Protecting AI Assets and Outputs with IP Strategies in a Changing World
Artificial intelligence (AI) technology is evolving at extraordinary speed, creating new challenges and opportunities for intellectual property (IP) protection. AI developers and deployers need to rethink their IP strategies to safeguard their AI models and outputs and to avoid infringing on others’ IP rights. This chapter discusses some of the key developments and considerations for IP protection of AI in the areas of trade secrets, copyrights, and patents.
Trade Secrets: A Flexible and Fast Option for AI Protection
Trade secrets are a form of IP protection that does not require government review. Instead, federal and state statutes protect any type of information that derives economic value because a company takes reasonable steps to protect its secrecy. The breadth of information covered by trade secret law affords AI developers and deployers the opportunity to protect aspects of AI for which patent or copyright protection may be ill-suited or unavailable, including algorithms, model parameters, such as number of nodes and weight values, and datasets selected for training, validation, and testing. Trade secrets also avoid the ownership issues that may arise with copyrights and patents, because trade secret ownership derives from a legal right to possess the information, irrespective of the origin of the information. By comparison, patents and copyrights define ownership through the concepts of inventorship and authorship—neither of which legally exists for information created solely by AI.
Trade secrets do not protect against independent development or reverse engineering of the same or similar information, only against improper acquisition of the secret information. Traditional examples of improper means include theft, espionage, or breach of contract. Recent cases provide support to expand the definition of “improper means” to include activities that encompass the manner by which a competitor may seek to obtain inputs to an AI model.1 These activities include scraping, prompt injection, or manipulation to obtain information, especially where such actions violated a website’s terms of use.2 Terms of use that explicitly forbid these activities may support arguments that the owner took reasonable measure to protect the information, particularly if the owner monitors activities with the AI platform and takes action when it identifies accounts violating the terms of use. Cases are beginning to arise that should provide insight into the scope of “readily ascertainable,” “reverse engineering,” and “improper means” when dealing with generative-AI models designed to provide information in response to queries, and into what impact the potential violation of a model’s terms of use has on the answers to these questions.3
Trade secret owners should consider whether or not they can define a trade secret with “reasonable particularity.” While this is not a statutory requirement, courts often impose it. The use of generic terms such as “artificial intelligence,” “machine learning,” “proprietary software,” “algorithm,” or “model,” or describing the function performed may not suffice.4 Trade secret owners may consider including language in the terms of use or licenses that acknowledges that these components derive value from the fact that the owners keep them confidential. That could provide an argument that the parties to the contract have received notice as to the scope of the trade secrets.
Copyrights: A Potential Source of Protection and Infringement for AI-Related Content
Copyrights grant an owner the exclusive right to reproduce, distribute, perform, display, or create derivative works of original works of authorship. Because the United States limits copyright protection to the human-created content and not the AI-generated content, copyrights can be particularly useful for protecting AI-related content that involves creative human contributions.5 Notably, other jurisdictions do recognize AI-generated works as original and eligible for protection.6 These jurisdictions could offer a path to register AI-generated content and police unauthorized publication of the content on the internet.
Separate from the issue of the availability of copyright to AI-generated content is the potential for copyright infringement when training an AI model with copyright-protected material. The number of existing cases on this issue shows that it resides at the forefront of AI and IP disputes. The allegations in these cases are twofold. One allegation raises concerns with AI models that ingested and copied copyrighted material for training purposes. The second focuses on the potential for the outputs to contain or resemble copyrighted material used in training.7 An obvious way to avoid such allegations is to obtain rights to the training data. But for large language models, this may be impractical and financially unrealistic. Thus, the outcomes of the pending cases will be important in showing whether AI-model developers achieve the results for which they have lobbied lawmakers and argued before the judiciary—that using copies of copyrighted materials to train a model is fair use.
The fair-use defense includes four factors to which courts apply varying weight.8 These factors are: (1) the purpose and character of the use, (2) the nature of the copyrighted work, (3) the amount and substantiality of the portion used, and (4) the effect of the use on the potential market for the copyrighted work.9 The first decision to address this defense in an AI context—Thomson Reuters v. ROSS Intelligence—did not side with the model developers.10 It should be noted that the ROSS court sought to limit its decision to non-generative AI models. Two cases addressing generative AI large language models (LLMs) issued opinions after the ROSS decision. Both of those cases held that the use of copyrighted material to train the LLMs was a highly transformative use.11 Both decisions held the four-factor fair use test supported a conclusion that the defense applied. The court decisions were swayed by different arguments and facts in each case. But both pointed to the use of guardrails to prevent the inclusion of copyrighted content in the LLM’s output as a important consideration.12 Although these decisions found that the fair use defense applied, both left room for arguments that specific facts in other cases could lead to a different result.
Patents: A Powerful but Challenging Option for Protecting AI-Generated Outputs
Patents are a form of IP protection that grant the owner the exclusive right to make, use, sell, or import a new and useful invention for a limited period of time. Patents can be a powerful tool for protecting AI-generated outputs provided that human involvement exists with the conception or reduction to practice of the invention, as AI cannot be named as an inventor.13 A key advantage of patents over trade secrets and copyrights is that independent development or reverse engineering are not defenses to patent infringement.
