Industry Transformation

AI in Media and Entertainment: Content Creation to Distribution

How AI moved through media: generative video and music, dubbing, recommendation, and the lawsuits and contracts that followed, from SAG-AFTRA's 2026 synthetic-performer test to the second Sony and Universal suit against Suno.

Gabriele Masetti · 2026-08-05

The tools moved faster than the industry's ability to argue about them

By the middle of 2026, nearly every stage of making and delivering film, television, and music has a generative AI product sitting inside it: text-to-video systems drafting shot sequences before a camera rolls, voice models re-recording dialogue in ninety languages overnight, and recommendation engines deciding what hundreds of millions of people watch next. None of this happened quietly.

Each layer of adoption has arrived paired with a lawsuit, a union fight, or a public backlash, and the entertainment industry's experience with AI over the past three years is less a story of smooth automation than of an industry renegotiating, clause by clause, who gets paid and credited when a machine does part of the work.

Generative video grows up, fast

OpenAI's Sora went from research preview to a genuine production tool in a single generation. Sora 2, released to select users in the US and Canada on September 30, 2025 alongside iOS and Android apps, generates 10-to-25-second clips with synchronized dialogue, sound effects, and ambient audio, and introduced "cameo" characters that let a user insert a verified likeness of themselves into any generated scene.

It didn't last as a standalone consumer brand: OpenAI shut the app down in April 2026 and set 24 September 2026 as the shutdown date for the Videos API and both sora-2 models, listing no replacement on its own deprecation page — a reminder that in this category, product cycles are measured in months, not years.

Runway has taken the opposite path, building deep enterprise relationships rather than a viral consumer app. In September 2024 it announced a partnership with Lionsgate to build a custom video generation model trained on the studio's own catalog of more than 20,000 film and television titles, intended strictly for the studio's internal pre- and post-production use — the first time a major studio handed its own library to an AI lab to build a bespoke model rather than licensing a general-purpose one.

Runway followed with a $308 million Series D in April 2025 at a $3 billion valuation, then a $315 million Series E in February 2026 at roughly $5.3 billion; its Gen-4.5 model, released in December 2025, ranked first among video models on the independent Artificial Analysis Video Arena, ahead of both Google and OpenAI. That lead did not survive the year: by September 2026 Google's Gemini Omni Flash headed the board at an Elo of 1233, Alibaba's Wan 3.0 four points behind, and Gen-4.5 was near the top of neither text-to-video nor image-to-video.

Google's competing Veo line has emphasized native audio and prompt fidelity — Veo 3 in mid-2025 and the Veo 3.1 update that followed — backed by Google's own compute and integration into Google's broader creative tools. Total funding into AI video generation companies reached roughly $3.08 billion in 2025, nearly double the $1.58 billion raised the year before, which tells you how much capital believes this is now core production infrastructure rather than a novelty.

What these tools are actually used for inside real productions is narrower than demo reels suggest: pre-visualization, storyboarding, background plate generation, and — per the Lionsgate arrangement — asset generation the studio owns outright rather than licenses per-use. Full AI-generated theatrical features remain the exception, not the norm, but the pre-production pipeline is already being reshaped.

Music: the fair-use fight that split the majors

No corner of media AI has produced sharper legal conflict than music generation. On June 24, 2024, the RIAA — representing Sony Music Entertainment, UMG Recordings, and Warner Records — filed copyright infringement suits against Suno (in the District of Massachusetts) and Uncharted Labs, maker of Udio (in the Southern District of New York), alleging mass unauthorized copying of copyrighted recordings to train the models. Both companies' defense rested on the same argument: that training on copyrighted works is transformative fair use under 17 U.S.C. § 107, not infringement.

The case then splintered in a way few predicted. By late 2025, Warner Music settled with Suno and Universal Music settled with Udio — both settlements reported to move toward licensed, revenue-sharing models rather than continued litigation, with the UMG-Udio deal in October 2025 described as creating a per-generation licensing template for future deals.

Sony's claims against Udio remain active, and the original Suno case is unresolved more than two years after filing. Suno kept licensing in the meantime — Warner, BMG and Believe all signed — and in September 2026 launched a v6 model family it said was trained on partner-licensed content plus its users' own creations.

Sony and Universal answered on 18 September 2026 with a second, 45-page suit in the same Massachusetts court, arguing the licences cure nothing because v6 was trained partly on the earlier models' outputs. "V6 is not a fresh start; it is the fruit of the same poisoned tree," the complaint says, putting the recordings at issue at 60,202. Suno called the claims "fundamentally flawed on both the facts and the law."

