Founding Partner
Knowing how to patent artificial intelligence starts with understanding that most AI patent rejections are not about the technology but rather how the claims are written. The difference between an approved AI patent and an abandoned application often comes down to specific language choices made before the application is ever filed.
Most AI patent applications that fail at the USPTO fail because the claims describe what the AI does without specifying how it does it, or because the claims read as pure abstract math to an examiner who is looking for a concrete technical improvement. Getting AI claims right requires a different drafting approach than almost any other technology area.
The team at Gearhart Law has put together this guide to the specific drafting strategies that give AI patent applications the best chance of surviving examination.
Why AI Claims Fail: The Section 101 Problem
The biggest obstacle in patenting artificial intelligence is not prior art. It is patent eligibility under Section 101. The USPTO applies a two-step framework to determine whether a claim is directed to something patentable or to an abstract idea.
- Step one asks whether the claim is directed to an abstract idea, a mathematical concept, or a mental process. Machine learning models (weights, activations, loss functions) map easily onto this category in an examiner’s eyes.
- Step two asks whether the claim adds something significantly more than the abstract idea, like a specific, concrete technical improvement that goes beyond just applying math to a general-purpose computer.
Most AI applications fail step two because the claims are written at too high a level of abstraction. They describe the result rather than the technical implementation. They use functional language to cover any possible way of achieving an outcome, without specifying the architecture, training approach, or system design that actually produces it.
What Makes an AI Claim Patent Eligible
An AI claim survives Section 101 scrutiny when it is tied to a specific technical improvement in how a computer system works, not just a description of a result the system achieves.
The claim needs to show that the AI does something technically different from what a generic computer would do and that this difference produces a concrete benefit. Some examples of what this looks like in practice:
- A claim that improves the speed or efficiency of a specific type of data processing by using a novel neural network architecture, not just any architecture that achieves faster processing, but the specific structural choices that make it faster
- A claim that reduces error rates in a defined classification task through a specific training methodology, with the specification showing exactly how that methodology works and why it produces better results than prior approaches
- A claim that improves the performance of a computer system in a defined technical context (network latency, memory usage, or processing load) through a specific AI-implemented method
What all of these have in common is specificity. The technical improvement is defined, measurable, and tied to the particular implementation described in the claims and specification.
System, Method, and CRM Claims: Why All Three Matter
Most AI patent applications should include three types of claims, each protecting the invention from a different angle.
- Method claims cover the steps of the AI process: the training procedure, the inference pipeline, or the specific sequence of operations that produce the result. Method claims are useful for blocking competitors who use the same process regardless of the platform it runs on.
- System claims cover the hardware and software architecture that implements the AI: the specific configuration of components, modules, or subsystems that work together. System claims are useful because they are often viewed more favorably by examiners than pure method claims, since they tie the invention to a tangible configuration rather than a process that could be characterized as abstract.
- Computer-readable medium (CRM) claims cover software stored on a non-transitory medium that, when executed, performs the claimed steps. CRM claims protect the software product itself, regardless of the hardware it runs on.
Filing all three in the same application gives you overlapping protection and multiple fallback positions if the examiner objects to one type. A well-structured AI patent application includes independent claims of each type, with dependent claims that add specificity at each level.
If you are working on an AI invention and want to make sure your claims are structured to survive examination, the patent attorneys at Gearhart Law can help you build an application from the ground up. Reach out for a free half-hour consultation.
How to Draft AI Claims That Describe How, Not Just What
The most common drafting mistake in AI patent applications is describing what the system produces rather than how it produces it. Examiners are trained to look for claims that could be performed mentally or with pen and paper, which makes purely functional AI claims vulnerable to rejection.
Here are the specific drafting approaches that make the difference:
1. Name the architecture
Instead of claiming a system that uses a neural network to classify data, claim the specific type of network, like a convolutional neural network with a defined layer structure, a transformer architecture with a specific attention mechanism, or a recurrent network with a specific gating configuration. Named architectures tied to specific technical functions are harder to characterize as abstract.
