Gameplay data or a training environment?
Two different products, routinely discussed as if they were one. The distinction decides what you are paid, how often, and which rights you have quietly given away.
Gameplay data is recorded play: files of what human players did, which a model learns from by reading them. A training environment is a playable copy of the game in which the lab's own agent takes actions, sees consequences and tries again. The first is finished the moment it is delivered; the second keeps producing new situations as long as it runs. For a studio that is the difference between a one-off fee and a recurring line — and they are separate rights, so one contract should not silently include the other.
| Gameplay data | Training environment | |
|---|---|---|
| What it is | Recorded play, as files | A running copy of the game |
| What the model does | Reads what humans did | Acts, fails, retries |
| Who it suits | World models, video prediction, imitation learning | Agents learning to act and plan |
| Delivery | A transfer, once | Hosting and operation, continuously |
| Revenue shape | A fee against an existing asset | Integration, monthly service, usage per agent |
| Studio effort | Export and clearance | Engineering, paid for by the lab |
| Exhausts when | The lab has read it | Not really — new seeds make new situations |
| Main risk | Rights clearance in the footage | Scope creep in the licensed use |
An agent cannot learn to recover from a mistake it only read about.
Recorded play is mostly competent play. It shows an agent what success looks like and almost never what the twelve ways of failing look like, or how to climb out of them. Training to act requires being wrong and living with the result, which only a live world can provide.
That is why the two markets coexist rather than compete. Recordings teach a model what a world looks like; an environment teaches an agent what it does back.
These are different grants, and one can swallow the other.
A data licence written as "use of the game for AI training" can be read as covering environment rights too. If you intend to license an environment later — possibly to a different lab — the data agreement has to say what it does not include.
The reverse matters as well: an environment licence should name whether the lab may keep and reuse the trajectories its agents generate. The contract checklist covers both.
Environments, not data.
We build and operate the playable side: the game made agent-ready, one isolated copy per lab, with scored tasks and metering. If a lab wants recordings as well, that is a separate agreement with the studio and we are not in it.
Request a fit check- Recurring rather than onceUsage pricing keeps paying while the lab trains.
- Repeatable across labsThe same world, licensed again on its own isolated copy.
- No footage to clearNothing is exported, so there is no recorded-player or music clearance question.
About the two products
Is licensing gameplay data the same as licensing a training environment?
No. Gameplay data is recorded play — a finished asset a model learns from by reading it. An environment is a playable copy of the game in which the lab’s own agent acts, fails and tries again. One is a sale, the other is a service with recurring revenue attached.
Which one pays better?
Data pays sooner; an environment pays longer. A data licence is usually a fee against an asset that already exists. An environment carries an integration fee, a monthly service fee and a usage fee per concurrent agent, so it grows with how hard the lab trains.
Can a studio do both with the same game?
Usually yes, but the contracts have to say so. They grant different rights, and a data licence written loosely can swallow the environment rights you would want to sell separately later.
Which do AI labs prefer?
It depends on the research. Teams training agents to act want environments, because an agent cannot learn to recover from a mistake it has only ever read about. Teams training world models and video-prediction systems often want recorded play as well.