Rapid advances in artificial intelligence (AI) technology, fluid market dynamics as new AI models become available, and a change in governmental viewpoints on AI have created an evolving AI landscape. This evolution is forcing a rethinking of existing intellectual property protection regimes, i.e., trade secrets, copyrights, and patents. Developers and deployers of AI technology will have to consider which of these offer the best protection for the AI models and their outputs and whether other protections in the form of contractual limitations are required to supplement the desired intellectual property protection. Developers and deployers will have to also consider the liabilities that may arise from running afoul of third-party intellectual property. This article discusses developments pertinent to AI in trade secrets, copyrights, and patents, and provides insight into how to integrate these developments into AI strategies.
IN THE NEAR TERM, TRADE SECRETS SUPPLEMENTED BY CONTRACTUAL PROVISIONS MAY OFFER THE BEST AVENUE OF PROTECTION
Many IP strategies focus on obtaining copyrights and patents to protect assets, because the registration of a copyright or grant of a patent provides a public recognition of a property right. But this comes at a cost of time and expense—factors that may be paramount in the fast-moving world of AI development. By comparison, trade secret protection offers an easier path to obtaining protection, because government approval is not required. Instead, federal and state trade secret statutes place the ability to maintain trade secret protection in the hands of the AI developers or deployers. Trade secret protection arises when an owner takes reasonable steps to protect information that derives independent economic value from not being generally well known or readily ascertainable through proper means.1 The information can be in any form or any type.2
The breadth of information that trade secret protection encompasses affords developers and deployers the opportunity to protect aspects of AI for which patent or copyright protection may be illsuited. This may include algorithms, model parameters such as number of nodes and weight values, and datasets selected for training, validation, and testing. These categories could run afoul of patent eligibility subject matter requirements and the creativity requirement of copyright.3 It will be important to consider whether the statutory requirements to obtain a patent or copyright will likely prevent their issuance, because the necessary disclosure of information during the patenting or copyright application process is contrary to the requirement to take reasonable steps to maintain the confidentiality of the information to establish a trade secret.4
Another benefit of trade secrets compared to copyrights and patents is clarity of ownership.5 Copyrights and patents vest ownership in authors and inventors, respectively. But the use of AI to develop information pertinent to the subject matter of the copyright or patent application may cloud the authorship and inventorship analyses. Ownership of a trade secret is derived based upon lawful possession of the information, as opposed to the manner by which the information was created.6 This avoids the ownership concerns that could arise if patent or copyright protection is sought.
A noted limitation of trade secrets is that they only provide protection against a competitor that acquires the trade secret by "improper means." Statutes define "improper means" as including "theft, bribery, misrepresentation, breach or inducement of breach of a duty to maintain secrecy, or espionage through electronic or other means."7
Recent events show that the term may also encompass activities, such as "scraping" or "prompt injection," that appear less nefarious than the terms recited in the statute.
DeepSeek's new AI model, which purportedly costs much less to train compared to leading models, sent shockwaves through the stock markets and resulted in about a trillion dollar combined loss in the value of key AI companies. A few days after the release of DeepSeek, one of these leading companies questioned whether DeepSeek used data from its model to train the newly released DeepSeek model. The company had noticed accounts believed to be associated with DeepSeek "scraping," i.e., bombarding, its AI model with millions of questions to obtain responses that would allow DeepSeek to access the underlying data associated with its model. It has been speculated that DeepSeek used this data to train the new version of its model. While the point of an AI system is to provide information in response to prompts, recent precedent has concluded that use of computer "scraping" to obtain more information than a human can obtain amounted to "improper means" in supporting a finding of trade secret misappropriation.8
The potential for an AI company to raise trade secret allegations based upon accessing an AI model's proprietary data is not hypothetical. OpenEvidence, Inc., has filed a complaint alleging trade secret misappropriation by Pathway Medical by "manipulat[ing] the OpenEvidence system into divulging its foundational code, both overtly and surreptitiously."9 OpenEvidence alleges that the defendants subverted safeguards by offering a series of questions that it described as "prompt injection" hacking to obtain "the set of instructions that define how the [OpenEvidence] AI model behaves and responds," i.e., the large language model's underlying algorithms.10 OpenEvidence further alleges that defendants accessed the OpenEvidence system in violation of OpenEvidence's terms of use.11 A decision in this case may 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 what impact the potential violations of a model's terms of use has on the answers to these questions. It may further provide rulings that offer clarity on reasonable steps to protect proprietary information associated with an AI model.
A distinction often raised when comparing a trade secret to a patent is a patent's ability to thwart competition from an independently developed product. But this may be a distinction without a difference when considering certain AI inventions. The black-box nature of AI models could limit the ability of a patentee to develop the reasonable basis required by a United States District Court before asserting patent infringement against an independently developed model. If a patentee does not understand how a model operates, it may need to show that a developer copied the patentee's technology. In certain instances, a patent owner may need to identify third-party access to the patent owner's information to provide a basis to allege patent infringement akin to what would be required to assert trade secret misappropriation.
