Can You Keep (An AI) Secret? The Role Of Trade Secrets In IP Protection Strategies For AI

Introduction

As artificial intelligence (AI) technology advances, and companies invest hundreds of millions of dollars to stay ahead of the competition, the strategies for protecting the related IP have received increasing attention. Some of the primary types of IP protection – patents and copyrights – are applicable to certain aspects of AI technology. However, various decisions by the U.S. Patent Office and U.S. Copyright Office limit the patentability and copyrightability of AI-generated inventions and works of authorship.

In light of the growing complexity and limitations on patent and copyright protection for AI, trade secret law is playing a vital and evolving role in the protection of certain AI technologies. Trade secret protection can be valuable. Many recent damage awards in trade secret cases have exceeded $100 million. Companies need to understand how trade secrets can be used as part of a comprehensive IP protection strategy for AI. However, as with patents and copyrights, there are limits to the scope of protection that can be secured by trade secrets. And the risk of loss of that protection must be considered.

Additionally, there are some unique challenges with AI trade secrets, including identification of the relevant aspects of an AI system and new technological threats that may reveal trade secrets, such as prompt injection and model distillation (explained below). In some cases, adversarial attacks via prompt injection or model distillation techniques are used to extract trade secrets from AI systems. Companies need to understand these issues and adopt trade secret protection strategies with these challenges in mind.

Also, emerging AI regulatory frameworks contemplate certain disclosures for safety, oversight, or transparency that may conflict with trade secret protections. Companies that intend to rely on trade secret protection need to be aware of and monitor such regulations.

This article will delve into these topics and other key emerging issues related to AI and trade secrets.

Challenges with Patent and Copyright Protection for AI

Many software-based systems are protected by a combination of types of IP. Two of the main types of IP used for such protection are patents and copyrights. Both have significant limitations with AI-related inventions. With the advent of generative AI, it is possible that the output of a generative AI tool can be an “original” work or an “invention.” In fact, the United States Patent and Trademark Office (USPTO) and the United States Copyright Office (USCO) have received applications naming an AI tool as an inventor on a patent application and as the author on copyright registrations. However, as detailed below, these applications and registrations have been rejected for lacking human inventorship or authorship.

Patents

Patents protect inventions, while copyright protects original works of authorship. Patents require an application to be filed with the USPTO and necessitate that the invention is new, useful, and non-obvious. Typically, a U.S. application is published within 18 months of filing (unless a nonpublication request is filed) if it has not already been granted as a patent. Either way, once the information is published, any contents of the application are no longer secret. While the subject matter of a patent application may be subject to trade secret protection before the patent is published or is issued, once one of those events occurs, the protection is lost.

Various impediments may limit the availability of patent protection for certain AI inventions. One impediment is what qualifies as patent eligible subject matter under 35 U.S.C. § 101. Abstract ideas are not patent eligible. Many Patent Examiners and courts have asserted that various AI innovations are unpatentable “abstract ideas” (e.g., mathematical formulas, algorithms or laws of nature). These issues are well known and apply to all software inventions, whether AI-related or not.

Another impediment particularly relevant to AI is the requirement for human inventorship. Patent applications must list the actual inventors of the subject matter to be patented. An AI tool cannot be listed as an inventor on a U.S. Patent application. Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022). Thus, any inventions created purely by AI are not likely patentable. Patents for AI-assisted inventions are not prohibited but significant human conception of the invention must exist to qualify. The USPTO issued initial guidance on this topic focusing on the Pannu test for joint inventorship. See, Inventorship Guidance for AI-Assisted Inventions (2024). However, it later rescinded the initial guidance and replaced it with updated guidance.¹ See Revised Inventorship Guidance for AI-Assisted Inventions (November 28, 2025) focusing on conception. According to the updated guidance, conception is “the formation in the mind of the inventor, of a definite and permanent idea of the complete and operative invention, as it is hereafter to be applied in practice.”  Conception is complete when “the inventor has a specific, settled idea, a particular solution to the problem at hand, not just a general goal or research plan.” The USPTO issued updated guidance dated December 5, 2025 addressing patent eligibility under 35 U.S.C. § 101, particularly when evaluating claims related to machine learning or artificial intelligence.

Copyrights

Copyright protects original works of authorship fixed in a tangible medium. Copyright protection automatically exists once a work is created and fixed in a tangible form. However, a federal registration can be obtained and, in fact, must be obtained before filing a lawsuit for copyright infringement. As part of the application, a copy of the work must be deposited. Once you submit a deposit copy to the USCO, it becomes part of the public record and can be viewed by members of the public upon request. However, in some cases, the applicant may submit “identifying material” instead of an entire copy of the work. Identifying material must adequately represent the authorship claimed in the application. Thus, at least some of the deposit will be public and at least that portion will not maintain trade secret status.

