Is this the thing the industry is angry about?
Fair question, and it deserves a straight answer rather than a reassurance. No, it is not — the transaction runs the other way. But some objections survive that answer, and pretending otherwise would be the fastest way to lose a conversation worth having.
The grievance in games is about work used for AI training without consent and without payment. Licensing a game as a training environment is the opposite transaction: the studio decides, per title, what a named lab may do; the lab pays an integration fee, a monthly service fee and a usage fee; the terms are written down, auditable and terminable; and nothing is scraped, because nothing is taken at all. That answers the main objection. It does not answer every objection, and the ones it leaves standing are worth naming out loud.
Two things that get discussed as one.
Most of the anger in the industry is aimed at the left-hand column. Almost none of the coverage distinguishes it from the right-hand one.
| Scraped for training | Licensed as an environment | |
|---|---|---|
| Who decided | Nobody asked you | You did, per title |
| What was paid | Nothing | Integration, service and usage fees |
| What was taken | Whatever was reachable | One licensed use, named in writing |
| Where it runs | Their infrastructure | Hardware you approved, under your terms |
| Generative output | Unrestricted | Restricted by the licence |
| Can you stop it | Not really | Termination and exit clauses |
| Who can audit it | No one | You can |
Four objections the distinction does not dissolve.
These are the ones that come up in the second half of a first call, usually after someone has been quiet for twenty minutes.
- 1
"A model trained on our game learns from our work."
True, and worth separating into two claims. An agent practising inside your simulator is learning the task, not learning to build simulators — what transfers is driving, planning and recovery, not your level design. The narrower worry is real: a model that has seen your assets could be prompted toward something that resembles them. That is a contract term, not a hope, and it is why the licence should forbid generative reproduction of your assets and characters outright.
- 2
"Our team will hate this."
Some of them might. In GDC's 2026 survey 52% of industry professionals said generative AI is having a negative impact, and the most negative disciplines were artists, designers and programmers. What usually lands is that nothing generative enters your pipeline: no art is generated, no role is replaced, nobody is asked to work with a model. What does not land, and should not be argued away, is a principled objection to the labs themselves.
- 3
"Players will read it as selling out."
Nothing here is public by default and no announcement is required. But secrecy is a poor plan. The durable version is a contract you would be comfortable describing in public: what you licensed, what you refused, what the lab may never do with it. A studio that can say that sentence out loud has nothing to manage.
- 4
"It legitimises them."
This one is not a misunderstanding and we will not talk anyone out of it. If a studio's position is that frontier labs should not be helped, then this deal is the wrong deal and no fee changes that. We would rather hear it in the first conversation than discover it in the third.
Consent is not a formality you sign away once.
The reason the industry is angry is that a great deal of creative work was used without anyone being asked. We think that was wrong, and a licence that is vague enough to be read the same way is not much better than no licence.
So the terms we help studios set are narrow on purpose: one named use, one named lab, no redistribution, no generative reproduction of your assets, deletion at the end of the term, and the right to audit where the build runs. If a lab will not accept that shape, that is information about the lab.
The clauses that enforce it- Nothing generative in your pipelineNo asset is generated, no role is replaced, no artist is asked to work with a model.
- Your build never leaves controlled hardwareIt runs where you agreed, for the lab you named, and is deleted when the term ends.
- No is a complete answerA fit check costs nothing and obliges nothing. Most conversations should end there.
Asked in most first calls
Is licensing a game to an AI lab the same as AI scraping?
No, and the difference is the whole transaction. Scraping takes work without asking and without paying. Licensing an environment starts with the studio deciding, names exactly what the lab may do, pays for it, and can be ended. Nothing is taken; something is sold.
Could a model trained on my game learn to make games like mine?
An agent learning to drive a truck inside your simulator is learning to drive, not learning to build simulators. The transferable worry is narrower: a model that saw your assets could be asked to reproduce something resembling them. That is why the licence should forbid it in writing, and most studios do.
Will my team object?
Some will, and the survey data says the most negative groups are exactly the disciplines you would involve. It helps that nothing generative happens in your pipeline — no art is generated, no job changes, no asset is fed to a model that outputs art. But if the objection is to AI labs existing, a contract does not answer it and leadership has to.
Do we have to announce it?
No. Nothing about an environment licence is public by default. The better reason to keep the contract tight is not secrecy but the opposite: if it does become public, you want to be able to say exactly what you licensed and exactly what you refused.
What if we think this legitimises AI labs?
That is a coherent position and no commercial argument dissolves it. If a studio holds it, it should not do this deal. We would rather hear it in the first conversation than in the third.