Can You Patent AI? What the USPTO Actually Grants in 2026

About 80,000 of the utility patent applications the U.S. Patent Office received in 2020 involved AI, and the USPTO reports that AI now appears in more than 18% of all utility patent applications it receives, figures that underscore how rapidly AI technology is transforming the U.S. patent system. Yet Section 101 subject matter eligibility remains the primary challenge in AI patent applications, with examiners applying heightened scrutiny to algorithm and software-based inventions compared to hardware-based patents that more readily demonstrate a practical application. Many founders filing AI patents lose at a hurdle they never saw coming, not because their technology is weak, but because their claims were drafted wrong.

This article explains exactly what it takes to answer the question "can you patent AI" with a yes that holds up. You will learn the legal standards that determine whether an AI invention qualifies for patent protection, how the USPTO's inventorship rules for AI-assisted inventions changed when its 2024 guidance was replaced in November 2025, how to approach algorithm and model patents, what examination looks like, and what to do before you file.

The Short Answer to Whether You Can Patent AI and Why It Is More Nuanced Than Yes or No

The correct answer is: you cannot patent artificial intelligence as a category, so when founders ask can you patent AI, the honest answer is it depends entirely on how the invention is framed, but you can patent what your AI system specifically does. The U.S. Patent Office does not grant patents on AI technology in the abstract, on the general concept of machine learning, or on the broad idea of a neural network, as these fall into the category of abstract ideas excluded from patent eligibility. It does grant patents on specific, technically grounded implementations of AI technology that solve a defined problem in a novel, non-obvious way.

3 Legal Hurdles Every AI Patent Application Must Clear | Source: USPTO AI Patent Activity Report; HLK-IP AI Patenting Report

A useful analogy: you cannot patent electricity, but you can patent the light bulb. The question is never "is this AI?" but "does this AI-powered system or method solve a specific technical problem in a significant manner that no prior art has addressed?" If your answer involves concrete technical results tied to a specific implementation, you likely have something worth protecting.

For a deeper look at how these principles apply to specific software architectures, see The Ultimate Cheat Sheet for Patents on Artificial Intelligence.

Why the USPTO Does Not Grant Patents on AI Itself, Only on What AI Does

Under 35 U.S.C. § 101, a patentable invention must fit one of four categories: a process, machine, manufacture, or composition of matter. The Supreme Court's ruling in Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014), established that implementing abstract ideas on a generic computer is not patent-eligible under United States patent law. That decision controls how examiners evaluate AI patent applications today. An AI invention framed only as "a computer uses machine learning to perform task X" almost certainly fails under Alice. An invention framed as "a system that uses a specific convolutional architecture to reduce false positives in real-time fraud detection by a measurable margin" has a fighting chance.

Before asking whether your AI invention is patentable, define the specific technical problem it solves and the specific technical result it achieves. That framing determines everything.

The Four Patent Categories and Which Ones AI Inventions Fit Into

Most ai innovations fall under two of the four statutory categories: process (a method of doing something using AI technology) and machine (a system or computer that implements the AI). Knowing which category applies is a prerequisite to answering can you patent AI for any specific invention. Design patents are rarely relevant to AI, and plant patents have no application here.

WIPO's study of AI patenting found that machine learning is the dominant AI technique disclosed in patents, included in more than one-third of all identified inventions, and that computer vision is mentioned in 49% of AI-related patents. The leading application fields were telecommunications and transportation (15% each) and life and medical sciences (12%). Recommendation engines, facial recognition, and speech processing are familiar examples.

The category you choose affects how broadly your patent claims can be written. Process claims capture the method regardless of what hardware runs it. Machine claims capture the configured system. Filing both types in the same application maximizes coverage and gives you prosecution flexibility if an examiner challenges one set of claims. Most registered patent attorneys drafting AI patent applications file both. For a detailed breakdown of how utility patent categories apply to AI inventions, see Utility Patents Explained: An Essential Guide for Every Inventor.

The Three Legal Hurdles Every AI Patent Application Must Clear

Every AI patent application must satisfy three independent legal requirements: subject matter eligibility under § 101, novelty under § 102, and non-obviousness under § 103. These are the three gates any founder must clear before the question of can you patent AI resolves in their favor. The first hurdle is where many AI applications stumble. According to the USPTO Office of the Chief Economist's Adjusting to Alice report, the likelihood of a first Office Action with a § 101 rejection rose 31% in the 18 months after the Supreme Court's Alice decision in the technology areas it affected.

