United States - Market Insights (Technology)

Introduction

AI is a diverse set of tools that are used by all other industries. Historically, machine learning has been used for classification and prediction, including natural language processing (NLP) and image recognition, such as facial recognition. Now, generative AI tools can create novel text, images, or video based on human prompts, and AI agents are performing many tasks autonomously. AI can replace many tasks that humans currently do because it can be cheaper, faster, and (hopefully!) more accurate. However, there are key limits to the work performed by AI, and those limits are evolving rapidly as developers find new ways to overcome the flaws.

Although AI implicates broad social and political issues, the analysis here focuses on some of the legal issues surrounding AI. In the technology sector, AI is: a valuable asset to be protected (e.g., by patents and trade secrets); a tool for businesses to achieve business goals; and a weapon bad actors use against legitimate businesses and consumers. This chapter addresses all three of these aspects: how to protect IP assets that involve AI; how to leverage the strengths of AI tools while mitigating downside risk; and how to limit exposure to bad actors who are using AI (e.g., cybersecurity).

Protecting AI-related innovation — patents

Inventions that use AI

The tools and techniques of AI, such as neural networks and convolution kernels, are well known. Therefore, patentable novelty typically involves other aspects, such as constructing novel inputs for an AI model, modifying the AI training in a substantial way, utilizing AI output in a creative way, or building AI hardware. For example, merely inserting AI into a process that was previously done without AI would not satisfy patentability, even if the process with AI is faster or better. On the other hand, if the inserted AI has been created specifically for a certain use case, it is likely to be patentable.

Are you using any novel data elements? Because of the widespread use of AI, it is difficult to identify truly new data elements. Are you manipulating the data elements in some ways before they are used for training? In many cases, patentability of an invention is based on innovative manipulations to create synthetic data elements. Typically, synthetic data elements apply mathematical operations to groups of raw data items. The set of possible raw data elements to use may be quite small, but there is virtually unlimited flexibility for creating synthetic data elements. In most cases, the training uses existing tools or methods, so the training itself does not add to patentability. However, if you have a new training process, that may add to patentability.

Generative AI creates new opportunities for innovation. First, innovators are finding new ways to engineer prompts so that the generative AI tool produces the desired output. If the methodology to generate prompts is sufficiently technical, it may be patentable. Second, current versions of generative AI tools have limitations, so there are opportunities to improve the AI tools in ways that are patentable. The improvements can be to: improve the inputs (e.g., retrieval-augmented generation (RAG)); alter the internal processing; or modify the output.

Avoid divided infringement

Applying AI typically occurs in two distinct phases. In the first phase, one or more developers train an AI model and test the model until it produces good results. In the second phase, the model is deployed (e.g., as part of a device or software application) and end users use the deployed model. A single patent claim that requires both training and using an AI model has a problem with divided infringement because there is rarely a single party that performs both the training and the usage.

One solution to the divided infringement problem for AI is to draft two distinct claim sets. A first claim set requires only training the AI model. A second claim set requires only using the trained AI model. The two claim sets (typically in two distinct patents) cover a wide variety of possible infringers.

Subject matter eligibility (section 101)

After the US Supreme Court decision in Alice v. CLS Bank in 2014, “subject matter eligibility” under section 101 of US Patent Law (35 USC § 101) has been a hot issue, particularly for software and AI. In addition, there has been a wide range of opinions about section 101 among patent examiners, appeal board members, and courts.

In 2025, the tide has been turning under United States Patent and Trademark Office (USPTO) Director Squires. In Ex Parte Desjardins (Appeal 2024-000567), the Patent Trial and Appeal Board (PTAB) pointed out that “sections 102, 103, and 112 are the traditional and appropriate tools to limit patent protection to its proper scope.” These sections focus on prior art and enablement rather than categorically eliminating certain subject matter areas. Director Squires also pointed out in the PTAB opinion that excluding patentability of AI innovations “jeopardizes America’s leadership in this critical emerging technology.” Based on this PTAB decision and its incorporation into the Manual of Patent Examining Procedure in 2025, expect fewer and better reasoned section 101 rejections as examiners learn to apply the new guidance.

Human and AI inventors

Inventions created solely by an AI system are not patentable in the United States. See Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022). The more common case is where an AI system contributes to an invention, but one or more humans also contribute to the invention.

Original guidance on AI inventors in February 2024

The USPTO addressed hybrid inventorship in its February 13, 2024 guidance. Because AI inventors cannot be listed in a US patent application, the USPTO focused instead on the contributions of people. The guidance helps determine which humans (if any) made contributions to an invention that are sufficient to be designated as inventors. The primary case is Pannu v. Iolab, 155 F.3d 1344 (Fed. Cir. 1998) (identifying what is required for a person to be designated as an inventor).

