Artificial intelligence is generating a wave of innovation, and with it, a wave of questions about IP protection. Can you patent AI models? Can an AI system be listed as an inventor? What if AI helped you develop your invention? These are no longer hypothetical questions. They’re ones the U.S. Patent and Trademark Office is actively grappling with, and the answers have changed significantly in the past few years.
At Goldstein Patent Law, we work with inventors and technology companies navigating the rapidly shifting landscape of AI patent law. Here’s what you need to know right now.
The Short Answer: It Depends on What You’re Patenting
Can you patent AI models in a general sense? Not as abstract systems or mathematical concepts. But can you patent specific AI-related innovations, including architectures, training methods, and real-world applications of AI? Absolutely, provided you meet the standard patentability requirements.
The question isn’t really whether AI technology is patentable in principle. It’s whether your specific claims are structured in a way that clears the legal hurdles the USPTO applies to this category of invention. Getting that structure right is where most AI patent applications succeed or fail.
The Core Legal Challenge: Patentable Subject Matter Under § 101
The biggest obstacle for anyone asking whether they can patent AI models is Section 101 of the Patent Act, which defines what kinds of inventions are eligible for patent protection. Under the Supreme Court’s Alice/Mayo framework, abstract ideas, mathematical concepts, and mental processes are not patentable on their own.
Many AI inventions run directly into this wall. A machine learning algorithm, at its core, is a mathematical process. A neural network architecture, described in the abstract, may look to an examiner like a mathematical concept dressed in technical language. The USPTO has rejected a large number of AI-related claims on exactly these grounds.
The path through § 101 requires showing that your claimed invention is directed to a practical application, not just the abstract idea or math itself. An AI system that detects anomalies in a network is abstract. An AI system that detects anomalies in industrial sensor readings to prevent equipment failure, integrated into a specific technical process, is more likely to clear the bar.
The USPTO’s 2024 Guidance on AI Patent Eligibility
In July 2024, the USPTO issued a guidance update on patent subject matter eligibility specifically addressing AI inventions. This was a significant development for anyone asking whether they can patent AI models, and it has meaningfully shaped how examiners approach these applications.
The 2024 update introduced a new set of examples (Examples 47 through 49) specifically addressing AI inventions, providing insight into how examiners apply the subject matter eligibility analysis to AI technologies.
Example 47 focuses on an artificial neural network designed to detect anomalies, illustrating how claims involving specific applications of neural networks can meet eligibility criteria by demonstrating an improvement in computer technology. Example 48 involves AI-based methods for analyzing speech signals and separating desired speech from background noise, showing how applying AI to solve specific technical problems in signal processing can be patent-eligible. Example 49 involves a method for personalizing medical treatment using an AI model, illustrating how AI integrated into practical applications providing concrete benefits can meet eligibility requirements.
An August 2025 memo from the USPTO to patent examiners further reinforced that an AI invention may be subject matter eligible when presented as a practical application.
The consistent thread across all of this guidance is that specificity and practical application are everything.
Who Can Be the Inventor? The Human Requirement
One of the most significant questions surrounding AI patent law is not just whether you can patent AI models, but who gets listed as the inventor when AI played a role in developing the invention.
The answer, under current U.S. law, is unambiguous: only humans can be inventors.
The USPTO’s revised inventorship guidance, released in November 2025, confirms that AI systems cannot legally be inventors. AI tools, whether generative, analytical, or model-based, are treated the same as laboratory equipment, software, or research databases: they may assist in the creation of inventions, but they cannot conceive those inventions.
The guidance finds that humans who use AI tools when developing new ideas do not forfeit their inventorship status in patent applications. The guidance likens AI systems to traditional equipment that assists in development, with the humans deploying those tools still being the ones who “conceived” the idea.
This is important practically. If you used AI tools to help develop or test your invention, that does not disqualify you as the inventor. The human who directed the process, interpreted the outputs, and exercised creative judgment is the inventor. The AI is the tool.
The 2024 USPTO guidance explicitly clarifies that for subject matter eligibility purposes, how an invention was developed, whether with or without AI assistance, is not relevant to the eligibility analysis. AI-assisted inventions are evaluated using the same Alice/Mayo framework as any other technology.
What Aspects of AI Can Actually Be Patented?
For inventors and developers asking whether they can patent AI models, it helps to break the question down by what specifically you’re trying to protect.
Specific neural network architectures. A novel architecture designed for a particular technical purpose, described with enough specificity, may be patentable. The key is tying it to a concrete technical improvement rather than describing it in purely abstract terms.
Training methods and processes. A new method for training a model, particularly one that solves a known technical problem in training efficiency, accuracy, or computational cost, can be patentable as a process claim.
AI applied to a specific technical field. Applications of AI to solve specific real-world problems in manufacturing, healthcare, logistics, signal processing, and similar fields tend to fare better in examination than broad claims about AI in the abstract.
AI-integrated systems. A system that combines hardware, software, and an AI component to accomplish a specific technical function is often more defensible than a pure software or algorithm claim.