2. Specify the training approach
Instead of claiming a model trained on data to produce outputs, describe the training methodology, like the loss function, the optimization approach, and the data augmentation technique, where those choices contribute to the technical improvement being claimed. Training process claims with specific technical parameters survive scrutiny better than generic learning claims.
3. Anchor claims to a technical problem
Every AI claim should identify a specific technical problem and explain how the claimed invention solves it at a technical level. Examiner guidance specifically looks for this nexus between the claimed method and a concrete technical benefit.
4. Use functional language as a supplement, not a foundation
Functional language is useful for covering design-arounds, but a claim built entirely on functional language and any method that achieves a result is a Section 101 target. Use it to broaden coverage around a structurally defined core, not to replace structural definition.
Ready to File Your AI Patent Application?
The difference between an AI patent that issues and one that goes abandoned after three rounds of Office Actions is almost always in the initial drafting. Claims written at the right level of technical specificity, supported by a specification that makes the technical improvement concrete and structured across system, method, and CRM formats, gives you the strongest possible foundation.
Gearhart Law’s team works with AI founders, software developers, and technology companies, offering hands-on AI and machine learning patent prosecution experience. Leave your details and we will be in touch, or call 908.273.0700 for a free half-hour consultation.
Frequently Asked Questions About Patenting Artificial Intelligence
1. Can you patent artificial intelligence?
Yes. AI inventions can be patented when the claims are directed to a specific technical application or improvement rather than an abstract idea or mathematical concept. The key is not whether the invention involves AI, but how the claims themselves are written. Gearhart Law’s patent team has experience prosecuting AI patent applications across a range of technology areas.
2. Why do so many AI patent applications get rejected?
Most AI patent rejections come from the USPTO’s Section 101 patent eligibility framework. Claims that describe what an AI system produces without specifying how it produces it tend to be characterized as abstract ideas or mathematical relationships. The fix is claims that tie the invention to a specific technical improvement in how a computer system works, supported by a specification that explains the technical details at a sufficient level of depth.
3. What types of claims should an AI patent application include?
A well-structured AI patent application includes system claims covering the hardware and software architecture, method claims covering the process steps, and computer-readable medium claims covering the software product. Each type protects from a different angle and serves as a fallback if the examiner objects to one format. Talk to the patent prosecution team at Gearhart Law about structuring your application correctly from the start.
4. How specific do AI claims need to be?
Specific enough to show that the claimed invention involves a concrete technical improvement, not just any method that achieves a result. This typically means naming the architecture type, specifying the training approach where it contributes to the invention, and tying the claims to a defined technical problem and solution. Generic claims that could cover any neural network trained on any data to produce any output will not survive examination.
5. What should an AI patent specification include?
The specification needs to support the full scope of the claims with actual technical detail, including the architecture, the training methodology, performance comparisons showing the technical improvement, and use cases that ground the invention in a specific application. A thin specification that describes results without explaining implementation makes claims vulnerable to both Section 101 rejections and written description rejections.
6. What is the current USPTO guidance on AI patents?
The USPTO issued updated guidance on AI patent eligibility in 2024 that provides specific examples of patent-eligible AI claims. The key question is whether your claims are tied to a concrete technical improvement rather than a general computational result. Aligning your claim language with the guidance examples before filing is one of the most practical ways to reduce Section 101 risk.
7. Can you patent a machine learning model?
Yes, in some cases. A claim to a specific machine learning model architecture that produces a concrete technical improvement (i.e., faster processing, lower error rates, reduced memory usage) in a defined application can be patent eligible. A claim to a generic model trained on data to classify inputs is likely to face a Section 101 rejection. The architecture, training methodology, and specific technical application all matter in how the claim is structured.
8. Do I need a patent attorney to file an AI patent application?
You are not legally required to use an attorney, but AI patent applications are among the most technically and legally complex filings at the USPTO. The claim drafting, specification strategy, and alignment with current USPTO guidance all require both technical and legal expertise. Working with an attorney who has hands-on AI prosecution experience gives your application the best possible foundation. Contact our office to schedule a free consultation today.