Union voice agreements and digital replicas: what a data buyer inherits

Why union terms reach buyers who never sign a union contract

SAG-AFTRA negotiates agreements covering performers across film, television, streaming, games, audiobooks and dubbing. The agreements are made between the union and producers, and a data buyer is not a party to any of them.

That does not keep a buyer clear of them. The restrictions attach to the recordings themselves: a producer that agreed to limits on reuse cannot grant a buyer more than it holds, so a buyer acquiring the audio acquires the limits along with it. The contract between buyer and producer may say nothing about the agreement underneath, which is exactly why the gap is discovered late.

There is a second, softer route. Union terms are the reference point that non-union deals get compared against. When a union agreement requires separate consent and separate payment for a digital replica of a performer, that expectation spreads into the rest of the market within a couple of years — including into corpora that have no union involvement at all.

The provisions themselves differ by agreement and by year, so the version that governed the session is the one that matters. Structurally, the replica clauses tend to do the same five things:

• Require consent before a digital replica of a performer is created, given for a described use rather than as a general permission.
• Require separate compensation for the replica, negotiated on its own terms rather than absorbed into the original fee.
• Limit the scope — often to the production the performer was engaged for, or to a defined category of uses.
• Define what counts as a replica, which is where the disputes will be, because a model trained on a performer is not a copy of any single performance.
• Provide a mechanism to bargain when a producer wants to go beyond what was agreed.

The chain, and where it breaks

Follow the rights from the performer to the data buyer and the failure points become visible:

• The performer consents to a producer, for a described use, under an agreement with a defined term and scope.
• The producer holds the recordings and whatever reuse rights the agreement granted it — which may be narrower than the producer believes.
• A data buyer acquires the recordings, and only what the producer could lawfully grant.
• A model is trained, and outputs reach the buyer's customers, which is usually a use nobody described at the session.

Dubbing is the case that catches people out

Localization work is where the gap shows up most often, because the sessions were commissioned for something specific. A performer hired to dub a production into one language may have consented to that production, in that language, for that distribution window. A corpus licence assumes a much broader grant than the session paperwork supports.

The same applies to audiobook and interactive work, where the engagement is tied to a title or a character. In each case the question is the same: what did the individual performer agree to, in writing, in their own name — as distinct from what the production company agreed to with the studio.

A buyer who receives a corpus described as "professionally recorded voice talent, full rights" is holding a description rather than a chain. The chain is a set of documents, and it either exists or it does not.

What to verify before you licence

The questions, ordered by how much they change the answer:

• Were the sessions union sessions? Ask for the agreement name and the year, and check that the year lines up with the session dates.
• Does the applicable agreement contain AI or digital replica provisions, and what do they require for the use you intend?
• Did the individual performer consent, in writing, to that use — or only to being recorded?
• Does the performer's grant run to the producer only, or to successors and assigns? A grant that stops at the producer is not one a buyer can rely on.
• Where the agreement requires notice to the union before a replica is created, was that notice given?

Contracting around the gap

Where the chain has a hole, the buyer's options are to fix it or to price it, and both are legitimate:

• Return to the performers for a fresh consent that names training and synthetic output. Slow, but it produces a corpus that can be warranted.
• Require a representation from the supplier about union status, with the supporting documents attached as a schedule rather than promised on request.
• Require notice if a performer withdraws or a union raises an objection, so the buyer learns early rather than from a demand letter.
• Keep union-affected material in its own batch or corpus, so a problem in one part does not require unpicking everything.

Where the industry is heading

Union agreements in this area are revised frequently, and each round has tightened the terms around replicas: more explicit consent, more specific scope, more separate payment. The direction is consistent enough that a corpus built on the loosest reading of an old agreement is likely to look deficient against the next one.

For a supplier, that argues for collecting the strongest consent available at the session rather than the minimum the current agreement allows. The marginal cost is a conversation; the alternative is a corpus that cannot be licensed in three years.

This is an operational map for procurement rather than legal advice. Union agreements are revised regularly, the terms in force for a given session govern it, and interpreting a specific agreement is work for counsel.