AI and Trade Secret Law: The Two Faces of Trade Secret Protection
Ancient Roman coins often bore an image of Janus, a two-faced deity. The two faces symbolized that Janus was the deity of both beginnings and endings in conflict, travel and trade.
Similarly, the law of trade secrets presents two faces when it comes to protection of generative AI. On the one hand, trade secrets law is one of the few IP regimes that will protect generative AI platforms and (in some cases) their output. But given recent revelations that employees of major companies have input their employer’s confidential information into third party generative AI platforms, AI use also creates a risk of loss of protection for a company’s trade secrets.
Trade Secret Protection for Generative AI (and Perhaps its Output)
Because generative AI creates output with little or no human intervention, developers have had a rough time finding protection for it under existing IP regimes such as copyright and patent. The output of generative AI is not protectible under copyright law because the lack of human action in creating the output means that the output is not a work of “expression,” which is a key requirement for copyright protection. Similarly, patent protection requires invention and novelty, and where a process or product generates an outcome or information with little or no human intervention these requirements are generally not met.
To qualify for protection as a “trade secret,” however, an item (whether information, methods, processes, programs or codes) does not have to meet these requirements. It does not have to be novel or the product of human invention like patentable subject matter, and does not have to be an original work of expression like a copyrighted work. It merely has to meet two criteria:
- the owner must take reasonable measures to keep it secret; and
- it must derive “independent economic value, actual or potential, from not being generally known to, and not being readily ascertainable through proper means by, the public[1]”.
In other words, unlike a copyrighted work (which needs to be an original work of expression) or patentable subject matter (which does not generally apply to software based inventions[2] given that they rely on abstract ideas and often lack an inventive step) a trade secret only needs to be kept reasonably secret, and has to have independent value by not being generally known. Due to this, a developer of an AI platform that takes reasonable measures to protect the data used to train the AI and the code supporting the platform may be entitled to claim trade secret protection for it.
So how can a developer of an AI platform maintain trade secret protection? By applying the same protective measures that have traditionally entitled other processes or information to have trade secret protection, such as:
- Ensuring that only employees and contractors with a need to access and use the platform, the training data, etc. have a right to use it. Access should not be company wide and should be protected by credentials, logged and tracked;
- Ensuring that the employees and contractors described above have executed agreements in which they acknowledge that the AI is a trade secret and is confidential, and (with respect to contractors and other third parties) that they are granted only a limited license to access and use the AI platform and training data for purposes of working for the company that owns the trade secret;
- Ensuring that the company has policies requiring that these measures be used when sharing information with subcontractors, vendors and other third parties. Such policies should require all third parties to maintain the confidentiality of the trade secrets and use them only for the purposes permitted by the trade secret owner.
The measures taken to protect the trade secret do not have to be absolute. They only need to be reasonable under the circumstances[3]. This generally means that the degree of protection to be afforded must be commensurate with the value of the trade secret, and is dependent on circumstances such as the sensitivity of the information and what is reasonably protective and yet still permits use of the trade secret (e.g., the company does not have to lock up all employee laptops every night).
With respect to the second element (that the item be “not generally known”), information does not need to be completely confidential to be a trade secret. A combination of elements that are publicly known will qualify for trade secret protection if the integration of publicly known information that makes up the trade secret is a unique, “effective, successful and valuable integration[4].” Trade secrets also only need to be minimally novel (i.e., not simply made up of general knowledge, skill or experience[5]), unlike patentable subject matter which has to meet a high bar for novelty.
In addition to the AI platform, the output of the AI platform may qualify as a trade secret if the users are bound by contractual obligations of nondisclosure and confidentiality. Trade secrets lose protection if they are publicly disclosed without any restriction on use or confidentiality, for example, in a legal filing, disclosure at public events, or publications[6]. However, courts will grant protection for trade secrets disclosed to third parties if the disclosure occurs pursuant to a confidentiality agreement (NDA) or similar restriction on disclosure[7]. If the developer of an AI platform requires its third party users (vendors, customers, etc.) to accept license and confidentiality terms that restrict the output of an AI platform to the users’ internal use, then the output may be protected under trade secret law.
