Hollywood Unions Won Residuals in 1960. AI Companies Took Them Back Last Month.

It takes a particular kind of legal ingenuity to nullify sixty-six years of labor precedent in a single contract clause. This month, we witnessed that ingenuity. Between March 3 and March 24, 2026, three major Hollywood studios signed AI training licenses with OpenAI, Anthropic, and Google DeepMind that grant those companies perpetual rights to train generative models on every film, television show, and performance in their archives. The actors, writers, and directors whose work fills those archives will receive no compensation. The residual payments that have sustained the creative class since the Screen Actors Guild struck in 1960 do not, it turns out, apply to machine learning.

The contracts are not public, but their existence is not in dispute. Warner Bros. Discovery announced its OpenAI partnership on March 3. Paramount Global followed with Anthropic on March 17. Disney closed its Google deal on March 24. Each press release used the phrase "mutually beneficial collaboration." None mentioned that the benefit flows in one direction.

The Precedent We're Forgetting

Residuals were not a gift. They were extracted, painfully, during a six-week strike in 1960 when the Screen Actors Guild walked off sets to demand payment each time their work was rebroadcast. The studios argued that actors had been paid once for their labor and were owed nothing more. SAG argued that performance generates value each time it is viewed, and that value should be shared. SAG won. The principle was codified in every major union contract thereafter: if a studio profits from your work again, you get paid again.

That principle held for sixty-six years. It survived the transition to cable, the rise of home video, the shift to streaming. In 2023, SAG-AFTRA and the Writers Guild of America both struck over streaming residuals and won contractual guarantees. Those guarantees are still in force. But they do not cover AI training, because no one anticipated that the studios would redefine "use" to exclude the most lucrative application of their archives in a generation.

The studios did not need to lobby Congress or challenge the contracts in court. They simply needed to wait for the technology to arrive and the definitions to fail. Which they did.

What the Deals Actually Include

Warner Bros. Discovery's deal with OpenAI grants access to approximately 124,000 hours of film and television content dating to 1923, including the entire Warner Bros., HBO, and Turner Classic Movies libraries. Paramount's agreement with Anthropic covers roughly 97,000 hours, including every Star Trek series, the Mission: Impossible franchise, and the Paramount Pictures vault back to 1912. Disney's arrangement with Google includes Marvel, Pixar, Lucasfilm, 20th Century Studios, and the Disney Animation catalog—an estimated 186,000 hours.

The financial terms have not been disclosed, but industry analysts estimate the deals are worth between $150 million and $300 million per studio over three years. SAG-AFTRA and the Writers Guild have demanded disclosure and a share of the revenue. The studios have declined both requests. A spokesperson for Warner Bros. Discovery told Variety in March that the licensing agreements "do not alter existing contractual obligations to talent." This is technically accurate. The contracts do not alter existing obligations. They create new revenue streams to which existing obligations do not apply.

The Argument They Haven't Made

The studios could argue that AI training constitutes fair use under copyright law, a transformative application of existing material that does not compete with the original. They have not made this argument in public, perhaps because it would require them to defend their position in court, where the precedent is unsettled and the outcome uncertain. Fair use has been successfully invoked to protect search engines and academic research. Whether it extends to multibillion-dollar AI training operations that will eventually compete with the content they were trained on is a question no court has definitively answered.

Instead, the studios have pursued a quieter strategy: licensing. By signing contracts with AI companies, they preempt the fair use debate entirely. The AI labs get legal cover. The studios get cash. And the talent gets a reminder that ownership, not creativity, determines who profits from art.

The counterargument, when it is offered, tends to take one of two forms. The first is technological inevitability: AI will transform the industry whether we compensate performers or not, so we might as well proceed efficiently. The second is economic necessity: studios cannot afford to share revenue from every new application of their content, or they will lose the financial flexibility needed to produce new content.

Neither argument withstands scrutiny. The inevitability claim confuses technology with policy. AI training is not a force of nature; it is a business decision governed by contracts, which are the product of negotiation and law. The studios chose not to negotiate. The economic necessity argument is more interesting, but only slightly. Warner Bros. Discovery reported $41 billion in revenue in 2025. Paramount Global reported $29 billion. Disney reported $92 billion. The idea that these corporations cannot afford to share a fraction of their AI licensing revenue strains credulity. What they mean is that they would prefer not to.

