One of these is a line to cite. The other is a gap to explain.
European rules now ask model providers to say publicly what they trained on. That turns the provenance of training material from an ethical argument into a documentation problem — and documentation is exactly what a licence produces.
Providers of general-purpose AI models placed on the EU market have to publish a sufficiently detailed public summary of the content used for training, following a template issued by the European Commission's AI Office, and to keep a policy for complying with EU copyright law. The template distinguishes content obtained commercially through agreements with rights holders from content scraped from the web, which has to be described by its main sources. A licensed game environment falls in the first category: a named agreement with a named rights holder. It is one entry. Scraped material is a list and an explanation.
Provenance stopped being a matter of taste.
Until recently, where training material came from was an argument between a lab and its critics. Making it a published document changes who the audience is: regulators, customers, journalists and the rights holders themselves can now read the same page.
The obligation follows the market rather than the lab. A model placed on the EU market carries the duty with it, whoever built it and wherever the training ran — which is why this is a live question for American labs with European users, not only for European ones.
It is in force, and the grace period is for old models only.
According to the legal commentary cited below, the transparency duties have applied since 2 August 2025, the AI Office's enforcement powers began on 2 August 2026, and models already on the market before that date have until 2 August 2027 to publish in line with the template. Summaries are expected to be kept current, including after further training.
Dates and penalties here are drawn from law-firm analysis rather than the Official Journal. Treat them as orientation and have counsel check the text before relying on them.
Three things a licence gives you that a crawl does not.
- 1
A counterparty with a name
A licence identifies the rights holder, the titles, the versions and the permitted use. Everything a disclosure needs is already written down, because it had to be written down to sign the deal.
- 2
A defined scope
Scraped material raises the question of what exactly was taken and under what reservation of rights. A licence answers it in advance: this use, this term, these limits.
- 3
An answer that does not change
Summaries have to be kept current. A licensed source is stable across updates; a crawl is a moving description of a moving thing.
A licence is not a compliance product.
Using benned does not make a model provider compliant, and anyone selling it that way is overselling. Compliance is the provider's own work across its whole corpus, and one licensed environment is one line in a long document.
The honest claim is narrower: for the part that came from a licensed game, the provenance is documented, attributable and stable. That is worth something precisely because it is not everything.
Request a pilot- For labsOne named agreement instead of a provenance question you answer in public.
- For studiosYou are the rights holder in that agreement — the position these rules are written around.
- For bothThe paper trail exists either way. The only question is whether it was made on purpose.
About the rules
Does the EU AI Act require labs to say what they trained on?
Providers of general-purpose AI models must publish a sufficiently detailed public summary of the content used for training, following a template issued by the European Commission’s AI Office, and must maintain a policy for complying with EU copyright law. It applies to models placed on the EU market regardless of where the training happened.
Does a licensed game environment have to be disclosed?
If it formed part of what a model was trained on, it belongs in the summary. The difference is that the template has a place for exactly this: content obtained commercially through agreements with rights holders. A named licence is a straightforward entry rather than something that needs explaining.
Does this apply to a lab outside the EU?
The obligation follows the market, not the lab. A model placed on the EU market brings the duty with it, whoever built it and wherever it was trained. For a US lab this is a European compliance question attached to a European customer base.
Does using benned make a lab compliant?
No, and nobody should sell it that way. Compliance is the model provider’s own work across its whole training corpus. A licensed environment makes one part of that corpus documented and attributable, which is a smaller claim and a true one.
What does this mean for a studio?
That the licensed route is the one with a paper trail on both sides. The studio is the rights holder in that agreement, which is the position the rules are written around — and a documented licence is the positive version of reserving your rights rather than the defensive one.
Where this comes from
This page is a plain-language summary of published legal commentary, not legal advice, and not a substitute for the Regulation itself. The primary texts are Regulation (EU) 2024/1689 and the AI Office's template for the public summary of training content.