The advent of generative artificial intelligence has given rise to a debate about how this innovative technology will affect intellectual property rights and trade secret protections in particular.
There are open questions about, for example, whether prompt injection hacking can constitute trade secret misappropriation and, more ominously, predictions about the potential demise of trade secrets on the grounds that generative AI will render all trade secrets readily ascertainable.
Some observers opine that iconic trade secrets, like the recipe for a popular soft drink or a secret manufacturing process, could lose all protection if an advanced AI program replicates it.[1]
Courts have not yet resolved these issues, and there are reasons to doubt the negative forecasts. Courts in the past have addressed the challenges of innovative technologies for trade secrets and, in so doing, have sided with trade secret owners. These cases offer a potential road map as to how courts might address trade secret protections in the context of generative AI.
What are trade secrets?
A trade secret is valuable information, not generally known or easily ascertainable, that is protected from disclosure by one company and that would be of great value if known to a competitor. And like any secret, once it is out, it is no longer a secret.
According to Title 18 of the U.S. Code, Section 1839(3)(A)–(B), to constitute a trade secret, its owner must "take[] reasonable measures to keep such information secret," and the information must "derive[] independent economic value" from both "not being generally known" and "not being readily ascertainable through proper means" by someone "who can obtain economic value from the disclosure or use of the information."[2]
Determining trade secret status is generally a mixed question of law and fact, because the reasonableness of confidentiality measures is usually a fact question.
According to Section 1839(5), the law protects trade secrets from misappropriation by parties who learn of the information by improper means.[3] Improper means includes the use of "theft, bribery, misappropriation, breach or inducement of a breach of a duty to maintain secrecy, or espionage through electronic or other means," as stated in Section 1839(6)(A).[4]
Importantly, according to Section 1839(6)(B), trade secrets are not protected from "reverse engineering, independent derivation, or any other lawful means of acquisition."[5]
In short, a trade secret is secret valuable information protected by reasonable measures that is neither generally known nor readily ascertainable by one who would gain from its disclosure or use.[6] And learning — or trying to learn — the secret by improper means makes it no less a trade secret.[7]
Only reasonable measures are required for trade secret protection.
According to Section 1836(b)(2)(IV)(AA), to protect trade secrets, owners are required to undertake reasonable measures to maintain secrecy.[8] It goes without saying that unreasonable measures — safeguards that lack any contextual reasonableness — are not required. Several cases illustrate this concept in the context of technological challenges to trade secrets.
For example, in 1970, in E.I. duPont deNemours & Co. v. Christopher, the U.S. Court of Appeals for the Fifth Circuit held that otherwise legal aerial photography of a DuPont chemical plant used to discover a secret manufacturing process for producing methanol was nonetheless an "improper method of discovering trade secrets exposed during construction of the [] plant."[9]
In so holding, the court explained that the law does "not require a person or corporation to take unreasonable precautions to prevent another from doing that which he ought not to do in the first place," let alone "guard against the unanticipated, the undetectable, or the unpreventable methods of espionage now available."[10]
In other words, the court found it unreasonable and unrealistic to require DuPont, for example, "to put a roof over the unfinished plant to guard its secret."[11]
The U.S. Supreme Court subsequently relied on Christopher for this improper means principle.[12] And in its 2024 decision in Compulife Software Inc. v. Newman, the U.S. Court of Appeals for the Eleventh Circuit also applied Christopher's reasoning when it held that the automated data scraping of millions of fully public insurance quotes from a competitor's website constituted improper means for purposes of trade secret misappropriation.[13]
The court found "that even if individual [insurance] quotes that are publicly available lack trade secret status, the whole compilation of them (which would be nearly impossible for a human to obtain through the website without scraping) can still be a trade secret."[14]
In response to an argument that website scraping activities "may be perfectly legitimate," the court explained that the defendants had "not take[n] innocent screenshots of a publicly available site" and instead had "copied the order of Compulife's copyrighted code and used that code to commit a scraping attack that acquired millions of variable-dependent insurance quotes." The court compared this conduct to the use of surreptitious aerial photography at issue in Christopher.
In Compulife Software v. Newman, an earlier decision from 2020 on the same claims, the Eleventh Circuit explained that against "a certain type of reconnaissance," even taking "no measures to protect" a trade secret may still be reasonable or the method of taking improper.[15] In other words, there was nothing reasonable required of Compulife to protect its datasets and copyrighted code from the intrusion and theft facilitated by sophisticated technology tools that far exceeded the capacity of a human being acting alone.