By comparison, little if any protection is available under copyright or trademark law for AI output. In the few cases decided on the matter, courts have held that the output of an AI platform cannot be patented because the Patent Act requires that inventors must be human beings[8], and that AI cannot create copyrighted works because only a human can create a work of expression[9].
The Other Face of Janus: Losing Protection for Trade Secrets Through Employee Use of Third Party AI
Unfortunately, although AI gives, AI can also take away. In 2023, Cyberhaven Labs reported that its product detected that 4.7% of employees who used ChatGPT for work had copied and pasted their employers’ confidential information into the platform[10]. Employees who enter confidential or trade secret information of their employers as input data into a generative AI platform may be disclosing the information to a public source, potentially subjecting such information to loss of trade secret status. Depending on the AI platform’s terms of use, such disclosure may even grant the AI platform developer a license to use such information and trade secrets for its own purposes.
Once again, trade secret protection requires that companies take reasonable protections under the circumstances to safeguard their trade secrets from disclosure. In the context of AI platform use, companies can protect their trade secrets by updating employee policies to limit what kind of information can be submitted to AI platforms and adopting technical means to monitor, block or limit use of generative AI[11]
Conclusion: Flip a Coin on AI Use, or Take Action to Protect Trade Secrets
Whether a company is a developer of AI platforms or a user of them, generative AI has great potential for gain or for loss. The company can help determine whether its development and use of generative AI becomes a protectible part of its intellectual property portfolio, or a leaky sieve for company trade secrets and confidential information, by adopting the measures described in this article. In the alternative, the company can “flip a coin” when it comes to AI development and use and hope for the best.
© 2024 Daniel J. Langin
[1] See 18 U.S.C. § 1839; see also Uniform Trade Secrets Act Section 1.(4) (“Trade secret” means information, including a formula, pattern, compilation, program, device, method, technique, or process, that: (i) derives independent economic value, actual or potential, from not being generally known to, and not being readily ascertainable by proper means by, other persons who can obtain economic value from its disclosure or use, and (ii) is the subject of efforts that are reasonable under the circumstances to maintain its secrecy.”).
[2] See Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014). In this case, the Supreme Court ruled that a business method application for a software-based escrow service was not patentable because the algorithms on which the software was based were abstract ideas, which are not patentable.
[3] See e.g., Pioneer Hi-Bred Int’l v. Holden Found Seeds, 35 F.3d 1226, 1235 (8th Cir. 1994); Gates Rubber Co. v. Bando Chem. Indus., Ltd., 9 F.3d 823, 848-49 (10th Cir. 1993).
[4] Rivendell Forest Prods. Ltd. v. Georgia-Pacific Corp., 28 F.3d 1042, 1046 (10th Cir. 1994); Hertz v. Luzenac Group, 576 F.3d 1103, 1109 (10th Cir. 2009).
[5] Kewanee Oil Co. v. Bicron Corp., 416 U.S. 470, 476 (1974); BladeRoom Grp. Ltd. v. Facebook, Inc., Case No. 5:15-cv-01370-EJD (N.D. Cal. Jan 02, 2018).
[6] See e.g., Art & Cook, Inc. v. Haber, 416 F.Supp.3d 191, 196 (E.D.N.Y. 2017).
[7] See, e.g., Rockwell Graphic Sys., Inc. v. DEV Industries, Inc., 925 F.2d 174 (7th Cir. 1991).
[8] Thaler v. Vidal, 43 F. 4th 1207 (Fed. Cir. 2022).
[9] Thaler v. Perlmutter, [citation needed] (D. D.C. August 18, 2023).
[10] https://www.cyberhaven.com/blog/4-2-of-workers-have-pasted-company-data-into-chatgpt/ (accessed March 5, 2024).
[11] For example, Samsung discovered that employees had been pasting strings of code directly into ChatGPT to help debug it and loading confidential meeting notes into it to create summaries. As a result Samsung adopted a rule limiting employee prompts to ChatGPT on work devices to 1024 bytes. “Oops: Samsung Employees Leaked Confidential Data to ChatGPT,” Gizmodo April 6, 2023 https://gizmodo.com/chatgpt-ai-samsung-employees-leak-data-1850307376 (accessed March 5, 2024). According to the same article, Verizon and J.P. Morgan Chase blocked employee access to ChatGPT entirely.