Human review in AI dubbing belongs wherever a real voice or face is copied

Prime Video learned this in public. Over one weekend at the end of November 2025, it uploaded AI-generated English and Spanish dubs for the anime Banana Fish, No Game No Life 0, and Vinland Saga, and after the backlash it took down the English dubs but left the similarly bad Spanish ones up. The damage went beyond quality. In some cases the AI tracks had replaced existing dubs, and trade coverage raised questions about whether dubbing rights were exceeded, whether contracts were breached, and whether the AI tracks improperly replaced paid human performances. Forbes put the saving at roughly $75 an hour in voice actor pay, something like $50,000 for the whole show. In the end, Banana Fish was left with no English language option at all.

Four things changed in 2026 that make this an urgent problem rather than a reputational one. SAG-AFTRA's new contracts require consent before a replica dubs a performer into another language. EU deepfake labelling became enforceable in August. A federal digital replica bill is sitting on the Senate calendar. And one of the best-known visual dubbing vendors has just gone into administration. My argument in this piece is simple. Machine translation and voice synthesis are not where studio-scale dubbing fails. It fails when a rights or review gap shows up at a predictable seam.Human review should go wherever a real person's voice or face is being copied, not wherever the language is hard.

A seven-step playbook for AI dubbing and lip-sync clearance

1. Sort the slate by recognisability before picking a pipelinepipelineAll active sales opportunities across the stages of the sales process, together with their combined potential value and probability of closing.View full definition →

Here is the rule to put on the whiteboard: *if the audience can recognise whose voice or face it is, a human with authority signs it off. If they can't, a native speaker checks it. If nobody can produce the paperwork, it doesn't ship.*

Put every title into one of three lanes:

• Red: the title uses a replica of an identifiable performer, either a cloned voice from the original cast or visual lip-sync on their faces. Netflix's production guidance asks for caution with changes that affect "a performance's emotional tone, delivery, or intent." That includes using visual ADR to lip-sync to new audio.
• Amber: stock synthetic voices on a catalog title that already has fans or an existing human dub. Banana Fish belonged here.
• Green: titles with no existing dub, no replica and stock licensed voices. This was the original Prime Video pilot, where AI-assisted dubbing was used for licensed movies and series that would not otherwise have been dubbed, and localization professionals worked alongside the AI to control quality.

The lane decides how much review a title gets, who signs, and which contracts you have to pull. How many languages you are dubbing into decides none of that.

2. Rebuild the chain of title for each new language version

A new language track is a new version of somebody's work. Perso, which sells AI dubbing (so read this as a vendor position), argues that under US law a dubbed version is a derivative work that only the rights holder can authorize, which means you need explicit written permission or a license that covers derivative rights. For licensed titles, read the dubbing clause in the licence for three things: who may produce versions, with which talent, and whether an existing dub may be replaced. Most of these licences were written before voice cloning existed. If the clause doesn't mention AI-generated versions, treat it as silent and ask the licensor.

3. Get consent that names the title, language, territory and term

Talent consent is now built into the contracts. Members ratified SAG-AFTRA's 2026 TV/Theatrical deal by a vote of 91.42% to 8.58%, and the agreement runs from July 1, 2026 to June 30, 2030. The union's own summary says the old contract let producers use a digital replica to perform in another language without consent, and the new one reverses that. According to one guide to the deal, the dubbing consent requirement applies to pictures that start principal photography on or after July 1, 2027. The TV animation contracts ratified this autumn carry the same reversal, and if producers use an independently created digital replica, the actor is owed at least their minimum session fee.

State law tightens this further. In California, AB 2602 makes a replica clause unenforceable for new performances if the replica stands in for work the person would otherwise have done, the clause has no reasonably specific description of intended uses, and the person had no lawyer or union representing them in the negotiation. Deepdub, which sells AI dubbing, recommends terms that specify the productions, languages, territories, and duration covered, and points out that a publicly available recording does not grant a license. Netflix adds that AI models trained for talent manipulation must be specific to one production and never reused on other projects without consent.

A worked example of replica exposure

Take a hypothetical case: a 10-episode series where six principal voices are cloned into four languages. Count each voice, in each language, in each episode as one replica use, and you get 6 × 4 × 10 = 240. The pending NO FAKES Act proposes statutory damages of $5,000 to $25,000 for each violative replica. If every use counted separately, exposure would run from $1.2 million to $6 million. Compare that with the roughly $50,000 Forbes estimated a human English dub of Banana Fish would have cost. The bill is not law, and how replicas get counted would be fought over in court. But the gap between those two numbers explains why the consent file, not the voice model, is what controls your exposure.

