We test these tools. These are the things we will not do.

A published, specific list of refusals is the only thing that makes the rest of this section worth reading. Each one is grounded in a position somebody else has published, so you can check it.

Last checked September 2026. Nothing on this page is legal advice. The grounds for each refusal are quoted and dated on our record of what the industry and the unions have said.

We will not create, train on or use a synthetic version of a real person’s face, body or voice without that person’s specific, informed, written agreement, freely given, paid at a rate they have agreed, and revocable. Consent for one production does not travel to another and is not implied by somebody having walked into shot.

The ground: SAG-AFTRA’s agreement ratified in June 2026 requires consent and an “articulable business reason” before a performer is scanned for a digital replica; Netflix’s guidance of August 2025 requires explicit documented consent. Voice cloning starts at $6 a month on ElevenLabs’ published price in September 2026, so no policy here can rest on cost as a barrier.

There is a difference between a tool that makes a job faster and a tool used to avoid employing the person who does the job. Where a client asks us to replace a paid person with a generated substitute on a project we are running, we decline the work.

The ground: Netflix’s fifth principle is that generative AI should not replace or generate union-covered work without consent. The honest counterweight: Bectu has published no member-facing AI guidance, and its Big Survey of August 2025 found around half the television workforce out of work while naming the post-strike contraction, commissioning collapse and freelance insecurity — not AI. The contraction came first.

We will not put a production’s unreleased scripts, rushes, stills, cast information or schedules into any service whose terms allow the provider to train on uploads, or whose data handling we cannot establish in writing.

The ground: Netflix requires that tools used on a production cannot store, reuse or train on production data inputs or outputs. This is where most real-world breaches happen, and they happen through defaults rather than decisions.

We prefer tools trained on licensed or clearly consented material, and the register says plainly when a vendor states nothing. We will not build a model on other people’s work we do not have the right to use.

The ground: training a model in another artist’s style without cleared rights is the first thing Netflix names as needing written approval. The BFI and CoSTAR Foresight Lab report of June 2025 found more than 100,000 film and television scripts already used in training without payment or permission, and after the UK government’s decision in March 2026 there is no text-and-data-mining exception to shelter behind.

The refusals

Next