For more than a century, cinema has asked one precise thing of the viewer: follow a path that has already been set. Even when a film left room for ambiguity, the shots, rhythm and order of scenes were decided before the encounter with the audience. Generative AI makes an old ambition of interactivity practical again: not simply choosing between two endings, but moving through a narrative world that can open variations, change perspective and remember previous choices.
This does not mean every film has to become a video game. The point is subtler: some works can become narrative systems, designed with rules, characters, locations and constraints clear enough to generate different episodes without losing identity. Showrunner, Fable’s platform, presents this idea through animated series built as simulations: viewers can watch a story and then intervene, develop their own scene or follow a deviation. It is still an early experiment, but it captures the question now reaching cinema: what happens when a work is no longer only a finished file?
AI mainly changes the production scale of alternatives. A traditional branching story requires every possibility to be written, shot, edited and checked in advance. That is why deviations remain few and expensive. A generative system can instead prepare many variants of dialogue, space, framing or transitions, provided there is a structure holding them together. The decisive element is not the number of scenes but coherence. If a character changes goals with every response or the aesthetic shifts for no reason, freedom quickly becomes noise.
This is where a new form of editing appears. The editor no longer only orders images that already exist; they also design the conditions under which new images may appear. They decide which points must not change, which information the viewer must receive, which deviations are allowed and which are denied. A scene might open through three points of view while a certain relationship remains ambiguous; the music might follow a theme while shot rhythm reacts to the time spent inside a sequence. Editing becomes a grammar of possibilities.
Research on controllable generative narratives insists on exactly this point: openness is not enough. Models need to work with narrative structures, genre constraints, relationships between characters and readable goals. Without authorial direction, a system tends to choose the most predictable or statistically plausible solution. In a story, however, the interesting gesture may be precisely the one that interrupts prediction. The role of directing therefore remains decisive: defining what the model should not solve too well.
The audience, meanwhile, does not automatically become a co-author. It can be explorer, witness, player or guest. These are different roles and need careful design. Too many choices can break emotion; no choices make interactivity unnecessary. The useful question is not ‘how much control do we want to give?’ but ‘at what moment does the audience’s choice actually create a new point of view?’ A variation has value when it reveals something the linear version could not show.
There are also practical limits. For a story to react, it needs contextual data, memory of decisions, content controls and clear rules around rights, voices, faces and narrative worlds. A character generating dialogue in real time cannot use a performer’s identity without authorization, and an open world cannot blur original material with a community-created version. The more modifiable a work becomes, the more legible its boundaries must be.
VARIABLE STORIES is therefore not the name of a future in which a machine invents films instead of authors. It is the hypothesis of a cinema that can continue to live after the first edit without giving up form. AI can make tests, branches and perspectives accessible that once would have cost too much; human responsibility remains choosing the question, designing the rules and deciding when a story must finally stop.