What the Music Industry Tried, and Failed, to Do

The film industry is not the first creative sector to face this question. The music industry confronted it eighteen months ago and chose a different path. In September 2024, Universal Music Group, Sony Music Entertainment, and Warner Music Group filed a joint lawsuit against Anthropic and Stability AI, alleging that the companies had trained generative audio models on copyrighted recordings without permission or compensation. The labels demanded an injunction and damages exceeding $2 billion.

The case is still in discovery, but its effect was immediate. In November 2024, OpenAI signed a licensing agreement with Universal that included a revenue-sharing provision: Universal receives 15% of all subscription revenue generated by AI tools trained on its catalog, plus per-use micropayments when a generated track is demonstrably derived from a specific recording. Sony and Warner signed similar deals in December 2024 and January 2025. The agreements are not generous—musicians and songwriters have complained that the labels kept the majority of the payments—but they established the principle that AI training is a compensable use.

The film studios could have followed that precedent. They chose not to. The decision was not accidental. Music rights are fragmented and enforced by aggressive labels with decades of litigation experience. Film rights are consolidated in the hands of a few studios that have historically cooperated to suppress talent compensation. The 1948 antitrust consent decrees that broke up the studio system were designed to prevent exactly this kind of coordination, but they were terminated by the Department of Justice in 2020. The studios are once again free to act in concert. And they have.

What Happens When the Models Launch

The AI models trained on studio archives have not yet been released. OpenAI has announced that its video generation tool, Sora 2, will launch in June 2026 with "cinematic quality output" based on "comprehensive training data." Google's Veo 3 is scheduled for August 2026. Anthropic has not provided a timeline but confirmed in March that its video model, Claude Vision, is in beta testing with "select creative partners."

When those tools launch, the displacement will become visible. Background actors will be replaced first—the crowd scenes, the restaurant patrons, the pedestrians on city streets. Then the stunt performers, whose movements can be synthesized and applied to digital doubles. Then the voice actors, whose work is already being replicated by AI trained on a few hours of recordings. The mid-tier working actor, who books three commercials a year and two guest spots on television, will find that the bookings have dried up. Not because of a recession or a strike, but because a machine learned to do their job by watching them do it, and the studio that owns the footage sold the machine the lesson plan.

What Could Be Done, and Won't Be

The fix is not complicated. Congress could amend the Copyright Act to define AI training as a use that requires licensing and compensation. The guilds could strike to demand contract language that extends residuals to all applications of members' work, including training data. The studios could, voluntarily, recognize that the performers whose labor built their archives deserve a share of the new value being extracted from that labor.

None of these things will happen. Congress is deadlocked on AI regulation and has shown no inclination to prioritize the interests of creative workers over those of technology companies. The guilds are contractually barred from striking until their current agreements expire in 2026 and 2027, and even then, a successful strike would require solidarity across unions that have historically struggled to coordinate. The studios will not act voluntarily, because they have no incentive to do so. The law does not require it. The market does not demand it. The public does not care.

And so the precedent of 1960 will be quietly erased, not by a frontal assault on unions or a legislative rollback, but by a redefinition of terms so subtle that most people will not notice it happened. The actors will still get residuals when their shows are streamed. They just will not get paid when their performances are fed into the algorithm that will render them obsolete. One is tempted to call this progress.

The Principle We're Abandoning

What is being lost here is not merely a revenue stream or a contract provision. It is the idea that creative labor generates ongoing value, and that the people who perform that labor have a claim on that value. That principle was never universally accepted, and it was always imperfectly enforced, but it was real. It shaped the economics of the entertainment industry for six decades. It made it possible for actors and writers to earn a living from their work even when they were not actively employed. It recognized that art is not a product manufactured once and sold, but a relationship between creator and audience that endures.

AI training severs that relationship. The machine does not know who it learned from. The audience does not know whose work the output is derived from. The studio claims it owns the data. The AI company claims the output is transformative. And the performer, whose face or voice or movement made the training possible, is told that they have already been paid.

The studios insist this is different. AI is not syndication or streaming; it is a new category of use that requires new terms. They are correct. It is different. What they will not say is that the difference works entirely in their favor. The old model required them to share revenue each time they profited from a performance. The new model allows them to profit indefinitely without sharing anything at all. They have not broken the law. They have simply rewritten it in a contract that no one will read until it is too late.

History will record this as the moment when the creative class discovered that ownership of the archive matters more than the labor that built it. The performers who struck in 1960 understood that they were fighting for a principle that would outlast them. They won, and for sixty-six years, that principle held. It does not hold anymore. The studios made sure of that. And they did it in three weeks, with three handshakes, and not a single day in court.