Using this guidance, applicants knew that they needed at least one substantial human inventor. Good practices included documenting the human contribution if appropriate. This would safeguard against potential future litigation that could invalidate patents. In particular, it can be useful to have contemporaneous emails or documents that clearly show the human inventive aspects. If possible, it is useful to document how the human inventors used an AI system as a tool.

Updated USPTO guidance on inventorship in November 2025 creates uncertainty about AI

The guidance released November 28, 2025 rescinds the previous guidance (i.e., a significant contribution is not required) and asserts that AI is just a tool “like a microscope.”

Because the USPTO has rescinded the requirement of a human making a significant contribution, and proposes that AI is “just a tool,” the current USPTO guidance is unclear what is intended when inventions have contributions from both people and AI systems.

There are at least three reasons to make business decisions based on the earlier guidance.

• The USPTO is bound by statutes and court decisions. Any guidance that is inconsistent with patent statutes or court decisions is not binding.
• As the new guidance points out, the USPTO presumes the named inventors are correct. Evaluating inventorship will likely occur only in court proceedings (e.g., patent litigation), so the USPTO guidance has no substantial impact.
• A person cannot claim inventorship rights for features the person did not conceive himself/herself. Each inventor must sign a declaration that he/she is the actual inventor.

As a practical matter, applicants should continue with inventive processes where humans make significant contributions. By keeping the people in the inventive process and creating internal documents to memorialize the human contributions, applicants will be prepared if inventorship is questioned later in litigation.

Patents versus trade secrets for AI

Protection of AI-related inventions

Patents are subject to examination and many patent laws, and this is even more burdensome for AI-related inventions. Because trade secrets are not subject to these requirements, they provide an attractive alternative to patents. In particular, trade secrets:

• do not require a human inventor;
• do not require proof of novelty;
• do not require proof of subject matter eligibility (i.e., no consideration of whether an AI invention is an “abstract idea”); and
• are not subject to the wide variation of patent examiners.

Advantages of trade secrets include: coverage of material that is not patentable; having no issues related to an assigned examiner; avoiding the uncertainty of subject matter eligibility rules or a potentially protracted timeline; and potential to extend protection indefinitely. Advantages of patents include: patent infringement does not require proof of misappropriation or intent; patents are enforceable regardless of reverse engineering or independent development; patents can be used by marketing; and researchers can publish their work when it is protected by patents.

Protection of AI-related data

In the context of AI, there are almost always at least two data sets to protect as trade secrets. First, the training data for the AI system should be protected. In fact, the training data may be the most valuable asset for many AI use cases. In particular, it can take a considerable investment of time and money to collect and classify the training data. And the training data can be supplemented in the future to build even better training sets.

Second, a trained AI model can be protected as a trade secret. A trained AI model can be considered a large batch of parameters (e.g., neural network weights), and these parameters are not visible when the model is used.

Using AI tools

People have been using AI for decades. Historically, people have used machine learning tools to classify data and make predictions, but since 2022 people have been interacting with generative AI systems such as ChatGPT, Claude, and Perplexity to do much more. The value of these tools is undeniable, but the downside risks are often overlooked.

Use of generative AI puts your data at risk

When using a tool like ChatGPT, Claude, Meta Llama, Microsoft Bing, or other generative AI system, it seems like a private conversation. It is not. Treat any use of a public AI tool as a public announcement to the world. It may be public and it may be used to train the model. This forfeits any trade secret protection for that information and potentially forfeits patent rights. Therefore, confidential information should never be entered in one of the public systems. Private, paid, or enterprise versions of AI tools may be better, but it depends on the contract terms. Does the contract guarantee confidentiality? Does the contract guarantee no use for training? Can you trust the vendor?

No attorney-client privilege for AI-generated content

A federal court decision held that there was no privilege for content generated by an AI system. The judge ruled that documents generated by an AI tool were not protected by attorney-client privilege or the work product doctrine, even though the content was later shared with attorneys. See United States v. Heppner, No. 25-cr-00503-JSR (S.D.N.Y. Feb. 10, 2026).

The court reasoned that the communication with an AI platform was not communication between a party and counsel and there was no reasonable expectation of confidentiality. Similar reasoning is expected in future cases.

AI and cybersecurity

Cybercriminals are using AI

Generative AI produces much higher quality text to deceive us, generative AI produces very high quality images and photos to deceive us, and generative AI can spoof the voices and mannerisms of people you know. Because of this, AI is helping the bad actors to execute attacks more quickly, more cheaply, and with higher quality. Criminals use AI to take advantage of the weakest link in cybersecurity: people. See Verizon 2025 DBIR, IBM 2025 Cost of a Data Breach Report, and FBI IC3 2024 Report.

Cybersecurity is using AI

The good guys are also using AI. AI helps cybersecurity tools identify risks and threats more quickly, AI helps cybersecurity tools identify a greater variety of threats (without requiring threat signatures), and AI enhanced cybersecurity can correlate anomalies and suspicious activity from many sources in the aggregate.