Consent Requirements for AI Training Data

A consent chain is the documented trail proving that every person appearing in a dataset granted the rights the buyer needs, before capture. It links each delivered episode to a signed release and records the legal basis, so the license a buyer receives is backed by artefacts rather than an assurance. That distinction matters because training data with murky consent is a legal and reputational liability that tends to surface only after a model has already shipped.

This is general information about how consent is typically documented for commercial training data, not legal advice. Confirm the specific requirements for your jurisdiction and use case with counsel.

Consent and license answer two different questions. Consent establishes that the people appearing in the footage agreed to be recorded and to how the footage would be used — it is about the participants. A license is the separate grant defining what the buyer of the resulting dataset can legally do with it — train commercially, redistribute, sublicense, and so on. A dataset can have a permissive commercial license and still carry an undocumented consent problem underneath it, which is exactly the risk a documented consent chain is built to close.

A model trained on footage with an undocumented or thin consent trail carries a liability that is invisible during evaluation and expensive to discover after deployment — there is no way to retroactively prove a participant agreed to commercial use once the model is already shipping. A verifiable, artefact-backed consent chain is what lets a commercial team actually ship a model trained on human demonstration data with that question already settled.

Every Firsthand contributor signs a written release granting commercial training rights before capture, and each batch ships with the consent artefacts and a data card listing jurisdictions of capture. See /provenance for the details.