Specialist Chapter: Procuring Patent Protection for AI Inventions

This is an Insight article, written by a selected partner as part of IAM's co-published content. Read more on Insight

In summary

This article discusses patenting strategies for artificial intelligence (AI) inventions in today’s AI landscape. It also reviews the current legal framework for patenting AI technology in the United States and examines two case studies of AI inventions.

Discussion points

  • How to identify AI inventions in today’s AI framework
  • How best to protect different types of AI inventions with patents
  • What are the obstacles in patenting AI?

Referenced in this article

  • Trinity Info Media, LLC v Covalent, Inc
  • Thaler v Vidal
  • In re Bd of Trustees of Leland Stanford Junior Univ
  • SAP Am, Inc v InvestPic, LLC
  • CardioNet, LLC v InfoBionic
  • McRO v Bandai
  • United States Patent and Trademark Office, 2019 Revised Patent Subject Matter Eligibility Guidance
  • Open AI

Today, generative AI is a groundbreaking field. Generative AI uses large language model (LLM) technology that helps with natural language processing and understanding. ChatGPT by OpenAI is a well-known chatbot that implements LLMs. Since its release in November 2022, ChatGPT became a pioneering generative AI model owing to its accessibility and ability to generate coherent and contextually relevant text in a conversational manner. OpenAI is also rolling out its commercialised platform known as GPT Store.[1] The GPT Store allows various generative pre-trained transformer (GPTs) in different categories to become searchable and customisable for a user to build their own generative AI applications. This marketplace to monetise customised GPTs is similar to the App Store framework, where customised GPTs are hosted, developed, promoted and evaluated on OpenAI platforms, providing a marketplace for tailored AI models and services for specific tasks.

Owing to advancements in AI, companies are pursuing their intellectual property rights to keep and obtain a competitive edge in the AI landscape. Since January 2020, over 20,000 US patent publications relating to a ‘neural network’ or ‘machine learning’ have been filed. In this article, we will discuss strategies to procure patent rights over AI technology: what are AI inventions? How best to protect them with patents? What are the obstacles?

Identifying AI inventions

Today, the AI framework has multiple actors. Identifying AI inventions in the AI framework, the associated actors and the owners are a crucial step in any intellectual property strategy.

Patenting AI frameworks

AI in a product

AI inventions can be integrated in different products, including robotic surgery tools, autonomous driving vehicles, and virtual and augmented reality. To obtain patent protection, an applicant should identify an innovative AI component within the AI product and assess whether the innovation lies within the AI component or within a particular application of the AI component within the product. Consider a vision neural network model that is engaged in autonomous driving to detect road conditions by capturing images of a surrounding environment. An applicant should assess whether the novelty lies in training and inference of the vision neural network model, or in how a control mechanism uses the output of the vision neural network model. A detailed case study of a robot surgical product involving AI is also discussed below.

AI as a module

AI inventions can be contained in a ‘module’. AI companies develop neural network modules that are then deployed on their servers. In this scenario, the AI architecture (design of an AI structure, including types and layers of a neural network), training and finetuning of AI (using data to teach and fine-tune the neural network to achieve certain functions), AI testing (using the trained neural network to perform certain functions) and AI deployment (releasing the neural network into the real world to perform tasks) are all conducted by the same company. Therefore, a patentee may pursue a patent strategy directed to building a new AI model architecture, training the AI model with a training objective that achieves a certain function and using the AI model to perform a new task at inference in the real world.

AI as a service

To streamline AI resource utilisation, downstream AI companies may use AI as a service (AIaaS). In a nutshell, AI companies may access and utilise AI tools, algorithms and models (eg, a commercialised GPT model subscription) provided by a third-party provider, without considerable investment in hardware, software and expertise. In this scenario, a patentee should identify components and functions within an AI product that are built on top of AIaaS and are owned or conducted by the patentee. For example, the patentee may pursue a patent strategy directed at how systems communicated with external AI models to perform certain tasks via application programming interfaces (APIs).

AI as a marketplace

In a foreseeable future of AI marketplaces, an AI ecosystem may be created that builds, trains, publishes, trades and uses customised AI models by multiple parties. For example, the GPT Store allows creators to build and sell their own GPTs. In this ecosystem, AI inventions may occur at different levels: building and pretraining of the original GPT (eg, performed by the original GPT provider), hosting and providing APIs that customise the GPT (eg, performed by the marketplace platform), facilitating the publication and transactions of customised GPTs at the AI marketplace (eg, performed by the marketplace platform) and building any customised AI infrastructure by purchasing or subscribing to customised GPTs from the marketplace (eg, performed by a downstream AI company). Therefore, a patentee may need to identify the appropriate actor within the AI ecosystem to obtain patent coverage and minimise divided infringement.