Copyright can be used to protect software, certain aspects of UIs, documentation and other creative works. Copyright does not extend to functional aspects of AI systems and processes, such as model architecture or training methods.

Content that is solely AI-generated is ineligible for copyright protection. Only humans can be recognized as authors and the work must be the product of human creativity. Merely prompting an AI tool is generally not sufficient to make a human an author. See, Thaler v. Perlmutter, 130 F.4th 1039 (confirming that works generated solely by AI without human authorship are not copyrightable, though the decision did not definitively resolve when prompting or other human involvement in AI-assisted creation becomes sufficiently creative for authorship). The USCO has issued guidance on AI-related works, confirming that most AI-generated works cannot be protected by copyright. See, Copyright Registration Guidance for Works Containing AI-Generated Materials (2023) (“Initial Guidance”)² and Copyright and Artificial Intelligence Part 2: Copyrightability.

For these and other reasons, certain aspects of AI systems may not be protectable under patent or copyright law. And filing a patent application or copyright registration may result in the work becoming public resulting in the loss of trade secret. In these cases, trade secrets may be worth consideration where feasible.

Trade Secrets

Easy to Protect, Easy to Lose

In contrast to patents and copyrights, trade secrets provide a more flexible form of IP protection. A trade secret is defined as information that derives independent economic value from its secrecy and is subject to reasonable efforts to maintain that secrecy. Trade secret protection does not necessitate formal registration. Unlike patents, which require public disclosure and have limited duration, trade secrets can protect information indefinitely, provided that reasonable measures are taken to maintain confidentiality. With commercially deployed systems, anything exposed to the user through use is typically not a trade secret. Anything “under the hood” (i.e., not exposed to the user) may be.

Traditional trade secret issues often involve reverse engineering, unauthorized access to computers, and theft of technology by employees or other (external) bad actors.

These issues also are relevant to AI and various AI-based trade secret lawsuits have resulted (as discussed below). Additionally, new theories of AI-specific misappropriation have resulted, for example, from lesser-known exploits such as prompt injection and distillation.

Trade Secret Law

In the United States, trade secret protection is governed by both federal and state law. The Defend Trade Secrets Act (DTSA), codified at 18 U.S.C. § 1836, provides a federal civil remedy for trade secret misappropriation and coexists with state-level protections based on the Uniform Trade Secrets Act (UTSA), which has been adopted in some form by most states. The Economic Espionage Act (EEA), 18 U.S.C. §§ 1831–1839, enables criminal prosecution of trade secret theft, particularly in cases involving foreign actors or national security concerns. Under the DTSA and state analogs (e.g., the UTSA), misappropriation occurs by acquisition of valuable secret information via “improper means.”

To successfully assert a trade secret claim under the DTSA or UTSA, a plaintiff must establish the following:

Existence of a Trade Secret: The information must have commercial value due to its secrecy and must not be generally known or easily discoverable.

Misappropriation: The trade secret must have been acquired through improper means (such as theft, bribery, or breach of duty) or used/disclosed without authorization.

Reasonable Measures: The owner must have taken reasonable steps to maintain the confidentiality of the information.

To defend against a trade secret claim, alleged infringers often argue that the information was publicly known or that the plaintiff failed to take adequate measures to preserve its secrecy. Defendants may try to show they independently developed or legally acquired the alleged trade secret, negating a claim they misappropriated it.

In some cases, a defendant may allege the trade secret was legally reverse engineered. The Supreme Court has held that products in the public domain may be reverse engineered “to discover and exploit the trade secret.” Bonito Boats, Inc. v. Thunder Craft Boats, Inc., 109 S. Ct. 971 (February 21, 1989). Additionally, comments to the Uniform Trade Secrets Act state that reverse engineering is legal when the analyzed product was obtained “by a fair and honest means, such as purchase of the item in the open market.” Thus, features of a product that are properly obtained and can be fairly accessed may not survive as trade secrets once the product is made publicly available. As discussed further below, the boundaries of permissible reverse engineering take on new dimensions in the AI context, where techniques such as prompt injection and model distillation may be used to extract information from AI systems.

Trade Secrets Can Be Valuable

A number of trade secret lawsuits have resulted in staggering damage awards. Trade secret awards have trended upward in recent years (especially post-DTSA in 2016) with more nine-figure verdicts driven by unjust enrichment theories, willful conduct allowing punitive damages (up to double under DTSA), and juries punishing perceived bad-faith theft.³ Such cases often involve jury verdicts that include compensatory damages (e.g., lost profits, unjust enrichment) and sometimes exemplary/punitive damages for willful conduct. Software, technology, and high-value proprietary processes dominate these high-stakes cases.