4. Clear the music cue by cue, including any translated lyrics

AudioShake, which sells stem separation, notes that dubbing needs clean dialogue stems and an intact M&E (music and effects) track before translation can begin. Its customers have used separation to remove commercial music where licences had expired or were never cleared for international release. That second use is the warning sign. Pulling stems out of a mix does not clear anything. For each territory, check that every sync licence covers the new version, and treat any cue that's sung on screen as a separate item. The Vocal Market, which licenses vocal data, lists what a localized song needs: the original composition, the translated lyrics (themselves a derivative work), the vocalist's performance rights in the new recording, and, if AI singing is used, rights to the training data behind the model. Your music team should bere-clearing each cue for every new territory and version the same way they handle a new sync deal.

5. Put human reviewers at the four seams machines miss

Deepdub's description of QC is a fair checklist: the reviewer checks the mixed version for sync drift, mistranslation, consistency of names and terminology, level problems, and compliance with the platform's delivery spec. I'd add four named owners:

• A native-speaker adapter for the script, because lines rewritten to fit mouth shapes change meaning.
• A dubbing director for performance. What went viral about Banana Fish was flat emotion, not mistranslation.
• Standards and practices for any change in dialogue. Flawless's tools were used on the thriller *Fall* to visually dub out swearing and secure a PG-13 rating, which shows the same tooling can move a rating in either direction.
• Talent and legal sign-off on every red-lane asset.

Write downwho signs off on each release before production starts, not after a clip goes viral.

6. Label and mark for the EU before the file leaves

Under the AI Act, providers must mark AI-generated outputs in machine-readable form, and deployers must tell viewers when content is a deepfake. These obligations have applied since 2 August 2026, and the Commission reads the territorial scope broadly: posting content on the globally accessible internet may trigger the labelling obligation, even if it isn't aimed at the EU market. Content created before 2 August 2026 does not need labelling, and systems already on the market have until 2 December 2026 to meet the machine-readable marking requirement. The Act lightens disclosure for works that are evidently artistic or fictional, but it doesn't remove it. Decide where your label goes (the audio track menu, the credits, or both) and record that decision. Prime Video's tracks went wrong partly because viewers found "AI beta" audio tracks attached to major anime titles without clear warnings.

7. Write vendor failure into the contract

Deadline reported on 30 October 2026 that Flawless, which spent about $100 million building SAG-AFTRA-endorsed visual dubbing tools, sold its intellectual property after its UK trading company went into administration. Its technology has been used on hundreds of Hollywood productions, though much of that work stayed quiet because studios declined to disclose their use of AI. If you hold voice or face models of your cast, ask where those models live now and who owns them. SAG-AFTRA's new terms protect performers from losing digital replica protections when ownership changes hands. Your vendor contracts should match that with three clauses: notice on any change of control, deletion or return of production-specific models, and no transfer of talent replicas without fresh consent. Netflix also requires that generative tools do not store, reuse, or train on production inputs or outputs.

Where do AI dubbing rollouts go wrong?

They go wrong when a team treats localization as a cost decision and ships without telling the audience, the talent or the licensor. These are the failures I see most often:

• Swapping out an existing human dub. Fans notice straight away, and licensors read it as going beyond the rights they granted.
• Counting language pairs instead of faces. A Japanese-to-English job with one cloned lead is riskier than a six-language job using stock voices.
• Treating a vendor's "commercial use" terms as clearance. A tool licence covers the tool's output. It doesn't cover your cast's voices or your licensor's music.
• Planning around the NO FAKES Act as if it were law. As of mid-October, one tracker notes the bill is not law, and state law and contracts govern cloned voices and faces today.
• Going straight to replacement and skipping licensing. Talent may sell you a clone at the right price: Voices, a voice marketplace, told Rest of World that voice AI jobs pay up to 85 times as much as traditional voice-over work. That's the company's own claim, but it tells you which way the market is moving.

What to do on Monday for your localization slate

• Tag every title in the 2027 localization plan red, amber or green using the recognisability rule.
• Ask business affairs for every talent agreement with a replica clause signed since 1 January 2025, and test each one against AB 2602's three conditions.
• Send each dubbing and lip-sync vendor three written questions: where your models are stored, what happens to them on insolvency or sale, and whether your inputs train their systems.
• List the AI-voiced tracks published in the EU after 2 August 2026 and confirm each one carries a label.
• Pull the cue sheets for your five most-localized titles and flag every song with sung or on-screen lyrics.
• Name one person per title who signs the release, and put that name in the delivery spec.

The cheapest point to clear a cloned voice is before the session that creates it. Once a dub is live and fans are clipping it, fixing it costs far more than the saving that justified it, as Prime Video found when the response to its Banana Fish dub was to pull the track.

The full course on this sector:AI in Media & Entertainment.