Patenting AI data

It is not a secret that AI models use and produce data. This data can be protected using patents.

Training data

Training data is a secret sauce that sets an AI model apart from its competitors. AI models can be trained on public datasets that are in a public domain, on a combination of public and private datasets, or only private datasets that are often kept as trade secrets. However, with the onset of regulation at state and federal levels directed at keeping AI models fair, unbiased, and responsible, keeping the training data a trade secret may no longer be a viable option. There are options, however, for companies seeking to obtain patent protection for their training data:

  • Unique data characteristics, including data structures and constituents may be claimed together with the AI model.
  • When an AI model combines different training datasets (eg, public and private or private and private), the claims may be directed at the combination or at specific steps for achieving the combinations.
  • AI models can process different datasets differently. Often, AI models are pre-trained by one company on one dataset and are fine-tuned on a specialised dataset for a particular purpose by a different company. This may result in different training data passing through different layers of the AI model, and claims may be directed to the type of training data that trains different layers and how the data is split during AI model pre-training and finetuning.
  • Sometimes synthetic data is created because training data is unavailable or is difficult to obtain. Synthetic data emulates scenarios in a real-world environment that are not covered by the original training dataset. Steps directed to identifying scenarios that are not covered by the original training dataset and creating synthetic data are all eligible for patent protection.
  • When an AI model is trained with training data, different training samples can be given different weights. In other words, not all samples are treated equally. Identifying which samples in the training dataset should be given more or less weight during training and which samples may be suppressed and ignored may also be covered by a patent.

AI inputs

A trained AI model receives data as input and generates an output as a result. The input to the AI model can largely be protected in ways similarly to the training data. In some instances, an input to an AI model can be a combined input from multiple sources, including an input from multiple AI models and an input augmented with retrieval augmented generation techniques. Patent protection may be directed to novel techniques for combining different data that serves as input to an AI model.

AI outputs

Today, AI outputs created by an AI model that is an inventor are not eligible for patent protection in United States.[2] On the other hand, AI outputs created by an AI model that serves as a tool in the inventive process are protectable. Protecting AI outputs as a product-by-process (eg, claiming a genome sequence generated by a particular AI model using particular input data) is also an option. Whether AI can or cannot be an inventor is a question that is being considered by the US Patent Office and Congress in the United States, and other bodies worldwide. Thus, one should be mindful to changes, if any, to AI inventorship. Further, there is ongoing research into LLM models that create three-dimensional outputs, impacting the medical devices, 3D printing, life sciences and material science industries. In addition to utility patents, the non-functional appearance of the three-dimensional outputs can be also be protected by design patents.

Overcome obstacles towards patenting AI

Obtaining patents for AI technologies present unique challenges under the current legal framework, particularly in meeting the eligibility requirements under section 101 of the Patent Act, because AI technologies often involve software and mathematical algorithms. The uncertainty in the current state of section 101 law is largely attributed to the two-step Alice framework for eligibility, established by the Supreme Court’s 2014 decision in Alice Corp Pty v CLS Bank Int’l, 573 US 208 (2014). Despite these challenges and uncertainty in section 101 law, we will delve into recent case law for guidance relevant to AI technologies and then explore practical considerations for successfully obtaining strong AI patents.

District courts: Rule 12(b)(6) motion to dismiss AI patent claims on section 101 grounds

In U.S. patent cases, Rule 12(b)(6) allows defendants to seek early dismissal under section 101. If granted, the case is dismissed without the plaintiff having the opportunity to engage in extensive discovery, which is crucial for gathering evidence and building a stronger case.

Recent cases (eg, Power Analytics, Health Discovery and Recentive), witnessed successful 12(b)(6) motions dismissing AI patent claims.[3] The courts rejected arguments based on differences from the human brain: for example, ‘machine learning algorithms are unique since they process information differently from how the human brain could or would’ or ‘humans could not perform the patented processes, because the data and algorithms are too complex’ (see, eg, Recentive at 17-20). These decisions underscore the importance of carefully crafting strong AI patent claims resistant to 101 challenges.

Federal Circuit

It is difficult to apply Alice with consistency as illustrated in Federal Circuit section 101 cases, hindering a unified interpretation. Despite this complexity, we distil guidance on effective claim drafting for AI technologies, addressing section 101 challenges: Trinity, Stanford and SAP found claims ineligible, emphasising abstract nature, while CardioNet, Thales and McRO found claims eligible because of tangible technological improvements.