Here are examples of just some of the largest damage awards in trade secret misappropriation cases from U.S. courts (under the DTSA or state equivalents such as the UTSA).

Appian Corp. v. Pegasystems Inc. (2022, Virginia state court): A jury awarded over $2 billion to Appian for trade secret misappropriation involving software espionage. This stands as one of the largest-ever jury verdicts in a trade secret case. The verdict was subsequently reversed by the Virginia Court of Appeals due to evidentiary and legal issues. On further appeal, the Virginia Supreme Court reversed the Court of Appeals in part and remanded the case for a new trial.

Propel Fuels, Inc. v. Phillips 66 Co. (2024/2025, California state court): A jury awarded $604.9 million in compensatory damages for willful misappropriation of trade secrets related to low-carbon fuel technology. A judge later added $195 million in exemplary damages, bringing the total verdict to around $800 million. This ranks among the highest recent verdicts.

TriZetto Group v. Syntel Inc. (2020, New York federal court): A jury awarded $855 million for trade secret misappropriation and copyright infringement involving healthcare software. Parts were later vacated or reduced on appeal, but it remains one of the top verdicts from the early 2020s.

Motorola Solutions v. Hytera Communications (2020, Illinois federal court): Roughly $764 million was awarded for theft of radio/communications technology trade secrets by a Chinese rival (later adjusted on appeal, with the Seventh Circuit in 2024 affirming in part and remanding).

Examples of AI Trade Secrets

AI technology varies but often involves some combination of data (including curated data sets), AI models, model weights, training methods, hyperparameters, internal tuning processes, algorithms, user interfaces and/or other components. Many people are aware that proprietary algorithms can incorporate trade secrets. But other items on this list also may involve trade secrets as discussed below.

Data

AI applications leverage vast quantities of data to train AI models. Many general-purpose AI tools are built on large language models (LLMs), which are AI models trained on extensive datasets of text, books, images, code, and other data. Data may come from public sources, proprietary sources, synthetic generation, or otherwise.

Trade secret protection in some aspects of data is often overlooked. Public data itself is generally not protectable by trade secrets. However, aspects of curated or proprietary data sets used for training may include trade secrets. Courts have clarified that trade secret protection may even be available where publicly available information is compiled, curated, or utilized in a proprietary manner and the arrangement derives value from its secrecy. For example, in Compulife Software, Inc. v. Newman, et al., No. 21-14074 (11th Cir. Aug. 1, 2024), the Court held that even if individual, publicly available insurance quotes lack trade secret status, the confidential rate database upon which the quotes are generated can be a trade secret and accessing the database improperly was a trade secret violation.

In United States v. Nosal, Nos. 14-10037 and 14-10275 (9th Cir. July 5, 2016), the Ninth Circuit determined that a database assembled from public sources constituted a trade secret due to the confidential methods employed in its compilation. The court emphasized that “it is the secrecy of the claimed trade secret as a whole that is determinative. The fact that some or all of the components of the trade secret are well-known does not preclude protection for a secret combination, compilation, or integration of the individual elements…” It explained that such a compilation may include data from public sources or a combination of proprietary and public sources. The Court further clarified that a compilation that affords a competitive advantage and is not readily ascertainable falls within the definition of a trade secret.

This rationale supports the applicability of trade secret law to certain AI-related datasets, even if they include public data, provided the dataset as a whole is secret and affords a competitive advantage.

Additionally, data needs to be prepared and processed to be used for AI training. Proprietary data processing techniques for AI training data may include trade secrets.

These are just some examples of AI-related data that may be protectable by trade secrets.

Training and Model Weights

The process of training AI models involves, among other things, identifying patterns in the data, extracting other information from the data and producing model weights. Model weights are the numerical parameters learned during training that define how the model processes input data to produce outputs. In neural networks, weights are the values in the connections between neurons, determining their influence on the output. Training the model includes specifying hyperparameters and iteratively adjusting model weights and other parameters to minimize errors, effectively “teaching” the model.

The following are some examples of the many aspects of training that may qualify for trade secrets:

• specific hyperparameter settings (e.g., learning rate, batch size) or optimization strategies (e.g., custom learning rate schedules) that improve model performance
• proprietary methods for training models
• custom algorithms for training, such as specialized optimization techniques or distributed training methods
• model architectures
• model weights
• inference optimization techniques (e.g., ones which enhance performance, speed, or cost-efficiency during deployment)
• AI models optimized for different applications, such as natural language processing, computer vision, recommendation systems, robotics and control systems, generative AI and other applications
• proprietary reinforcement learning techniques
• methods for fine-tuning pre-trained models for specific tasks, including the selection of datasets or tuning strategies

Inputs and Outputs

It is interesting to note that even inputs to and outputs from AI models may qualify as trade secrets, depending on how they are handled. Each AI system is designed and operates differently. Some treat inputs confidentially and only use the input to process your prompt. Others store them and use them to retrain the AI model. In fact, some terms of service (ToS) require you to expressly grant to the AI tool provider a license to your inputs and/or outputs. For those that do not require a license, if the AI system generates a unique solution for a business and the tool provider does not retain or reuse the result, the business may claim trade secret protection. By contrast, if the provider keeps copies of outputs, or if identical outputs are generated for other users, the claim to secrecy is weakened or lost.

