What consent for AI training actually looks like

Specific: consent to the use you are actually making

Consent has to name the use. Agreeing to appear on a podcast is not agreeing to train a model. Agreeing to research is not agreeing to a commercial system.

Data-protection law makes the same point. Valid consent under the GDPR is specific and unambiguous, given for defined purposes. A blanket permission to do anything with a recording is weak precisely because it is not specific. Name the purpose: training, developing, and evaluating AI models, including commercial ones.

Informed: the person understood what they agreed to

Informed consent means the person knew what was being collected, why, and what would happen to it. For voice data this matters more, because a voice can be used to build a model that speaks.

Some jurisdictions make this explicit. Illinois’ Biometric Information Privacy Act requires written notice and a written release before a company collects a voiceprint. The theme is consistent across regimes. Tell the person plainly, then get their agreement. Consent buried in dense terms nobody reads is not informed consent.

Auditable: you can prove it later

Consent you cannot evidence is, for practical purposes, consent you do not have. A buyer diligencing a dataset will ask how it was obtained. “Trust us” is not an answer.

Auditable consent means a record: who agreed, to what wording, when, and for which asset. It travels with the material. When a rights question arises months later, the record answers it. This is the difference between a claim of consent and proof of it.

Two permissions: the content licence and the voice

A common mistake is to treat consent as one thing. It is usually two.

The content licence is the right to use the recording — the file, the words, the performance. Voice and likeness consent is the right to use an identifiable person’s voice or image. They can belong to different people. The podcaster may hold the recording; the guest holds their own voice.

Durable consent captures both. On fiund, the AI-training licence sits on the asset, and where a person is identifiable, voice and likeness consent is handled separately. Owners keep ownership and approve buyers. Provenance records tie the two permissions to the file so a buyer can see the whole chain.