The practical effect is a music industry with no unified position, and a company that is licensing partner and defendant at once.

Either outcome sets precedent that other content industries — video, image, voice — are watching closely, because the legal question is identical: does training a model on copyrighted work require a license.

VFX and de-aging: older than the current AI moment, but now inseparable from it

De-aging is the one corner of this story where "AI" is easy to overstate. For Indiana Jones and the Dial of Destiny (2023), Industrial Light & Magic de-aged Harrison Ford using a toolset the studio called "ILM FaceSwap," combining machine-learning-based reference footage drawn from hundreds of hours of the earlier Indiana Jones films with a full CG head and on-set photography — more than 100 artists worked the shots, and the studio has been explicit that this was not a push-button AI process.

A more automated approach came with Here (2024), where Metaphysic's real-time, in-camera de-aging system altered Tom Hanks and Robin Wright's ages live on set without a traditional VFX post-production pass — the same company behind the viral "DeepTomCruise" deepfakes, now selling the underlying technology directly to studios.

The gap between ILM's artist-heavy hybrid and Metaphysic's live automation marks how unevenly "AI VFX" is distributed: some studios augment craft workflows, others replace pipeline steps outright.

Dubbing and localization: the clearest commercial win, so far

If there's a segment of media AI where the economics are least contested, it's localization. ElevenLabs' Dubbing v2, its upgraded AI dubbing model, translates and re-voices content across more than 90 languages while conditioning on the original audio performance rather than just a transcript — meaning it aims to preserve a speaker's hesitation, urgency, or warmth, not just their words.

For studios and broadcasters, the company pairs the model with ElevenProductions, a service layer adding human translators, professional voice casting, and audio mixing on top of the AI output rather than replacing that human layer outright. The commercial logic is straightforward: human dubbing workflows for a major studio title routinely cost hundreds of thousands of dollars per title per language, and a title that would previously reach a handful of dubbed markets can now reach dozens.

ElevenLabs has rolled out API access to select enterprise clients first rather than to the general public, prioritising studio and platform relationships over broad self-serve growth — the opposite sequencing from the consumer-first path Sora took, and the one still standing.

Recommendation: the AI layer nobody notices because it works

Generative tools get the headlines, but the AI system with the largest actual measured effect on the industry is one that predates the current wave by more than a decade: Netflix's recommendation engine. Netflix's own public estimates have put the value of its personalization system at over $1 billion annually in reduced subscriber churn, and the company has found that personalized recommendations increase content consumption three to four times over simply surfacing the most popular titles.

The least controversial application of AI in media — nobody strikes over a recommendation algorithm — is arguably the one with the deepest structural effect on what gets made, since a platform's greenlighting decisions increasingly follow what its own recommendation data says will retain subscribers rather than what wins critical acclaim.

The 2023 strikes: where the guardrails actually got written

The clearest evidence that Hollywood's labor unions saw this coming is the 2023 WGA and SAG-AFTRA strikes, both fought in significant part over AI terms rather than pay alone. The WGA strike began May 2, 2023, and ended September 24, 2023, with a contract establishing that neither traditional nor generative AI counts as a "writer" — meaning AI-generated text cannot be credited or treated as literary material — while reserving the union's right to prohibit its members' scripts being used to train AI, and requiring studios to disclose when AI-generated material is handed to writers as source material.

SAG-AFTRA's strike ran until a tentative deal was reached November 8, 2023, and its contract created distinct, defined categories for "digital replicas" and "synthetic performers," requiring informed consent and bargained compensation whenever a performer's digital likeness is generated or an existing performance is digitally altered, plus recurring meetings between the union and producers to address AI training and use going forward.

Those provisions have already been tested. In 2025, SAG-AFTRA filed an unfair labor practice charge against Llama Productions over an AI-generated recreation of James Earl Jones's voice used for Darth Vader in the video game Fortnite, arguing the deployment of a deceased performer's voice without bargaining violated the contract and displaced work from living performers — a case that shows the digital-replica rules apply even to work licensed from a performer's estate, not just to living actors' unauthorized likenesses.

2026: a test, not a ban

Those provisions have since been rewritten. On 4 June 2026 SAG-AFTRA announced that members had ratified a four-year TV/theatrical contract with the major studios, by 91.4% of ballots cast on a 19.3% turnout. What replaced the 2023 language is a test rather than a prohibition: a producer may use a synthetic performer only where it brings "significant additional value" compared with a live actor or that actor's digital replica.

