Licensed Game Data for AI Training: Rights, Provenance, and Delivery

Licensed game data for AI training is customer-authorized source material — code, 3D assets, audio, virtual worlds, and derived captures — delivered with a license manifest, provenance, and a redistribution story that a second team can audit. It is not “we found a ROM” and it is not “the engine EULA seems fine.” This article is operational, not legal advice. Counsel still has to read the actual contracts. The dataset still has to make those contracts visible.

Without that visibility, a technically excellent capture becomes unusable the moment a model ships, a partner asks for a chain of title, or a marketplace asset’s training clause disagrees with your mix.

Authorization is per layer

Games are stacks. Authorization to capture RGB from a build is not authorization to train on the soundtrack, the marketplace tree, or the unreleased character pipeline.

Customer authorization should name the legal entity, the build, and the purpose (training, evaluation, demonstration). Vague “research use” will be interpreted differently by every later buyer of the model.

License manifest

A manifest is a table, not a paragraph in a slide.

{
  "dataset_version": "harbor-wm-0.3.1",
  "customer": "redacted",
  "purpose": ["internal_training", "internal_eval"],
  "redistribution": "none",
  "items": [
    {
      "id": "mesh.dock_crate",
      "class": "3d_asset",
      "source": "customer_internal",
      "training": true,
      "redistribute": false
    },
    {
      "id": "sfx.wave_loop",
      "class": "audio",
      "source": "marketplace",
      "sku": "AUD-441",
      "training": false,
      "redistribute": false
    }
  ]
}

If an item cannot be classified, it does not enter training. Silence is not a license. For open benchmarks, read the actual dataset card; “released on Hugging Face” is not a rights opinion. Khronos’ glTF 2.0 standardizes file structure, not copyright.

Provenance

Provenance records origin and transformations: who captured the episode, which build SHA, which fork of the annotation schema, whether a VLM captioned the chunk, whether frames were transcoded. World-model pipelines that add CoT captions (ActWorld) create a new copyright-and-error layer. Derived labels need their own provenance even when the pixels were authorized.

Failure mode: mixing customer traces with publicly scraped gameplay “to boost diversity.” You have destroyed the authorization story for the whole mix unless the scrape is isolated and documented — and scraping is usually the wrong plan anyway.

Redistribution and confidentiality

Delivery modes differ:

• On-customer-prem capture, no export. The dataset never becomes your asset.
• Restricted export. Traces leave, identifiable assets do not.
• Train-only weights. Even derived data stays closed.
• Eval-only public subset. Requires a second, harsher rights pass.

Confidential frontier work often looks like the first two. That is compatible with strong models and incompatible with “we will open-source the replay.” Say so in the manifest.

Do not confuse open research corpora (GameFactory, VWE-Bench, and similar) with a customer title. Those corpora have their own cards and still need item-level reading. See also training-ready game datasets.

Delivery that can be audited

Ship:

1. The data.
2. The schema.
3. The manifest.
4. A changelog.
5. A known-issues list (unclear SKUs, missing stems, watermarked debug views).

QA is matching IDs: every mesh in a scene graph exists in the manifest; every excluded audio stem is actually muted in the delivered video. A delivery that includes muted audio in the mux but marks audio as excluded is a future incident.

Common failure modes

A practical intake questionnaire — still not legal advice — that stops a program from starting blind:

1. Who is the authorizing entity, and can they bind the listed IP?
2. Which builds and locales are in scope?
3. May we train, evaluate, distill, and/or demonstrate?
4. May traces leave the customer VPC?
5. Which asset classes are excluded (music, faces, UGC, unreleased heroes)?
6. What happens to checkpoints if the contract ends?

If any answer is “we will figure it out later,” the dataset is not licensed. It is delayed risk.

Confidentiality versus openness

Frontier customers often need the opposite of a public corpus. That is compatible with good science inside the walls: versioned schemas, holdout scenes, replay hashes. It is incompatible with copying an AlayaWorld or VWE-Bench recipe and shipping the customer’s GameVerse equivalent to Hugging Face.

When a public research corpus is used as a method reference, keep it in a different namespace from customer bytes. Mixing namespaces is how provenance dies.

What this page is not

This is not a license. It is not an opinion on fair use, trained-weight extraction, or any jurisdiction’s AI act. It is a description of fields a serious game-data program should not omit. For the technical capture layer, use game data for world models. For meshes, use 3D asset datasets.