Improvements to existing AI technology. If your innovation improves how an existing AI system performs, you may be able to patent the improvement independently of the underlying model. This is similar to how improvement patents work across any technology category, as we cover in our post on patenting improvements to existing products.
What Cannot Be Patented
Understanding the limits is as important as understanding the possibilities.
Abstract algorithms. A mathematical formula or statistical method, even if it underpins a sophisticated AI model, is not patentable on its own. The claim must do something with it in a concrete, practical context.
Generic AI applications. Claiming “using AI to analyze data” or “using machine learning to improve results” without meaningful technical specificity will almost certainly be rejected under § 101.
AI as an inventor. As discussed above, listing an AI system as an inventor on a U.S. patent application is not permitted under current law. This is settled regardless of how significant the AI’s contribution was to the development process.
Ideas without implementation. A concept for an AI system, without an enabling description of how it actually works, will fail the written description and enablement requirements regardless of § 101 issues.
The Non-Obviousness Hurdle for AI Patents
Even if an AI-related invention clears § 101, it still has to satisfy the non-obviousness requirement under 35 U.S.C. § 103. In a field that moves as fast as AI, this analysis is particularly nuanced. What was a novel architectural approach two years ago may already be well-established in the field.
The PHOSITA standard (Person Having Ordinary Skill In The Art) is applied at the time of the invention, not today. But in AI, ordinary skill in the art is a moving target. A claim that might have been non-obvious in 2020 could be considered obvious in light of how rapidly techniques like transformer architectures, diffusion models, and reinforcement learning from human feedback have proliferated.
Conducting a thorough prior art search before filing is critical. Our post on non-obviousness in patent law covers the full legal framework in detail, including the secondary considerations that can help establish non-obviousness even when prior art is dense.
Strategic Considerations for AI Patent Portfolios
For companies building AI-powered products, the question of whether they can patent AI models is often just the beginning. A broader IP strategy matters.
Layer your protections. Patent protection for technical implementations, trade secret protection for training data and model weights, and copyright for software code can work together. No single protection covers everything, and understanding what each covers is essential. Our overview of the difference between patent and trademark is a useful starting point for thinking about IP more broadly.
File early. The U.S. operates on a first-to-file system. In a field evolving as quickly as AI, timing matters enormously. A provisional application can establish your filing date while you refine your claims and strategy.
Think about claim scope carefully. Overly broad claims will be rejected. Overly narrow claims may not provide meaningful protection as the technology evolves. Getting this balance right requires someone who understands both the technology and the legal landscape.
Consider continuation strategy. As your AI product evolves, continuation applications allow you to capture new features and improvements under the priority date of your original filing. This is a powerful tool for maintaining relevance in a fast-moving field.
How to Start Your AI Patent Application
The patent application process for AI inventions follows the same general steps as any patent application, with the added complexity of § 101 analysis. A thorough prior art search, careful claim drafting with practical application front and center, and anticipation of likely examiner objections are all especially important in this space.
If you’re early in development and not yet ready to commit to a full nonprovisional, a provisional application can lock in your filing date while you continue building. The USPTO’s AI-related resources are also worth reviewing for current examiner guidance.
Frequently Asked Questions
Can you patent an AI model itself, such as a large language model or image generation model?
Not in the abstract. But specific technical implementations, training methods, or applications of such models may be patentable if they are tied to a concrete practical application and meet all other patentability requirements. The model as a general concept is not patentable; a specific, novel, non-obvious application of it may be.
Can an AI system be listed as the inventor on a patent application?
No. Under current U.S. patent law, only natural persons (humans) can be listed as inventors. AI tools used in the development process are treated similarly to laboratory equipment and do not qualify as inventors.
What if AI generated the core idea for my invention?
The inventorship analysis focuses on who conceived the invention in the legal sense: who exercised creative judgment, directed the process, and recognized the significance of the output. If a human played that role, even using AI tools extensively, the human is the inventor. You should discuss the specific facts of your situation with a patent attorney before filing.
Are AI training datasets protectable?
Datasets themselves are generally not patentable. However, novel methods for curating, processing, or using a dataset in an AI training process may be. Copyright may also provide some protection for original elements of a dataset, though this is a contested and evolving area.
How long does it take to get a patent on an AI invention?
The same general timelines apply as for any utility patent: roughly two to three years for standard examination, or six to twelve months through the USPTO’s Track One expedited program. AI applications may face additional rounds of Office Actions due to § 101 subject matter eligibility issues, which can extend the timeline.
Protecting Your AI Innovation Starts Here
The answer to whether you can patent AI models is not a simple yes or no. It depends on what you’re claiming, how you’re claiming it, and how well your application navigates the subject matter eligibility framework that the USPTO applies to AI technology. The rules are still evolving, and the decisions you make in drafting your application now will shape your protection for the full term of the patent.
Contact Goldstein Patent Law to schedule a consultation. We’ll evaluate your AI innovation, assess the prior art landscape, and build a filing strategy designed to hold up under examination.