These cases illustrate the propensity of courts to side with trade secret owners when trade secrets are threatened by technology innovations: aerial surveillance and automated bots. These specific technology innovations presented novel improper means for stealing trade secrets and both cases questioned the reasonableness of secrecy measures in the context of technological change.
In both circumstances, the Eleventh Circuit concluded that the trade secret owner had engaged in reasonable efforts to safeguard its trade secrets when threatened with enhanced intrusion opportunities posed by new technologies.
Generative AI platforms are not repositories of existing secret information.
According to Section 1839(3)(B), to be a trade secret, the information cannot be generally known or readily ascertainable,[16] meaning that the information is not easily accessible and is "difficult, if not impossible" to obtain.[17]
Put another way, readily ascertainable means that the information is easy to discover from another source through legitimate methods, such as in reviewing trade journals, reference resources or other published materials. Readily ascertainable relates to whether the trade secret owner has adequately safeguarded the information at issue with the specific objective of maintaining confidentiality.
Some observers contend that generative AI will render trade secrets readily ascertainable, thereby eviscerating trade secret status and legal protection. They reason that generative AI will facilitate the reproduction of a party's trade secrets through skillful prompt engineering.[18] But that conclusion misapprehends how generative AI works.
To begin, generative AI platforms are not all-knowing or exhaustive repositories of existing secret information. Generative AI outputs are not the equivalent of extracted text or data harvested from an organization's internal secret files and systems.
Instead, generative AI platforms generate new text based on the application of pattern creation algorithms to large language models and datasets that themselves are based on publicly available information. At present, these outputs are more apt to generate new trade secrets rather than to make any existing trade secrets readily ascertainable, because generative AI outputs reflect newly created data, not mere recall of existing data or heretofore secret information.
Importantly, datasets used to train generative AI models are not rife with trade secret information. Generative AI programs are trained on large volumes of information gleaned from the internet and other public sources. But these sources of public information and readily accessible data typically do not include confidential information and trade secrets.
As stated in the 2006 decision in BondPro Corp. v. Siemens Power Generation Inc., in the U.S. Court of Appeals for the Seventh Circuit, it is axiomatic that "[a] trade secret that becomes public knowledge is no longer a trade secret."[19] And, as stated in the 2006 decision in Oja v. U.S. Army Corps of Engineers, in the U.S. Court of Appeals for the Ninth Circuit, "[o]nce information has been published on a website ... there is no further act required ... to make the information available to the public."[20]
Such "publication of [] information over the internet" almost certainly destroys trade secret status.[21] The same principle could apply to secret information that is disclosed through prompts entered into a public generative AI platform.[22] However, the datasets that support large language models, insofar as they are developed from the internet and other public repositories, do not widely include trade secrets. The existence of nonsecret information in training data for generative AI models is therefore no threat to trade secrets.
Generative AI's outputs are also imperfect. The input-output exchange, much like a human conversation, is full of bias, variety, imprecision and error.[23] Any law student who has talked to a professor, or a lawyer who questions a witness, knows just how hard it is to ask the right question and elicit a useful answer.
Generative AI is no different: Notwithstanding the volume of information it might access, it mimics human language with all its imperfections. And generative AI platforms may hallucinate.[24] The outputs may be factually incorrect, "imagined," or simply unhelpful.
While a generative AI platform may produce a response to a prompt seeking a specific trade secret, such as the recipe for a soft drink, the answer's accuracy or proximity to the actual recipe cannot be ascertained. And multiple prompts for the same information may yield slightly different answers for each separate prompt. The various outputs may be interesting, but they are not a bona fide or confirmed reproduction of an as-yet-undisclosed trade secret.
In sum, while generative AI may increase access to information, and especially aid in the efficient synthesis of large amounts of information, it neither relies on trade secrets, regurgitates them on command, nor reliably ascertains them.
Conclusion
Insofar as generative AI language models and datasets are built from the internet and other public data repositories, they do not widely include trade secrets and confidential information. It is possible that generative AI might be used in the future to misappropriate trade secrets, and the scope of such uses and threats is not yet fully known.
However, courts to date have protected trade secret owners when technological advances have facilitated new and different improper means for trade secret theft. There is no reason to assume that courts will reverse their prior approach of protecting trade secret owners in the face of new technologies that might offer new "improper means," or that might challenge safeguards for secrecy that before now were deemed reasonable.
Joel Bush is a partner and Kurtis Anderson is an associate at Kilpatrick Townsend & Stockton LLP.
The opinions expressed are those of the author(s) and do not necessarily reflect the views of their employer, its clients, or Portfolio Media Inc., or any of its or their respective affiliates. This article is for general information purposes and is not intended to be and should not be taken as legal advice.
