The Copyright Question Hollywood Can No Longer Dodge

An AI-generated establishing shot can today be photorealistic, cinematically lit, and quickly produced. For the marketing team that is a win. For the legal department, the E&O insurer, and international distribution, the real work begins right here. Because an image that looks good is far from an image you may own, license, insure, and exploit worldwide.

This gap between visual persuasiveness and legal robustness has become the central question of the industry. It is not decided in the editing room, but in the chain of title, in the insurance application, and in the distribution agreement, and it can no longer be postponed.

Why a beautiful asset is not yet a safe asset

The core of the problem lies in training. Generative models come into being by processing vast amounts of data, including protected works. The U.S. Copyright Office classifies this clearly in Part 3: “several stages in the development of generative AI involve using copyrighted works in ways that implicate the owners' exclusive rights” (U.S. Copyright Office, Part 3: Generative AI Training). Whether this use is permitted depends in the US on the fair-use test, and the agency makes clear that there is no blanket exemption: “some uses of copyrighted works for generative AI training will qualify as fair use, and some will not” (U.S. Copyright Office, Part 3). Training is indeed often transformative, but the assessment depends on the individual case: on work, source, purpose, and output controls; the knowing use of piracy-based datasets weighs against fair use (U.S. Copyright Office, Part 3).

For the production this means: the legal question already sits in the model, not first in the generated image. If no one can prove which works a system was trained on, a residual risk remains that would ultimately have to be paid for by the buyer, distributor, or investor.

A second problem concerns ownership of the result. In the US the rule is: “copyright protection in the United States requires human authorship” (U.S. Copyright Office, Part 2: Copyrightability). A purely prompt-generated result is, in the agency's view, not protectable, because “prompts alone do not provide sufficient human control to make users of an AI system the authors of the output” (U.S. Copyright Office, Part 2). Protection arises only where humans visibly shape the work: according to the agency, authors can claim rights in the creative selection, arrangement, or editing of the results (U.S. Copyright Office, Part 2). Anyone who fails to document this risks an ownership gap in the middle of the work.

It is precisely here that a legal detail turns into an economic risk. The chain of title is the seamless proof that a production controls all rights to all elements, and it is usually a prerequisite for the errors-and-omissions insurance, which in turn is usually a prerequisite for distribution. Trade accounts describe how underwriting relies on this documentation; whether and on what terms cover is granted, however, depends on policy, exclusions, disclosure, and underwriting (Thoolie, Errors & Omissions Insurance for Filmmakers, trade account, no insurer commitment). Gaps can thereby trigger exclusions, follow-up questions, higher premiums, or a refusal.

An AI element without documented human authorship and without a cleared training provenance is thus a potential break in this chain. The AI Tool Kit of the Archival Producers Alliance, hosted by the Producers Guild of America, therefore recommends a “Cue Sheet” that tracks all AI-generated content and its metadata, as well as transparency toward the team, distributors, and partners; it explicitly states that this information is also useful for taking out E&O insurance (Producers Guild / Archival Producers Alliance, AI Tool Kit). In parallel, the insurance market is moving: as of January 1, 2026, the Insurance Services Office (Verisk) introduced standardized exclusions for generative AI in general liability (CGL) (forms CG 40 47, CG 40 48, CG 35 08), where the broadly construed wording “arising out of” can apply even with indirect AI involvement (Fenwick, The End of 'Silent AI'?). According to the source, these exclusions concern general liability and do not prove that every film E&O policy excludes AI; individual tech E&O policies, however, can exclude precisely the AI-generated risks such as IP infringements, so that a promised indemnity could be left without cover (Honigman, The AI Insurance Gap). Insurability is therefore no automatic given: AI assets can trigger additional queries, conditions, or costs.

This becomes controllable through documented asset provenance. The open standard C2PA describes, with Content Credentials, a cryptographically signed manifest that records who created or altered a piece of content, when, and with which tool; at the same time the standard itself draws a limit and delivers “no value judgments about whether the provenance data is good or bad” (C2PA, Content Credentials Explainer). Provenance is a documentation tool, not proof of law: it supports a production log but replaces neither a license nor rights clearance.

The Midjourney case study: when the output itself becomes the evidence

How concrete these risks are is shown by the first major Hollywood AI case. In June 2025, Disney and Universal sued Midjourney and called the service a “bottomless pit of plagiarism” because it produces unauthorized copies of protected characters such as Darth Vader and Elsa (AP News, Disney and Universal sue AI firm Midjourney). In September 2025 Warner Bros. Discovery joined, spoke of “brazen theft” with characters such as Superman, Batman, and Bugs Bunny, and demanded up to 150,000 US dollars per infringed work (Mashable, Midjourney pushes to expose studios' own AI practices). Midjourney defends itself with fair use and argues that the studios internally use comparable methods; within the discovery dispute, a magistrate judge in mid-June 2026 limited the studios' disclosure obligations to consumer-facing AI products (Mashable, Midjourney pushes to expose studios' own AI practices). As of July 2026 the case is pending; a ruling is still outstanding.

The decisive point for producers: not only the training data but also the output can become a problem when it recognizably reproduces protected characters or works, the difference between a model that shows “a space helmet” and one that recognizably delivers a protected character.