A final consideration of the benefits of trade secrets for AI is understanding whether an owner can define its trade secret with reasonable particularity. As trade secrets do not include a set of claims like patents, courts have adopted a requirement that the owner of a trade secret define the trade secret at the outset of a litigation.12 The definition must be sufficient to inform the defendant of what information is at issue. "Merely describ[ing] the end results of or functions performed by the claimed trade secrets" may not suffice.13 This may create issues with describing the trade secrets associated with an AI model, because a court may find recitation of general terms like "artificial intelligence," "machine learning," "proprietary software," "algorithm," or "model" do not provide a defendant notice of the information alleged to have been misappropriated.
Thoughtfully crafted contracts and terms of use may provide a path to meet the obligation to show the existence of trade secrets. Requiring users to agree to contracts or terms of use that explain that the various AI components such as the algorithms, system prompts, and training data compilation are confidential and have value because of that status may be persuasive in a dispute if the defendant asserts the plaintiff has not provided notice of the actual trade secrets. These documents may also identify activities that are forbidden, e.g., scraping or prompt injection, to show that the competitor used "improper means" to acquire the information. In contracts, an entity may consider whether the inclusion of a non-compete provision and a provision prohibiting reverse engineering are feasible.
AI RAISES COPYRIGHT QUESTIONS WITH BOTH THE POTENTIAL TO PROTECT AI-RELATED CONTENT AND POTENTIAL INFRINGEMENT WHEN DEVELOPING AN AI MODEL— BUT THE USCO AND COURTS ARE BEGINNING TO ANSWER THE QUESTIONS
Copyright issues from a protection and infringement perspective have been at the forefront of the intersection of AI and IP. The United States Copyright Office (USCO) has issued several reports on the former, explaining that human authorship remains the crux of copyrightability and the protections it affords. This position is consistent with U.S. court precedent, including a recent decision that upheld a USCO denial of copyright registration to a visual work created solely by an artificial intelligence program.14 The USCO also explained that, at present, prompts alone do not meet the requirements of authorship to support registration of a copyright, because they "do not provide sufficient human control to make users of an AI system the authors of the output."15 But the USCO recognized the evolving nature of AI; it did not foreclose that "[t]here may come a time when prompts can sufficiently control expressive elements in AI-generated outputs to reflect human authorship."16 It should be appreciated that not all copyright authorities align on whether AI-generated materials qualify for copyright protection. Notably, the Beijing Internet Court in China ruled that AI-generated works are original and constitute graphic works that reflect authorship sufficient for registration.17
But, all is not lost for obtaining U.S. copyright protection for works that include some AI-generated material. The USCO acknowledges that copyrightability may arise for aspects of a work that include a sufficient human contribution.18 This may include human alteration of AI-generated materials.19 It also may include compilation work such as source code generation. Code writers often use AI to assist in drafting portions of a code.20 While the portion generated by AI is not copyrightable, the portions generated by the human coder should be eligible for copyright protection.21
Much ink has been spilled on AI and copyright infringement. This is a result of the concerns of content owners that AI training protocols infringe copyrights by ingesting and generating copies of protected material without permission. For generative AI, there are additional concerns that the output of those models may result in acts of copyright infringement by reproducing portions of copyrighted material in the AI output. These cases continue to wind through the U.S. court system, but we are beginning to receive some signs as to the direction in which the law may develop, particularly with respect to the defense of fair use.
The copyright statute identifies four factors of the fair use defense without addressing the weight to provide each. In application, courts have elevated some of the factors to higher importance. The factors are:
- The purpose and character of the use, including whether it is commercial;
- The nature of the copyrighted work;
- The amount and substantiality of the portion used in relationship to the copyrighted work as a whole; and
- The effect of the use upon the potential market for the value of the copyrighted work.
These factors were the focus of the court's analysis in Thomson Reuters Enterprises Centre GmbH v. ROSS Intelligence. 22
The Thomson Reuters court was presented with allegations that ROSS engaged in copyright infringement through its use of third-party generated content that was largely based upon Westlaw headnotes to train an AI model to identify legal cases in response to user questions. After deciding that the Westlaw headnotes qualified as copyrighted material, the court delved into the four fair use factors and concluded that two supported ROSS and two supported Thomson Reuters. But the factors supporting Thomson Reuters—particularly factor four—were entitled to more weight, compelling a finding that the fair use defense did not apply.23
The court stressed that a key fact driving its decision was that the AI model it analyzed was nongenerative, in that it was providing a list of cases that were pertinent to the query as opposed to generating a narrative response.24 This will be a distinction that courts dealing with generative AI models may need to address. But there are some aspects of the analysis that may shape the outcomes in other cases. In view of the commercial nature of ROSS's endeavor, coupled with the fact that ROSS's use of the copyrighted material to train a non-generative AI model did not result in a transformed secondary work, the first factor supported Thomson Reuters. Importantly, the court rejected ROSS's reliance on several cases that applied the fair use defense, because in those cases "the copying was necessary for competitors to innovate."25
To view the full article click here
Footnotes
1. 18 U.S.C. § 1839(3).