Federal Circuit: claims not eligible

Trinity Info Media, LLC v Covalent, Inc, 72 F.4th 1355 (Fed. Cir. 2023) (Trinity)

The court held that a human being incapable of matching processing speed does not make an abstract process patent eligible. The Court explained that ‘Trinity’s asserted claims can be directed to an abstract idea even if the claims require generic computer components or require operations that a human could not perform as quickly as a computer.’[4]

In re Bd of Trustees of Leland Stanford Junior Univ, 991 F.3d 1245 (Fed. Cir. 2021) (Stanford)

The patent describes a method for more accurate prediction than prior art in genetic sequencing. The Court determined this is an improvement to an abstract idea, not a technological improvement. The Court concluded that the claims were directed to abstract ideas: ‘the use of mathematical calculations and statistical modelling,’ and this was ‘merely an enhancement to the abstract mathematical calculation of haplotype phase itself.’[5] The Court distinguished McRO and CardioNet, holding that they ‘involve practical, technological improvements extending beyond improving the accuracy of a mathematically calculated statistical prediction’.[6]

SAP Am, Inc v InvestPic, LLC, 898 F.3d 1161 (Fed. Cir. 2018) (SAP)

The Court found claims directed to statistically analysing investment information and reporting the results to be abstract.[7] Specifically, the Court distinguished McRO on the grounds that McRO was directed ‘to the creation of something physical’, unlike the quantitative predictions in SAP.

Federal Circuit: claims eligible

CardioNet, LLC v InfoBionic, 955 F.3d 1358, 1368 (Fed. Cir. 2020) (CardioNet)

The District Court determined on a Rule 12(b)(6) motion that a medical device patent was ineligible as it was directed at an abstract idea. The Court reversed, finding that the claims are instead directed to ‘an improved cardiac monitoring device’, confirmed by various specific technological improvements detailed in the written description. While the claims provide a statistical prediction, the Court found that they provide an improvement to cardiac monitoring technology as opposed to an abstract idea by providing an improved prediction of heart arrhythmia based on heart monitoring data.

Thales Visionix v United States, 850 F.3d 1343, 1348 (Fed. Cir. 2017) (Thales)

The claims are related to an inertial tracking system. The Court found that the claims, which admittedly included mathematics, were patent eligible, where ‘the application of physics create an improved technique for measuring movement of an object on a moving platform’.[8] The Court found that the claims here resulted in a system that reduces errors in an inertial system that tracks an object on a moving platform.

McRO v Bandai, 837 F. 3D 1299, 1311 (Fed. Cir. 2016) (McRO)

Often referred to as the ‘animation invention’ case, the claims related to automated lip synchronisation for animated characters. The Court found the claims patent-eligible, emphasising the specific improvement in computer animation and the use of rules to automate a previously manual process.

USPTO Guidance

The USPTO has provided examples for guidance on patenting AI inventions, including its 2019 Revised Patent Subject Matter Eligibility Guidance and ex parte Hannun, 2018-003323 (1 April 2019) (designated as ‘informative’) (Hannun). These examples were provided in 2019 and incorporated into the Manual of Patent Examining Procedure in 2020, but do not reflect the latest developments from the Federal Circuit. We expect that the USPTO will provide new guidance in 2024, following President Biden’s executive order (EO) on AI issued on 30 October 2023.

Practical considerations for patent application drafting

Use of machine learning alone doesn’t make claims eligible (see Power Analytics, Health Discovery and Recentive). Examiners, aligning with court decisions, will increasingly treat generic machine learning models and iterative training methods akin to generic computer components for section 101. Mere distinctions from a human brain may not suffice for AI claims’ eligibility.

Provide details. Courts seek specifics about the model and functions, for example: how the machine learning engine is configured and any particular structure and processes for performing the functions (eg, how to compare the real-time and predicted values, how to pick the threshold values and how to update the model).

Avoid description only in broad functional terms with little guidance on model parameters or training technique. Be careful with the description of using ‘any suitable’ machine learning or iterative training techniques. Instead, describe specific functions, parameters and training techniques, and emphasise the inventiveness of these specific features.

Emphasise the AI invention’s link with ‘something physical’. Courts underscore the importance of physical improvements, including the use of mathematics to achieve improvements in physical things (see CardioNet, McRO and Thales). For instance, frame outputs as ‘generated’ (eg, audio, images, videos, text converted from speech and code) rather than ‘predictions’ where applicable (see Stanford and SAP). For example, in AI for material discovery, outputs can be framed as alloy compositions and treatment parameters.