[1] Kyle Jahner, "Trade Secrets Law is Awkward Fit in AI Prompt-Hacking Lawsuit," Bloomberg Law (March 14, 2025), https://news.bloomberglaw.com/ip-law/trade-secrets-law-is-awkward-fit-in-ai-prompt-hacking-lawsuit; John G. Sprinkling, Trade Secrets in the Artificial Intelligence Era, 76 S.C. L. Rev. 181, 207-8 (2024) (predicting that Coca-Cola will lose "all protection" for its formula because the recipe will become readily ascertainable by "an advanced AI program").
[2] 18 U.S.C. § 1839(3)(A)–(B).
[3] Id. at § 1839(5).
[4] Id. at § 1839(6)(A).
[5] Id. at § 1839(6)(B).
[6] Id. at § 1839.
[7] The Uniform Trade Secrets Act ("UTSA"), which has been adopted with slight variations in every state except New York, contains materially and substantively identical definitions. See Compulife Software Inc. v. Newman , 959 F.3d 1288, 1311 n.13 (11th Cir. 2020) (Compulife 1) (identifying the "largely identical" and "substantially equivalent" definitions between the DTSA and Florida's UTSA).
[8] 18 U.S.C. § 1836(b)(2)(IV)(AA).
[9] 431 F.2d 1012, 1017 (5th Cir. 1970).
[10] Id. at 1016–17.
[11] Id. at 1016.
[12] See Kewannee Oil Co. v. Bicron Corp. , 416 U.S. 470, 475–76 (1974).
[13] Compulife Software, Inc. v. Newman , 111 F.4th 1147, 1163 (11th Cir. 2024) (Compulife 2).
[14] Id. at 1162.
[15] Compulife 1, 959 F.3d at 1312.
[16] 18 U.S.C. § 1839(3)(B).
[17] Allstate Ins. Co. v. Fougere , 79 F.4th 172, 189 (1st Cir. 2023) (affirming finding that spreadsheets containing large amounts of publicly available data were too extensive to be "readily ascertainable" from alternative sources).
[18] See, e.g., John G. Sprinkling, Trade Secrets in the Artificial Intelligence Era, 76 S.C. L. Rev. 181, 207 (2024); David S. Levine, Generative Artificial Intelligence and Trade Secrecy, 3 J. Free Speech L. 559, 580 (2023) (predicting that Gen AI will make "accessibility" a non-issue because obscure information will be easier to find).
[19] BondPro Corp. v. Siemens Power Generation, Inc. , 463 F.3d 702, 706 (7th Cir. 2006) (holding that a trade secret published in a patent application was no longer a trade secret).
[20] Oja v. U.S. Army Corps of Engineers , 440 F.3d 1122, 1131 (9th Cir. 2006).
[21] DVD Copy Control Assn., Inc. v. Bunner , 116 Cal. App. 4th 241, 10 Cal. Rptr. 3d 185, 193 (2004) (finding no trade secret protections for a program that was posted on "hundreds of [w]eb sites"); see also Arkeyo, LLC v. Cummins Allison Corp. , 342 F.Supp. 3d 622, 632 (E.D. Pa. 2017) ("The posting of materials on the internet without any confidentiality protections makes the information publicly available and renders the materials incapable of trade secret status.").
[22] Lewis Maddison, "Samsung workers made a major error by using ChatGPT," Tech Radar (April 4, 2023).
[23] Chad Boutin, "There is More to Ai Bias Than Biased Data, NIST Report Highlights," NIST (March 16, 2022), https://www.nist.gov/news-events/news/2022/03/theres-more-ai-bias-biased-data-nist-report-highlights; The AI Zone, "The Randomness of AI: Understanding and Embracing Uncertainty," Medium (August 6, 2024), https://medium.com/@the_AI_ZONE/the-randomness-of-ai-understanding-and-embracing-uncertainty-db15d728b829#:~:text=What%20is%20Randomness%20in%20AI,training%2C%20or%20decision%2Dmaking; Benji Edwards, "AI search engines cite incorrect news sources at an alarming 60% rate, study says," Ars Technica (March 13, 2025), https://arstechnica.com/ai/2025/03/ai-search-engines-give-incorrect-answers-at-an-alarming-60-rate-study-says/.
[24] Lisa Lacy, "Hallucinations: Why AI Makes Stuff Up, and What's Being Done About It," CNET (April 1, 2024), https://www.cnet.com/tech/hallucinations-why-ai-makes-stuff-up-and-whats-being-done-about-it/.