But questions exist as to how patent offices and courts will deal with fundamental patent concepts such as the patent-eligible subject matter requirement, the prior art, the knowledge of the person of ordinary skill in the art, and the patentability analysis in view of AI. The USPTO guidance on patent-eligible subject matter explains that the existing principles suffice to address AI. Concerns exist that the existing standard presents too high a barrier for certain AI advancements, causing some in the industry to advocate for Congress to address the issue through legislation.14
There remain several patent concepts for which the USPTO has sought public comment but has not provided guidance.15 These include the impact of AI on prior art, knowledge of the skilled artisan, and patentability analysis such as obviousness, written description, and enablement. The USPTO’s and, more importantly, courts’ views of these concepts may fundamentally change the analysis of anticipation, obviousness, enablement, and written description. However, key courts are beginning to weigh in on these issues. In the first Federal Circuit ruling on patent eligibility issues concerning machine learning, the court held that “claims that do no more than apply established methods of machine learning to a new data environment” are not patent eligible.16 As the Federal Circuit and other courts issue decisions involving AI technologies, the analyses will continue to evolve.
Conclusion
AI technology is changing the IP landscape, creating new opportunities and challenges for IP protection and enforcement. AI developers and deployers need to adapt their IP strategies to protect their AI assets and outputs, and to avoid infringing on others’ IP rights. Trade secrets, copyrights, and patents may offer different advantages and disadvantages for protecting AI-related IP, depending on the type, nature, and scope of the AI information, model, or output. AI developers and deployers should consider using multiple approaches to protect their AI investments and to monitor the developments in the IP law and practice as they pertain to AI-generated content.
1 See Compulife Software, Inc. v. Newman, 111 F.4th 1147 (11th Cir. 2024); UAB “Planned5D” v. Facebook, Inc., 19-cv-03132, 2020 WL 4260733, *7 (N.D. Cal. July 24, 2020); OpenEvidence Inc. v. Pathway Medical, Inc., 25-cv-10471, Document 1 at ¶ 2 (D. Mass Feb. 26, 2025).
3 See, e.g., OpenEvidence Inc. v. Pathway Medical, Inc., 25-cv-10471, Document 1 at ¶ 2 (D. Mass Feb. 26, 2025).
4 T2 Modus LLC v. Williams-Arowolo, No. 4:22-CV-00263, 2023 WL 6221429, at *5 (E.D. Tex. Sept. 25, 2023) (“Merely describ[ing] the end results of or functions performed by the claimed trade secrets” may not suffice); see also, Yammine v. Toolbox For HR, 21-CV-00093, 2023 WL 6259412, at *6 (D. Az. Aug. 8, 2023).
5 Thaler v. Perlmutter, No. 23-5322, 2025 WL 839178, at *1 (D.C. Mar. 18, 2025) (holding the Copyright Act of 1976 requires human authorship to be eligible for registration).
6 Li v. Liu (2023), Beijing Internet Court Civil Judgment, (2023) Jing 0491 Min Chu No. 11279.
7 The United States Copyright Office highlighted two other potential copyright infringement concerns in the third part of its series on copyright and artificial intelligence, which was published in pre-publication form in May of 2025. See Copyright and Artificial Intelligence Part 3: Generative AI Training, Pre-Publication Version, May 2025. In the report, it raised concerns that downloading AI model “weights” may result in generating a copy of a digital representation of copyrighted material and that use of “retrieval-augmented generation” or “RAG” may result in generating copies of copyrighted materials. See id. at 28 and 30, respectively.
8 Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. ROSS Intelligence Inc., 1:20-cv-00613-SB, 2025 WL 458520, at *7 (D. Del. Feb. 11, 2025)
10 Id. at *10. Of note, ROSS has been granted an interlocutory appeal in the Third Circuit; the outcome of this appeal will likely have significant impact on future cases involving AI models and training practices.
11 Bartz et. al. v. Anthropic PBC, C 24-05417, Doc. 231 (N.D. Cal. June 23, 2025); Kadrey et. al. v. Meta Platforms, Inc., 23-cv-03417, Doc. 598 (N.D. Cal. June 25, 2025).
12 Bartz et. al. v. Anthropic PBC, C 24-05417, Doc. 231, at 11-12 (N.D. Cal. June 23, 2025); Kadrey et. al. v. Meta Platforms, Inc., 23-cv-03417, Doc. 598, at 12 (N.D. Cal. June 25, 2025); see also Copyright and Artificial Intelligence Part 3: Generative AI Training, Pre-Publication Version, May 2025, at 62, 74. See also Concord Music Group, Inc. v. Anthropic PBC, in which Anthropic agreed to implement measures that prevent the inclusion of copyrighted material in its model’s output, and the court cited the agreement in denying the request for a preliminary injunction. 5:24-cv-03811-EKL, 2025 WL 904333 *3, *6 (N.D. Ca. Mar. 25, 2025). Such a safeguard would likely moot some of a content owner’s arguments under factor 3.
13 Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022).
14 Ryan Davis, Law360, “White House Urged to Back Patent Eligibility Bill to Aid AI,” published March 12, 2025.
15 Request for Comments Regarding the Impact of the Proliferation of Artificial Intelligence on Prior Art, the Knowledge of a Person Having Ordinary Skill in the Art, and Determinations of Patentability Made in View of the Foregoing, 89 Fed. Reg. 34217 (April 30, 2024).
16 Recentive Analytics, Inc. v. Fox Corp., 2025 WL 1142021, at *4 (Fed. Cir. Apr. 18, 2025).