2. Id
3. 35 U.S.C. § 101 (patent eligible subject matter includes "any new and useful processes, machines, manufacture, or composition of matter, or any new and useful improvement thereof."); Feist Publications, Inc. v. Rural Telephone Service Company, Inc., 499 U.S. 340, 345, 111 S.Ct. 1282, 1287 (1991) ("The sine qua non of copyright is originality. . . . Original, as the term is used in copyright, means only that the work was independently created by the author . . . and that it possesses at least some minimal degree of creativity.").
4. 18 U.S.C. § 1839(3)(A).
5. Patent and copyright ownership derive from inventorship and authorship respectively. The use of AI may complicate the inventorship and authorship analyses.
6. "[T]he term 'owner,' with respect to a trade secret, means the person or entity in whom or in which rightful legal or equitable title to, or license in, the trade secret is reposed." 18 U.S.C. § 1839(4).
7. 18 U.S.C. § 1839(6).
8. Compulife Software, Inc. v. Newman, 111 F.4th 1147 (11th Cir. 2024).
9. OpenEvidence Inc. v. Pathway Medical, Inc., 12-cv10471, Document 1 at ¶ 2 (D. Mass Feb. 26, 2025).
10. Id. at ¶ 3.
11. Id. at ¶¶ 4, 8.
12. Oakwood Lab'ys LLC v. Thanoo, 999 F.3d 892, 906 (3d Cir. 2021).
13. T2 Modus LLC v. Williams-Arowolo, No. 4:22-CV00263, 2023 WL 6221429, at *5 (E.D. Tex. Sept. 25, 2023); see also, Yammine v. Toolbox For HR, 21-CV00093, 2023 WL 6259412, at *6 (D. Az. Aug. 8, 2023).
14. 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).
15. USCO, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025) at 18, https:// www.copyright.gov/ai/Copyright-and-ArtificialIntelligence-Part-2-Copyrightability-Report.pdf.
16. Id. at 21.
17. Li v. Liu (2023), Beijing Internet Court Civil Judgment, (2023) Jing 0491 Min Chu No. 11279.
18. USCO, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025) at 9 ("In sum, the use of a machine as a tool does not negate copyright protection, but the resulting work is copyrightable only if it contains sufficient human-authored expressive elements.").
19. Id. at 27 ("[T]he inclusion of elements of AI-generated content in a larger human-authored work does not affect the copyrightability of the larger human-authored work as a whole.").
20. Code developers should consider any licensing obligations that will apply to code generated by an AI model. Those licensing obligations may require a developer to provide access to the portions of code containing any AI-generated code.
21. In order to show copyrightability of the human-generated aspects of the source code, the code developer will need to be able to track the portions of the code generated by AI so that portion of the code is not claimed as copyrighted. If it is not feasible to segregate humangenerated versus AI-generated portions of the source code, trade secret may be a better option for protecting the code.
22. Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. ROSS Intelligence Inc., 1:20-cv00613-SB, 2025 WL 458520 (D. Del. Feb. 11, 2025) (Order).
23. Id. at *7.
24. Id. at **7-*8.
25. Thomson Reuters Enterprise Centre GmbH and West Publishing Corp. v. ROSS Intelligence Inc., 1:20-cv00613-SB, 2025 WL 458520, at *8 (D. Del. Feb. 11, 2025) (emphasis in original). The court noted that cases ROSS cited focused on making intermediate copies of portions of source code to develop functional capabilities or gain access to unprotected elements, e.g., copying code that allowed programs to interface with the software. Id.
Originally Published by Intellectual Property & Technology Law Journal
Visit us at mayerbrown.com
Mayer Brown is a global services provider comprising associated legal practices that are separate entities, including Mayer Brown LLP (Illinois, USA), Mayer Brown International LLP (England & Wales), Mayer Brown (a Hong Kong partnership) and Tauil & Chequer Advogados (a Brazilian law partnership) and non-legal service providers, which provide consultancy services (collectively, the "Mayer Brown Practices"). The Mayer Brown Practices are established in various jurisdictions and may be a legal person or a partnership. PK Wong & Nair LLC ("PKWN") is the constituent Singapore law practice of our licensed joint law venture in Singapore, Mayer Brown PK Wong & Nair Pte. Ltd. Details of the individual Mayer Brown Practices and PKWN can be found in the Legal Notices section of our website. "Mayer Brown" and the Mayer Brown logo are the trademarks of Mayer Brown.
© Copyright 2025. The Mayer Brown Practices. All rights reserved.
This Mayer Brown article provides information and comments on legal issues and developments of interest. The foregoing is not a comprehensive treatment of the subject matter covered and is not intended to provide legal advice. Readers should seek specific legal advice before taking any action with respect to the matters discussed herein.