Articulate additional advantages beyond improved prediction accuracy. Merely stating prediction or enhanced prediction without tying it to physical improvements may not suffice for eligibility, as seen in Stanford and SAP.

Stay adaptable to ongoing developments. Notably, President Biden’s EO on AI, issued on 30 October 2023, requires the USPTO director to – in an effort to address innovation in AI and critical and emerging technologies – publish guidance addressing inventorship and the use of AI and other considerations at the intersection of AI and IP on patent eligibility. Keep abreast of legislative changes and court decisions to enhance AI patenting strategies. Align drafting with recent legal developments for robust AI patent portfolios.

AI case studies

Below are case studies that showcase patent strategies for AI inventions.

Case study 1: AI assisted surgical system

Consider an AI company training an AI model to assist in a well-known type of laparoscopic surgery. The trained model reviews a video of a surgery in real time and makes surgical recommendations. The recommendations are displayed via icons on a graphical user interface that is viewable by the medical team. Eventually, portions of the surgery may be controlled by the surgical system in response to the trained model’s recommendations and predictions.

To shape a patent strategy for this AI surgical system, the first question is whether the company wishes to disclose sufficient detail of the AI architecture and training to enable others to develop a similar model without undue experimentation. If the company is unable to disclose sufficient detail, a utility application may eventually be rejected by the USPTO for lack of enablement or as an abstract idea under section 101. One option is to focus on protecting aspects other than the model itself. For example, the icon displayed on the graphical user interface may be eligible for a design patent. Design patents protect the design of the icon and do not require details of the AI model to be disclosed. Alternatively, if the company is able to disclose sufficient detail but does not want to publish these details prior to obtaining patent protection, the company can file a utility patent application with a non-publication request. While the non-publication must be withdrawn if and when the company decides to file internationally, filing a non-publication request is a way of keeping the technology a secret until a patent issues.

Assuming the company is able to disclose sufficient details regarding the model, the next key question relates to patent eligibility and is, what makes this invention special? Is there something unique about the training data or the model itself? Does this invention improve the functioning of a computer or improve a technical field? What is the system doing that could not be done by a human? The improvements should be detailed in the specification of a utility patent application. If the model is trained in a conventional manner using publicly available datasets to generate a recommendation that could be generated by a human, then it would likely be beneficial to include utility claims describing the icon output or describing the iteration of the surgical system that controls portions of the procedure in response to the model’s recommendation and predictions. If the model itself or the training of the model is unique, then patent claims can be drafted focusing on the use of the model as well as the training of the model.

Case study 2: AI outputs

Assume the company uses AI to improve a previously known 3D product by inputting requests and refining the AI model’s output. Unlike case study 1, meeting the enablement requirement will likely not be difficult if known manufacturing techniques can be used to make the improved product. Moreover, patent eligibility is less of a concern when the claimed invention is an improved 3D product. But the claim strategy may need to be carefully assessed in view of inventorship. For example, as the AI model itself cannot be an inventor, patent claims shall not be directed to AI output data alone. If patent claims are directed to the final product, however, the AI model merely served as a tool for the personnel who provided inputs and refined the AI model’s output, and it is the personnel who discovered the improved product using the AI model. As such, the personnel who utilise the AI model to arrive at the final product are identified as inventors when claims are pursued for this improved product. Moreover, the company could pursue design patent protection for the improved product if the improved product involved a new design.


[1] GPT store, ‘Discover the GPTs and plugins of ChatGPT’ (https://gptstore.ai/).

[2] Thaler v Vidal, 43 F.4th 1207 (Fed. Cir. 2022)

[3] Power Analytics Corporation v Operation Technology, Inc, C.A. No. 16-1955 (C.D. Cal. July 13, 2017) (finding claims using a ‘machine learning engine’ to be ineligible since the patent ‘does not specify how the engine is configured’); Health Discovery Corp v Intel Corp, 577 F. Supp. 3d 570 (W.D. Tex. 2021) (holding ineligible claims on a machine learning algorithm as directed solely to unpatentable mathematical ideas); Recentive v Fox, DDE-1-22-cv-01545 (D. Del. Sep. 2023)(finding claims using machine learning algorithms ineligible).

[4] Trinity at 1364.

[5] Stanford at 1250-1251.

[6] id. at 1251

[7] SAP at 1161.

[8] Thales at 1349.

Unlock unlimited access to all IAM content