The phrase is undefined in the contract, which is where the fight moves next. Producers must give the union notice and a chance to bargain first, and failing agreement the union may arbitrate for damages not capped at what a human would have been paid. A blunter remedy, the "Tilly Tax" — a payment into a union fund for every non-human performer — was a nonstarter for the AMPTP, studio sources told TheWrap.

Likeness, consent, and the fights still unresolved

The strike contracts didn't end the likeness disputes; they mostly defined the terms on which the next ones would be fought. In 2024, Scarlett Johansson publicly objected after OpenAI released a ChatGPT voice, "Sky," that many listeners felt closely resembled her own — OpenAI withdrew the voice, denying it was modeled on her but declining to keep using it amid the dispute, with SAG-AFTRA publicly backing her position.

Separately, voice actors including work associated with A-list names filed a proposed class action against the AI startup LOVO in New York federal court in 2024, alleging their voices were licensed under false pretenses (one plaintiff says he was told his recording was for "academic research only" before it appeared as a commercial synthetic voice) and used without proper consent.

In Germany, a Berlin court ruled in August 2025 that unauthorized AI voice cloning violates personality rights, ordering a YouTuber to pay a professional voice actor €4,000 in damages — one of the first clear European court rulings applying existing personality-rights law directly to AI voice clones. In the US, Tennessee's ELVIS Act, enacted in 2024, became the first state law explicitly extending right-of-publicity protection to AI-generated voice clones, with criminal and civil remedies.

The likeness argument reached its most visible flashpoint in September 2025 with Tilly Norwood, a fully AI-generated "actress" created by Eline van der Velden, after reports that talent agents were in discussions to represent her. SAG-AFTRA responded that Norwood "is not an actor, it's a character generated by a computer program that was trained on the work of countless professional performers," warning producers that using synthetic performers without bargaining would violate union contracts.

Actors including Whoopi Goldberg, Emily Blunt, and Melissa Barrera criticized the project publicly, while van der Velden defended it as "a piece of art" and said dozens more AI characters were already in development — a genuinely unresolved standoff between a union contract that requires bargaining over synthetic performers and a growing supply of AI talent nobody is bargaining to use.

The standoff produced a film rather than a resolution. In July 2026 van der Velden's London studio, Particle 6, said it was in production on "Misaligned", a comedy-drama Norwood will "star" in, shot by human crew using AI tools. Representation proved harder than production: WME and Gersh balked, and van der Velden said the agency conversations had stopped.

What audiences make of all this

The industry's other open question is simpler and less legal: do audiences actually want AI-made content, once they know it's AI-made. Coca-Cola's answer, twice now, has been instructive. Its AI-generated 2024 holiday commercial — built by three AI production studios using four different generative models, riffing on the company's beloved 1995 "Holidays Are Coming" ad — drew immediate backlash for looking "soulless," with viewers spotting distorted lights and nonsensical background shapes within hours of release.

Coca-Cola ran a similar AI-generated holiday spot again in 2025 and drew renewed criticism. The lesson repeated in both cycles: for content explicitly tied to nostalgia and craft — the exact qualities audiences associate with human authorship — AI production is a reputational risk even when the technology executes cleanly, because the objection isn't really about visible errors, it's about what the medium is understood to represent.

The tension that doesn't resolve

Put the pieces together and the pattern is consistent: wherever AI clearly lowers cost or expands reach without touching anyone's specific creative labor or likeness — localization, recommendation, pre-visualization — adoption has been fast and largely uncontested. Wherever it substitutes for a specific, identifiable human contribution — a performer's face or voice, a songwriter's catalog, a visible creative signature audiences associate with craft — it has triggered lawsuits, strikes, contract clauses, or public revolt, often all four.

The RIAA litigation's split verdict — some labels settling into licensing deals while others keep litigating for an outright win — suggests the industry itself hasn't decided whether to price AI training as a new revenue stream or fence it out entirely. Tilly Norwood and the Llama Productions charge showed that a hard-won contract doesn't end the argument; it moves it to case-by-case fights over what counts as bargaining, consent, and a "synthetic performer". The 2026 contract is the next turn of that screw: no ban, just a test expensive to argue and an undefined phrase left to arbitrators for four years. The technology is not slowing down to wait for those questions to settle.