Precisely for AI style transfer a distinction is important. Under US law a style as such is not automatically protected; the court in Kadrey v. Meta found: “style is not copyrightable—only expression is” (Loeb & Loeb, Kadrey v. Meta). Anyone who asks a system to work “in the style of” does not, therefore, infringe rights on that basis alone. It becomes risky when the output takes over concrete protected elements of expression or recognizable characters, or when AI products displace a market for works of the same kind; the U.S. Copyright Office warns of a “serious risk of diluting markets for works of the same kind as in their training data” (U.S. Copyright Office, Part 3). This assessment concerns the US legal situation and cannot be transferred wholesale to other countries.

One legal space, three answers: US, EU, UK

For internationally exploitable films a national view is not enough: three key reference markets treat the same technology differently.

In the US the line is taking shape through court decisions. In Thomson Reuters v. Ross the court decided in February 2025 that the use was not fair use: “Ross' use of the headnotes did not constitute fair use as a matter of law” (Loeb & Loeb, Thomson Reuters v. Ross Intelligence). The judge emphasized expressly, however: “only non-generative AI is before me today” (Loeb & Loeb, Thomson Reuters v. Ross). In Kadrey v. Meta the court in June 2025 rated the training as fair use, but allowed the “market dilution” theory as a possible future point of attack (Loeb & Loeb, Kadrey v. Meta). And in the Anthropic matter, Judge Alsup preliminarily approved a settlement of 1.5 billion US dollars on September 25, 2025; he had previously classified the training as fair use but rated the storage of more than seven million piracy-based books as an infringement (Reuters, US judge preliminarily approves $1.5 billion Anthropic settlement). In the music field, too, the picture is mixed: the RIAA-backed lawsuits against Suno and Udio from June 2024 are partly settled, partly still pending (RIAA, Record Companies Bring Landmark Cases against Suno and Udio).

The EU relies on written law. The DSM Directive (EU) 2019/790 contains, in Article 4, an exception for text and data mining of lawfully accessible works, which applies only insofar as the use “has not been expressly reserved by their rightholders in an appropriate manner, such as machine-readable means” (EUR-Lex, Directive (EU) 2019/790). This machine-readable rights reservation is the lever with which rightholders can prohibit training use. The EU AI Act additionally obliges providers of general-purpose models to set up a copyright policy and to provide a public summary of the training content; these obligations for general-purpose AI models have applied since August 2, 2025, to newly placed models, while enforcement and fines take effect from August 2, 2026; for providers of models that were already on the market before August 2, 2025, the source provides a transition period until August 2, 2027 (Crowell & Moring, Code of Practice for General-Purpose AI Models Published). The accompanying General-Purpose AI Code of Practice was published on July 10, 2025, and, with its copyright chapter, offers a voluntary path to compliance (European Commission, The General-Purpose AI Code of Practice).

The United Kingdom is still undecided. Under current law the data-mining exception applies only to non-commercial research (Section 29A CDPA); commercial training without a license is not covered by it (GOV.UK, Report on Copyright and Artificial Intelligence). After a consultation from December 17, 2024, to February 25, 2025, with 11,520 responses, the initially preferred broad exception with opt-out is no longer the preferred line (GOV.UK, Report on Copyright and Artificial Intelligence). Just how much territoriality counts was shown by the Getty case: on November 4, 2025, the High Court decided that a model such as Stable Diffusion is not an “infringing copy”; Getty had withdrawn central copyright allegations because the training could not be proven in the United Kingdom (The Guardian, AI firm wins high court ruling after photo agency's copyright claim). Specialist analyses draw from this the lesson that territoriality counts profoundly and that UK secondary law sets limits (Mayer Brown, Getty Images v Stability AI).

Counterarguments, limits, and risks

It would be disingenuous to paint the situation only as a threat. Not every AI use is unlawful: several US courts have recognized training as fair use under certain conditions, and the Copyright Office recommends letting the licensing market grow without state intervention (U.S. Copyright Office, Part 3). Moreover, tools are emerging that are meant to lower the risk. Adobe advertises Firefly with the claim that it uses “commercially-safe datasets, including licensed and public domain content” and does not train on customer data (Adobe for Business, Firefly AI Approach). That is a vendor statement, not a court-confirmed legal guarantee, but it marks the direction: licensed training models, in-house or private models on cleared assets, as well as contractual indemnities.

The limits, however, remain real. Vendor commitments and indemnities are worth only as much as their cover, and the insurance market is currently moving away from silent AI cover (Honigman, The AI Insurance Gap). Provenance metadata can be removed, and even where training is permissible, an output can reproduce protected elements, as the Midjourney case shows. Added to this is international fragmentation: an asset that seems defensible in one market can be vulnerable in another, and for a globally exploitable production that is not a marginal issue but a core risk.

Conclusion

The industry has long pretended that the AI question was above all an aesthetic one. It is not. Whether an AI result looks convincing is the easiest of all tests. The hard ones come afterward: Is the provenance documented? Are the underlying rights cleared and licensed? Can the result be insured? And does it stand up to the legal situation in every target market?

Anyone who asks these questions only at delivery asks them too late. The robust answer arises early: through licensed or in-house training models, documented human authorship, a clean production log on asset provenance, early disclosure toward the insurer and the completion bond, as well as indemnities whose cover has been verified. Hollywood need not fear the copyright question but answer it before the camera rolls. This is not legal advice; the concrete legal situation must be qualified by market and date.