possible to simulate these decisions in advance. A studio could estimate how audiences would respond to a new lead in a franchise sequel before the announcement, including which cohorts would become more interested and which new audiences the actor would bring with them. A rights holder could test whether the audience for one franchise has any real affinity for another, and which content would connect the two. A team developing a spin-off could compare a shortlist of leads against the IP’s existing and potential fans before making an offer.
Each of these is a question about cause and effect within a specific audience, which is the kind of question world models are designed to answer.
Studios already try to answer some of these questions before launch, usually through surveys, panels and focus groups. These methods capture what a few hundred people say they would do, at a single point in time. A world model learns from what much larger audiences actually do and keeps learning as they change, which makes it possible to explore far more scenarios than a research program can cover.
How we build IP world models
Each IP world model is trained in three stages. It first learns from observation, then learns the specific IP, and finally learns from actions.
In the first stage, we pre-train the model on a large observational corpus of content–audience interactions. The corpus is built from public platform data at scale, including a near-complete index of YouTube and engagement signals from TikTok, Instagram, Reddit and community platforms. From it, the model learns joint representations of content and audiences: multimodal representations of content attributes such as theme, tone, format, character and visual style, and behavioral representations of what different audiences watch, discuss and share. Pre-training also gives the model population-level priors on viewing behavior, such as which genres Gen-Z gravitates toward, or how genre preferences shift across aggregated demographic segments.
In the second stage, we fine-tune the pre-trained model on an IP-specific corpus. This adapts its representations to the IP’s titles, characters, eras and tone, and to the full range of ways audiences engage with it, from film and series to games, merchandise and fan communities. For a long-running science fiction franchise, the fine-tuned model might learn that affinity is driven by its optimism about the future, the sense of belonging that comes from a crew, the depth of its canon and the comfort of rewatching familiar episodes, and that different cohorts weigh these drivers differently. The result is a structured map of the fandom, from core fans to adjacent cohorts whose tastes overlap with the IP but who have not yet engaged with it.
In the third stage, we train on actions. We continuously run controlled A/B experiments with digital audiences across platforms, exposing specific cohorts to specific content and measuring the response across the funnel, from view rate and watch time to engagement and conversion. This interventional data conditions the model on actions as well as observations, linking each intervention to a measured outcome, which is what allows it to simulate decisions that haven’t been made yet. Training follows a continual learning paradigm: new experimental results are fed back into the model on an ongoing basis, so each IP world model adapts to shifts in style, trends and audience behavior, and its simulations reflect the audience as it is today.
All data is aggregated at the cohort level. We do not track or identify individual users.
Experiments
Audience prediction should benefit from the same approach that has worked across foundation models: learn the general patterns first, then specialize. A model that has learned how people engage with content in general starts with a strong sense of how audiences behave. Fine-tuning it on a single IP lets it build on that understanding, so it doesn’t have to learn general audience behavior from scratch. To test this, we compared three ways of building a model for an IP: a general model with no IP-specific training, a model trained only on the IP’s own data, and the general model fine-tuned on the IP.
Setup. Every model got the same task: predict which audiences will respond to a video it has never seen. It’s the same task we benchmarked in Part I and Part II of this series.
- The content: 72 videos from eight IPs we’ve run campaigns for, spanning an interview series, kids’ animation, indie film, food and more.
- The test: each video was held back from training, and the model’s predictions were checked against real campaign results.
- The measure: how often the model correctly picks the higher-affinity audience.
Results. Fine-tuning the general model on each IP gave the best results.
Fine-tuning improved predictions for 53% of the videos, and no IP got noticeably worse. The biggest gains came where the general model struggled most: for an animated children’s series, accuracy rose by more than seven points. A model that already understands audiences in general learns a specific IP better than one starting from nothing.
Learning from an IP’s own data alone worked well in some cases and badly in others. For IPs built around a single show with a consistent audience, such as an interview series where each episode was tested across a wide range of channels, the IP-only model beat the general model by up to six points, as expected. For an IP whose videos, and consequently audiences, came from varied genre categories, each tested on only a handful of channels, it fell 21 points behind. IPs that span many formats and genres may be better served by starting from the general model, since there is less consistent signal for a model trained on the IP alone to learn from. For these IPs, the priors learned in general pre-training help explain how and why non-intuitive genres respond to the content. The difference comes down to how much of the IP’s world the model has seen. A model trained on the IP alone only pays off once the IP has been tested deeply and consistently, across several pieces of content and many of the same audiences. Until then, starting from general knowledge works better.
How much data an IP world model needs. We also looked at how quickly accuracy improves as we run more tests for an IP. We retrained models on growing slices of three IPs’ data, varying both the number of videos tested and the number of channels each video was tested on.
Accuracy rises with both, but the number of different videos matters most. A model trained on a single video barely beats a coin flip, at 54–58%, however many channels that video was tested on. With 15 videos it reaches 72% even when each was tested on only five channels, and 81% when each was tested on up to 200, at which point it overtakes the general model (78.7%). Testing more pieces of content teaches the model more about an audience than testing one piece more widely.
Starting from general knowledge gets there much faster. For a celebrity interview podcast, the general model that had never seen the IP started at 71%. Fine-tuning it on a single video lifted that to about 77%, while a model trained on that one video alone scored 58%. Within two to four videos, the fine-tuned model matched the general model trained on all of our data, including this IP’s, and with 15 videos it reached 83%. An independent film and documentary IP followed the same pattern: one video took the fine-tuned model from 76% to about 80%, it caught up with the fully trained general model at around eight videos, and it reached 91% by fifteen.
In practice, a new IP doesn’t need a long history of campaigns before its world model is useful. A small, targeted set of tests on top of general audience knowledge can match a model trained on everything we have, and every test after that makes it better.
Next steps. Our future research focuses on extending IP world models from predicting single outcomes to simulating full scenarios, so teams can explore how an audience responds to a series of decisions over time.
