Which games fit, and which ones labs keep passing on.
The question every studio asks first, answered before you spend a meeting on it. Three properties decide almost everything, and none of them is how well the game sold.
Three things, in this order: realism, so that what an agent learns transfers beyond the game; breadth, meaning many situations and many valid ways to reach a goal; and customisability, so maps, rules, weather and starting states can be varied and agents cannot simply memorise one world. Release date and sales figures barely matter — a back-catalogue simulator with deep systems is usually worth more to a lab than a recent hit without them.
What each property is actually asking
- 1
Does it behave like reality?
Not whether it looks real. Whether a heavy load takes longer to stop, whether rain changes the outcome, whether a mistake costs something. A lab is buying consequence, and consequence is a simulation property, not a rendering one.
- 2
Can it be played well in more than one way?
A world with one optimal line gets solved and stops teaching. A world where a good plan can fail for four different reasons keeps producing signal long after the agent is competent.
- 3
Can the world be changed?
Through an editor, config files, mods or engine access — any route counts. Labs vary conditions per seed so agents generalise instead of memorising. A world that is always the same is worth training in once.
Where the demand sits today
A rough map, not a verdict. A deep strategy game beats a shallow simulator, and an exception in a weak category is still an exception worth a conversation.
| Genre | Fit | Why |
|---|---|---|
| Driving, flight and transport simulators | Strong fit | Physics that transfer, long routes, continuous consequence. |
| Construction, farming and logistics | Strong fit | Scarce resources, uncertain timing, plans that fail slowly. |
| City builders and management sims | Strong fit | Many interacting systems, no single correct strategy. |
| Economic and trading simulations | Good fit | Long horizons and adversarial dynamics, if the market is deep. |
| Open worlds with real physics | Good fit | Depends on whether the systems or the story carry the game. |
| Strategy games with hidden information | Depends | Valuable when the simulation is deep, thin when it is a ruleset. |
| Shooters, platformers, story games | Weak fit | Authored moments and reflex leave little to generalise from. |
| Puzzle and casual titles | Weak fit | Already solved, or too narrow to be worth hosting. |
Four things studios worry about that labs do not.
Most of the reasons a studio talks itself out of this conversation turn out not to be selection criteria at all.
Request a fit check- How it soldCommercial performance measured whether players wanted it. Labs are asking something unrelated.
- How it looksDated art is not a problem. Many labs run on structured game state and never see the screen at all.
- Whether the team still existsA shipped build and the rights to license it are enough for a black-box environment.
- Whether anyone still plays itTraining copies run on separate hardware with no connection to live servers or players.
About fit
What kind of games do AI labs want to train on?
Realistic simulations with many situations and many valid strategies: driving, flight, transport, construction, farming, logistics, city building, and management or economic simulations. The world has to behave enough like reality that what an agent learns transfers out of it.
Is my game too old to license for AI training?
Almost certainly not. Labs select on depth of simulation, not release date. A ten-year-old simulator with real systems underneath it is usually a better environment than a recent hit without them, and its commercial life is over, which makes the licence easier to grant.
Do shooters, platformers or story games fit?
Rarely, as training environments. They are built around authored moments and reflex, so there is little for a planning agent to generalise from. Some labs do want them for perception or world-model work, but the demand is narrower and the fees follow.
What if my game has no procedural generation?
It can still fit. What matters is whether maps, rules, weather, difficulty and starting states can be varied at all — through a level editor, config files, mods or engine access. Without any variation agents memorise one world, which limits what the environment is worth.
How do I find out whether my game qualifies?
benned runs a free fit check under NDA: we assess realism, breadth of situations and actions, how customisable the world is, and whether the technical route is black box or engine integration. It takes a conversation and a build.