AI-generated video and the return of interactive film

Interactive film has been “about to happen” for nearly sixty years. Every generation of the idea has hit the same wall: video is expensive to produce, so every branch has to be filmed, so the number of branches stays small, so the player’s choices feel thin. AI-generated video is the first technology that attacks the wall directly. If frames can be produced on demand, the branch does not need to exist before the player chooses it. That is why generated video, and the real-time world models growing out of it, is where interactive storytelling is heading. It is also why the demos deserve scepticism. Text-based generative games work today; generated video is the future, and the gap between those two states is mostly about latency, consistency and control. This piece looks at the lineage, the current research, what “interactive” would really require, and the hybrid designs an AI game designer can build before the hard problems are solved.

A short lineage of interactive film

The idea predates video games. Kinoautomat, a Czechoslovak comedy shown at Expo 67 in Montreal, stopped at nine points so a moderator could ask the audience to vote between two scenes. The whole room got one choice, and, famously, the branches tended to converge on the same ending. That pattern, many choices funnelling back to a small number of outcomes, would follow interactive film for decades.

In 1983, Dragon’s Lair put a laserdisc inside an arcade cabinet. Cinematronics released it with animation by former Disney animator Don Bluth’s studio, and it looked unlike anything else in an arcade. It played, though, as a sequence of timed inputs: press the right direction at the right moment or watch a death animation. The beauty was pre-rendered; the interactivity was a reflex test.

The CD-ROM era brought full-motion video (FMV) games. Night Trap (1992, Sega CD) became notorious less for its design than for its part in the US congressional hearings on video game violence. FMV games of that period mostly shared a problem: grainy, expensive footage with limited agency, and the format faded.

The revival came from design rather than technology. Sam Barlow’s Her Story (2015) gave the player a police database of short video clips of one woman’s interviews, searchable by keyword. Nothing branched. The interactivity was in what you chose to look for and in what you deduced. It remains one of the best arguments that interactive video does not require the video itself to change.

Then Netflix released Black Mirror: Bandersnatch in December 2018, a branching film with multiple endings delivered to millions of viewers on ordinary televisions. It showed both the appeal and the cost: an enormous amount of footage shot for paths most viewers never saw. In 2025 Netflix removed Bandersnatch and its remaining interactive titles from the service.

Work	Year	What was interactive	The limit it hit
Kinoautomat	1967	Audience votes at fixed points	One choice per room; branches converge
Dragon’s Lair	1983	Timed inputs over laserdisc animation	Choices were reflexes, not decisions
Night Trap	1992	Switching between camera feeds	Footage cost and low agency
Her Story	2015	Searching a database of fixed clips	Deliberately fixed content
Bandersnatch	2018	Binary choices on a streaming service	Every branch had to be shot

Each of these is a different answer to the same constraint: the footage must exist before the player arrives.

From text-to-video to world models

Two strands of AI research are relevant, and it helps to keep them apart.

Text-to-video models generate a clip from a prompt. Google’s Veo 3, announced in May 2025, added generated sound and dialogue to the video, and OpenAI’s Sora 2, released at the end of September 2025, did the same. These models are remarkable at producing a shot. They are not designed for interaction: you wait for a clip of a few seconds, and you cannot steer it while it plays.

World models generate frames continuously in response to actions. You press forward and the next frame shows you having moved. A few public examples mark the progress:

• GameNGen (Google Research, 2024) showed in the paper Diffusion Models Are Real-Time Game Engines that a neural model could simulate Doom interactively, frame by frame, with no conventional game engine running.
• Oasis from Decart and Etched, released in October 2024, was a playable, fully generated world in the style of Minecraft, available in a browser. It was impressive and also notably dreamlike: look away and back, and the scenery could change.
• Muse, Microsoft’s World and Human Action Model, was published in Nature in February 2025. It was trained on gameplay from Ninja Theory’s Bleeding Edge and framed explicitly as a tool for gameplay ideation.
• Genie 3 from Google DeepMind, announced in August 2025, generates navigable worlds from a text prompt in real time at 24 frames per second and 720p, staying consistent for a few minutes. It also supports text prompts that change the world while you are in it. In January 2026 Google opened an experimental prototype built on it, Project Genie, to Google AI Ultra subscribers in the US.

The direction is clear. In roughly two years, research moved from simulating one existing game to generating new, explorable worlds from a sentence. What none of these systems yet offers is a stable, rule-bound world that you could build a long game in.

What would make video interactive

“Interactive” is doing a lot of work in most announcements. It helps to break it into three properties. A system needs all three before it can carry a game rather than a demo.

Latency. The time between an input and its visible effect. Games feel responsive when this is very short; anything beyond a fraction of a second for movement starts to feel like steering a boat. Clip-based text-to-video is far outside this range. World models are designed around it, which is exactly why they matter more than better clip generators.

