The pitch is immediately appealing. Instead of a shopkeeper with eleven recorded lines, a shopkeeper you can talk to about anything, who remembers you, who reacts to what you did in the last town.
Demonstrations of this are genuinely impressive. Shipped implementations have been rarer and more modest, and the reasons are not technical limitations that will pass. They are design problems.
The four problems
1. Infinite dialogue is not the same as good dialogue
Written game dialogue is compressed, purposeful and edited. Every line does work: establishing character, conveying information the player needs, setting tone, or being funny. A writer cuts the eleven lines that do not.
Generated dialogue is fluent, plausible and unedited. It fills space rather than doing work. Players describe the effect consistently: the conversation is impressive for about ninety seconds and then feels weightless, because nothing said matters.
The scarce resource in a game was never quantity of dialogue. It was dialogue worth hearing.
2. A character who can say anything can say the wrong thing
This is the objection that stops most commercial deployments, and it is a serious one.
If a character is generated at runtime, the studio cannot guarantee what it says. That creates exposure that has nothing to do with quality:
- Age ratings depend on content classification. A system that can produce unpredictable content is difficult to rate, and a rating breach is a commercial catastrophe.
- Players will deliberately try to make characters say offensive things, and will publish the results.
- Licensed characters have contractual constraints on portrayal. A licensor will not accept a character who improvises.
- Localisation review assumes reviewable text. Generated text cannot be reviewed before it exists.
Mitigations exist — constrained generation, output filtering, restricted topic scope — and every one of them narrows the range that made the feature interesting in the first place.
3. It breaks the contract between player and game
Games teach players what matters through consistency. If a character responds to something, it is significant. If a door can be opened, there is something behind it.
A character who responds meaningfully to everything destroys that signal. Players spend an hour discussing farming techniques with a blacksmith, discover it changed nothing, and correctly conclude that conversation is decoration. Worse, they may miss the actual quest, because it no longer looks different from the improvisation.
Authored dialogue is a filter. Removing the filter does not add freedom; it removes information.
4. Latency and cost, which are unglamorous and decisive
A conversation needs a response inside roughly 300 ms to feel like a conversation. A cloud round trip plus generation plus voice synthesis is frequently slower, and if you add a visible thinking pause you have made an interaction that is worse than instant canned lines.
The cost side is worse. A studio with two million players, each holding a few conversations per session, is looking at an inference bill with no ceiling, on a product usually sold once. Subscription and live-service titles have a mechanism to fund it. A sixty-pound single-purchase game does not.
Generation is good at variety and bad at significance. Games run on significance. Any successful use of this technology has to add variety in places where significance is not required.
Where it does work
Which points directly at the applications that are shipping.
Ambient life. Crowd chatter, market vendors, guards commenting on weather, overheard conversations in a tavern. Currently these repeat until players can recite them. Generated variety here is a straightforward improvement, because none of it is load-bearing. If a guard says something slightly odd, nothing breaks.
Reactive barks. A character commenting on the specific thing that just happened — your outfit, your last fight, the fact that you are on fire. Short, contextual, low stakes.
Repopulating a persistent world. Massively multiplayer worlds have enormous surface area and small writing teams. Filling the gaps between authored content is a real problem that generation genuinely helps with.
Development tooling. This is where most of the actual value has landed, and it gets the least attention. Placeholder dialogue, first-draft barks, localisation drafts, generated test conversations, quick voice mockups for prototyping. Nothing ships as generated, but the writing team moves faster, and the shipped writing is better because the authors spent their time on the lines that mattered.
Simulation depth rather than dialogue. Characters with generated goals, schedules and relationships, expressing themselves through behaviour rather than speech. Arguably the most promising direction, because behaviour is easier to constrain than language and reads as more meaningful.
The hybrid pattern that seems to work
Studios that have shipped something successful converge on roughly the same architecture.
- Authored spine. Every plot-relevant line is written, reviewed, recorded and fixed. The story is not improvised.
- Generated leaves. Flavour, ambience and reactive commentary are generated within tight constraints.
- A clear visual or audio distinction so players learn which conversations matter. This sounds inelegant and is essential.
- Generation constrained to selection where possible. Rather than producing free text, the model picks from a large authored pool and adapts it lightly. All output is reviewable in advance.
- Local models for latency and cost. Small models running on the player's hardware, which removes both the round trip and the per-conversation bill.
- Full offline generation. Increasingly common: run generation during development, ship the results as content. You get variety with none of the runtime risk. Not really generative gameplay, and it produces the best current results.
What to watch for
Two developments would change the analysis.
Capable small local models. If a model good enough for character dialogue runs comfortably on a console alongside the game, the latency and cost objections both disappear. Progress here has been fast, and this is the constraint most likely to fall.
Reliable constrained generation. Techniques that make it genuinely impossible for a character to leave defined boundaries — not filtered afterwards, but structurally unable to. That would address the ratings and licensing problem, which is currently the hard commercial blocker.
Neither would solve the significance problem, which is a design issue rather than a technical one. That one requires designers to work out how to give players freedom of expression without teaching them that nothing they say counts. It is an interesting problem, and nobody has solved it yet.
The realistic near future
Games with noticeably livelier background worlds, characters who acknowledge specifics about you, and far less repeated dialogue. Plots, companions and the conversations you remember will remain written by people, because those need to mean something, and meaning is authored.
The technology will most likely have its largest effect where players never see it: in the tools, shortening the distance between a writer's idea and a playable scene. That is a real contribution. It is just not the demo.