Importantly, businesses should also consider the risk that their own trade secrets may be compromised when employees input proprietary information (e.g., source code, business strategies, or confidential data) into third-party AI tools. If the AI provider's ToS permit retention or reuse of inputs, or if the provider's systems lack adequate confidentiality safeguards, the act of submitting such information may constitute a failure to maintain reasonable measures to protect secrecy, potentially jeopardizing trade secret status.

The foregoing is illustrative only. Many other aspects of AI systems may involve trade secrets.

Sample AI Trade Secret Litigations

Many AI-related trade secret lawsuits have arisen under classic fact patterns such as employee theft, corporate espionage, and illegal reverse engineering. The following are some examples.

Employee Theft

Tesla, Inc. v. Proception, Inc., 5:25-cv-04963, (N.D. Cal.)

The complaint alleges that Proception, founded by former Tesla engineers, misappropriated trade secrets related to Tesla's AI humanoid robot "Optimus," to develop its human-like robotic hand. Tesla alleges that the employee downloaded thousands of files related to AI-driven robotics workflows and sensor calibration data. Tesla argued that these materials were of great value to the company, having resulted from years of engineering research and many millions of dollars invested. These files allegedly appeared in technical documentation at the competitor company, suggesting direct misappropriation.

Palantir Technologies Inc. v. Guardian AI, Inc., 1:25-cv-01977, (S.D.N.Y.)

Palantir alleged that former employees at Palantir’s healthcare division improperly took proprietary knowledge, which they used to launch Guardian AI. Palantir alleges that Guardian’s AI platform, designed to assist healthcare providers with insurance denial claims, was essentially based on Palantir’s own healthcare denial management tools, including various confidential models and data relating to AI-powered workflows and agents.

X.AI Corp. v. OpenAI, Inc., 3:25-cv-08133, (N.D. Cal.)

The complaint alleged that former xAI employees took and retained xAI trade secrets while departing for OpenAI, and that OpenAI was liable for trade-secret misappropriation. On February 24, 2026, the Northern District of California dismissed the complaint with leave to amend, holding that xAI failed to plausibly allege misconduct by OpenAI itself, including inducement or use of stolen trade secrets.

Illegal Reverse Engineering

C3.ai, Inc. v. Cummins, Inc., C.A. No. N23C-11-106 EMD CCLD (Del. Super. Ct. Aug. 16, 2024)

C3.ai alleged that Cummins improperly reverse engineered proprietary AI software provided under a Master Subscription and Services Agreement in order to develop internal tools. The Delaware Superior Court denied Cummins's motion to dismiss, concluding that C3.ai had plausibly alleged the existence and misappropriation of trade secrets. The case illustrates the importance of contractual restrictions on reverse engineering and use of AI software.

West Technology Group, LLC & CX360, Inc. v. Otter.ai Inc., No. 5:24-mc-80113-VKD (N.D. Cal. filed May 10, 2024). Plaintiffs sought discovery from Otter.ai relating to allegations that a former employee used AI transcription technology to record confidential meetings and expose proprietary information. The matter illustrates emerging trade-secret and confidentiality risks associated with third-party AI transcription tools.

Sanas.AI, Inc. v. Krisp Technologies, Inc., No. 25-cv-05666-RS (N.D. Cal. Dec. 1, 2025): The court denied a motion to dismiss trade secret claims involving AI-based accent-conversion technology, concluding the complaint sufficiently alleged protectable trade secrets and misappropriation.

The foregoing cases illustrate how traditional trade secret doctrines such as employee mobility, contractual restrictions and reverse engineering apply in the AI context. However, AI technology also presents novel mechanisms by which trade secrets may be exposed or extracted, raising questions that existing legal frameworks have not yet fully addressed.

Footnotes

1. The stated basis for rescission of the initial guidance was that Pannu is inapplicable when only one natural person is involved in developing an invention with AI assistance because AI systems are not persons and therefore cannot be “joint inventors” so there is no joint inventorship question to analyze.

2. For a summary of the Initial Guidance, see Copyright Office Guidance on AI.

3. However, blockbuster awards frequently get scrutinized, reduced, or overturned on appeal for issues like causation, evidence, or double recovery.

To read this article in full, please click here.