Consistency. Whether the world stays the same when you are not looking at it. Object permanence, a stable layout, characters who look the same in every scene, a door that stays open once you open it. Current world models describe their consistency horizon in minutes. A game needs hours, and saves, and the ability to return to a place days later.

Control. Whether the player’s action reliably produces the intended result, and whether the designer can enforce rules. In a conventional game, a locked door is locked because the code says so. In a pure world model, a locked door is locked only as long as the model keeps believing it. There is no inventory, no hit points, no win condition unless something outside the model tracks them.

Those three properties are, in effect, the same problems text-based games solved first. In a text game, the answer is to keep the rules and state in ordinary code and let the model narrate. Video will probably end up in the same place.

Hybrid approaches you can build now

If you are designing games with AI, you do not need to wait for a perfect world model to experiment. The useful designs today combine video generation with conventional state, the same split that already makes text-based AI games work.

Generated cutscenes. The game runs on ordinary rules, in text, 2D or 3D, and at key moments a clip is generated or selected to show what happened. Because cutscenes are not interactive, latency matters less, and you can pre-generate and review them. This is the most practical starting point.

Branching generated video. A Bandersnatch-style structure where the branches are generated rather than filmed. The tree is authored; the footage is produced per branch and reviewed before release. You get many more branches for the same budget, and the cost moves from the shoot to the edit. The hard part is keeping characters and locations consistent across hundreds of clips, which is where reference images and tight style guides matter.

Video over game state. The most promising idea. A conventional state machine holds the truth: where the player is, what they carry, what has happened. Each turn, the game passes a description of that state to a video model, which renders the scene. The video is a presentation layer, exactly as narration is in a text game. If the model renders something wrong, the state is still right, and the next frame can correct it.

// The game owns the truth; the video model only renders it.
type Scene = {
  location: "harbour" | "lighthouse" | "cellar";
  timeOfDay: "dusk" | "night";
  characters: { id: string; mood: "calm" | "afraid" | "angry" }[];
  event: string | null; // e.g. "lantern_breaks"
};

function shotPrompt(s: Scene, styleGuide: string): string {
  const cast = s.characters.map(c => `${c.id} (${c.mood})`).join(", ");
  return [
    styleGuide,                    // fixed palette, lens, character references
    `Location: ${s.location} at ${s.timeOfDay}.`,
    cast ? `On screen: ${cast}.` : "No characters on screen.",
    s.event ? `Action: ${s.event.replace(/_/g, " ")}.` : "Ambient shot, slow push in.",
  ].join(" ");
}

Database storytelling. The Her Story approach with generated material: a large, reviewed library of short clips the player searches or assembles. The interactivity lives in deduction, which suits a generative pipeline because nothing needs to be produced live.

World models as prototyping tools. Microsoft framed Muse around ideation, and that is the honest use for now: generate an explorable sketch of a place to test mood and layout, then build the real thing with an engine.

Design questions that do not go away

Even with perfect AI video, interactive film has unresolved design problems.

Choice versus authorship. The more freedom a player has, the less a director can shape pacing, framing and emotional arc. Film’s strengths come from control; games’ strengths come from agency. Generative video does not resolve this tension. It only makes the freedom cheaper.

The convergence problem. Kinoautomat converged on one ending partly by design. Generated branches can multiply without end, but stories need endings. Someone still has to decide what the possible outcomes are, and the code has to steer towards them.

Spectacle fatigue. A generated world that looks new every time stops being remarkable quickly. The interest has to come from a loop: a goal, constraints and consequences, the same as any game.

Ethical questions

AI-generated video raises sharper issues than generated text or audio, and an AI game designer inherits all of them.

• Likeness and consent. Video models can produce realistic people. Generating real individuals, including actors, without permission is a harm, and the same consent principles that performers have fought for in voice work apply here.
• Training data and provenance. How video models were trained is contested, and the legal position is unsettled. Keep a record of the tools and terms behind every clip you ship.
• Disclosure. Players should know when footage is synthetic, particularly in anything with a documentary or realistic style.
• Harmful generation at runtime. If a player can steer a video model live, they can try to steer it somewhere you would not approve. Runtime video needs the same moderation thinking as runtime text, with higher stakes because images persuade more than words.
• Compute and cost. Real-time video generation is computationally heavy. Who pays for it, and what that means for access, will shape who gets to make and play these games.

Takeaways

• Interactive film has always been limited by one thing: every branch had to exist before the player chose it. AI-generated video is the first technology aimed at that limit.
• Text-to-video makes shots; world models such as GameNGen, Oasis, Muse and Genie 3 generate frames in response to input, which matters more for games.
• Real interactivity needs low latency, long-horizon consistency and reliable control; current systems manage the first in research settings and are still short on the other two.
• The pattern that works for text games will likely work for video: conventional code owns state and rules, and the model renders.
• Build AI video hybrids now: generated cutscenes, branching generated video reviewed before release, video rendered over game state, or searchable clip libraries.
• Treat likeness, consent, disclosure and runtime moderation as design requirements from the first prototype, not afterthoughts.