In the creative industries, some copyright owners are exploring how they can make new content by combining their existing intellectual property assets with generative artificial intelligence.
In February, Microsoft Corp. announced that it has developed an AI model capable of generating visuals and gameplay based on footage from one of its Xbox titles.1 Activision Blizzard Inc., the developer of World of Warcraft and Diablo, has also reportedly trained an AI model on its games in order to generate concept art.2 Similar efforts are reportedly underway in the world of film.
These initiatives can serve multiple business goals. They can reduce costs and eliminate tedious tasks associated with, for example, game development or visual effects design. Using proprietary IP assets may also, at least theoretically, result in output that is a closer match qualitatively to the works a company has previously made, whether aesthetically or, for works involving data and facts, in terms of accuracy.
Legally, such initiatives may also help reduce some of the risk of liability.
Copyright owners considering such practices should still be aware that they are entering largely uncharted waters, however. The primary unknowns in this space include whether the owner can claim a copyright in the AI's output in view of a circuit court decision from March and U.S. Copyright Office guidance from earlier this year — even when the owner supplies the inputs — and whether there remains potential exposure to third parties.
Thankfully, there are a number of best practices and considerations that can help.
Ownership of Output
Copyright owners may be concerned that even if they train an AI model on their IP, they will not own anything that the AI generates. But owners likely have more rights than they believe in this scenario. Understanding the contours of the ownership rules regarding AI is key.
The U.S. Court of Appeals for the District of Columbia Circuit, in its March 25 Thaler v. Perlmutter decision, and the Copyright Office have held that copyright protection requires human authorship, and that purely AI-generated works thus do not qualify for protection.3
Extending this principle to various use cases, the Copyright Office has concluded that even where a human being has directed generative AI through complex prompts, without more, they generally will not own the AI's output because they have not exercised sufficient control over its expressive elements.4
In its January report, the Copyright Office noted that nuances arise when a user prompts an AI system with a work that itself contains protectable expression, such as by using a human-made drawing or poem as the prompt. In such cases, if the user owns the prompt, they may be able to assert their rights in both the prompt and in the portions of the output in which their original expressive elements are perceptible.5
The Copyright Office's report did not address, however, what happens when a copyright owner has trained an AI model entirely on its own human-made works. In that situation, it is less clear that the copyright owner could claim protection in any part of the output, even where the output includes perceptible expressive elements from the training corpus.
Notably, the training data used for many AI models consists of human-authored works, and yet the Copyright Office has taken the general position that AI-generated materials are not copyrightable.
Thus, paradoxically, copyright owners may have more success claiming ownership over AI output where they have prompted the AI with their IP rather than trained the AI with the same IP, even if the output is the same in both cases.
On a practical level, even in cases where the copyright owner cannot obtain a registration for AI output that is based on their existing IP, that does not mean that the output is without protection. In that scenario, the copyright owner may have a claim against third parties that would copy the AI system's output based on their rights in their original IP. For example, an animation studio that trains an AI model on footage from its cartoons may be able to prevent third parties from copying AI-generated output featuring the copyrighted characters from the original cartoons.
Moreover, AI users can claim ownership in the original expression they add to AI-generated materials — for example, human-written narrative added to AI-generated images — and can also claim ownership over the original selection and arrangement of AI-generated materials.6
In fact, in February, it was announced that, based on such a selection and arrangement theory, the Copyright Office reportedly approved a registration for an entirely AI-generated work where the author had combined roughly three dozen images that he had used AI to generate.7
Risks of Copyright Infringement
Companies that train AI models on their own IP may also be seeking to reduce exposure to third-party infringement claims. The success of this strategy depends on a number of factors.
First, many works involve a bundle of rights of multiple different people. A film, for example, may include licensed music, third-party branding and footage of actors with publicity rights; a sound recording may embody a musical work written by one or more third-party songwriters.
The IP owner may also be subject to other restrictions, such as those arising from deals with distributors. Before such works are reproduced through AI, it is important to analyze whether existing licenses and other contracts allow such use.
Second, even if the company trains an AI model on its wholly owned IP, it is important to understand whether third-party-owned works were also included in the model's training corpus — e.g., at an earlier stage. If so, those third-party-owned works could be reproduced in the AI's output.
Consider the same example of an animation studio that has trained a model on its cartoons. If no other works were used to train the model, the studio may have more comfort in its right to use anything that the AI generates as compared to when the studio's cartoons were only one portion of the training data alongside, say, cartoons created by a rival.
Third, even if a company has restricted the universe of information available to an AI system to the company's own IP, that does not mean that subsequent misuse of the AI system cannot lead to infringement. The AI may still be influenced by prompts embodying third-party content, for example. Accordingly, it is important to instruct employees on best practices to avoid such misuse.
Practical Tips
Based on the above considerations, here are a few takeaways for copyright owners who want to leverage the capabilities of generative AI to create new works based on their existing IP.
- Whether the owner can claim a copyright in the AI's output depends on a number of factors, such as whether the owner's works were used to prompt the AI versus merely to train the AI model and whether postgeneration modifications were made to the output.
- For practical purposes, regardless of whether the owner can claim copyright in the AI's output directly, in some cases it may be enough to rely on existing rights in the original works.
- To reduce exposure, the owner should ensure it has the ability, contractually and otherwise, to use its IP assets with generative AI.
- The owner should also understand what other data, if any, was used to train the AI model, and should instruct its employees on best practices to avoid infringing third-party rights.
Associate Parmida Enkeshafi was a contributing author to this client alert.
Footnotes
1 Microsoft, Introducing Muse: Our first generative AI model designed for gameplay ideation (Feb. 19, 2025).
2 New York Times, A.I. May Help Design Your Favorite Video Game Character (May 22, 2023).
3 Copyright Office, Copyright Registration Guidance: Works Containing Material Generated by Artificial Intelligence, 88 Fed. Reg. 16,190 (Mar. 16, 2023); Thaler v. Perlmutter No. 23-5233, --- F.4th ---- (D.C. Cir. Mar. 18, 2025).
4 See e.g., Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 2025), at 19; Copyright Office, Second Request for Reconsideration for Refusal to Register SURYAST (SR # 1-11016599571; Correspondence ID: 1-5PR2XKJ) (Dec. 11, 2023); Copyright Office, Second Request for Reconsideration for Refusal to Register Théâtre D'opéra Spatial (SR # 1-11743923581; Correspondence ID: 1-5T5320R) (Sept. 5, 2023); Copyright Office, Zarya of the Dawn (Registration # VAu001480196) (Feb. 21, 2023).
5 Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (Jan. 2025), at 23-24.
6 Id. at 24-27.
7 CNET, This Company Got a Copyright for an Image Made Entirely With AI. Here's How (Feb. 10, 2025).
Originally published by Law360
The content of this article is intended to provide a general guide to the subject matter. Specialist advice should be sought about your specific circumstances.