§ 102 novelty and § 103 non-obviousness are equally important but addressed through prior art analysis rather than claim drafting strategy. The prior art landscape in AI is moving fast. WIPO has documented an eightfold increase in generative AI patent publications since 2017, and its July 2026 update found that more than 56,000 new generative AI patent families were published in 2024 and 2025 combined, more than in the entire 2014 to 2023 period, reflecting how rapidly AI innovations are reshaping the global patent landscape. Commissioning a prior art search before drafting claims is not optional; it is the baseline for any credible AI patent strategy.

Can You Patent AI? What Artificial Intelligence and Patent Law Actually Require From Inventors Right Now

The legal framework governing artificial intelligence and patent law rests on a Federal Circuit ruling and USPTO guidance that together define who can invent and what they must have done within the U.S. patent system.

How U.S. Law Arrived at Today's AI Inventorship Rules: 4 Defining Moments | Source: USPTO AI Inventorship Guidance, 89 FR 10043 (Feb. 2024; rescinded and replaced by revised guidance, 90 FR 54636, Nov. 2025); Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022); Enfish v. Microsoft, 822 F.3d 1327 (Fed. Cir. 2016)

The USPTO's AI Inventorship Guidance and What Changed in November 2025

In February 2024, the U.S. Patent Office issued inventorship guidance for AI-assisted inventions (89 FR 10043). It confirmed that AI-assisted inventions are not categorically unpatentable, but it asked whether a human made a "significant contribution" to each claim, using the Pannu joint-inventorship factors. On November 28, 2025, the USPTO rescinded that guidance in its entirety and replaced it with revised guidance (90 FR 54636). For founders who have been asking can you patent AI when tools like large language models are involved in development, the current answer is still yes: the same inventorship standard applies to every invention, whether or not AI was used, and only natural persons can be named as inventors.

"Conception" is the touchstone of inventorship under the revised guidance. In patent law it means the complete and definite mental idea of the invention as it will be claimed, a specific, settled idea of a particular solution to the problem at hand, not just the idea that AI technology could help solve a problem. The revised guidance treats AI systems as instruments used by human inventors, comparable to laboratory equipment, computer software, and research databases. Using AI as a tool does not eliminate inventorship; the analysis turns on what the human actually conceived, decided, and directed. The USPTO presumes that the inventors named in the application are the actual inventors, and an application that names an AI system as an inventor will be rejected.

Document every human decision made during development, including what problem you identified, what approach you chose, what you directed the AI system to do, and how human ingenuity shaped the final claimed invention, because that documentation becomes your inventorship record if the application is challenged. For a full breakdown of what the USPTO now requires, see AI Patent Requirements Explained: What Are the Conditions of Patentability?

Conception, Significant Contribution, and How Patent Examiners Apply Them

Under the revised guidance, the question depends on who was involved. When a single person used AI, the test is conception alone: did that natural person form the complete and operative invention? The "significant contribution" test from Pannu now applies only when deciding whether multiple natural persons qualify as joint inventors; it is not used to weigh a human against an AI tool, because an AI system cannot be a joint inventor. Either way, a person who merely operates an AI system, accepts its outputs without creative direction, or prompts a general-purpose AI and files whatever it produces has not conceived the invention. A natural person who conceives the problem, defines the parameters, interprets outputs, and applies the results to a specific technical application has.

The practical evidence looks the same under either test. Did the person formulate the problem the AI was used to solve? Did they design the AI's parameters or training approach? Did they recognize and refine the AI's output into the claimed invention? A SaaS founder who specifies that their fraud detection model must flag real-time payment anomalies below a defined latency threshold, selects the training architecture, and validates the methodology against a defined benchmark has conceived the claimed system. Someone who prompts a general AI and files its response has not.

Build an inventor log that records each team member's specific contribution to the conception of each technical feature. This is the evidence that supports inventorship when an examiner challenges it.

How the Federal Circuit Has Shaped the Human Inventor Requirement

In Thaler v. Vidal, 43 F.4th 1207 (Fed. Cir. 2022), the Federal Circuit ruled that under 35 U.S.C. § 100(f) of the Patent Act, an "inventor" must be a natural person, specifically a human being. The case arose from inventor Stephen Thaler's attempt to list his AI system, DABUS, as the sole inventor on patent applications. The court rejected this, holding that the plain text of the Patent Act limits inventorship to natural persons, specifically human beings who can be identified as the source of the claimed invention. The Supreme Court declined to hear the patent case appeal in 2023, letting the Federal Circuit's ruling stand as controlling authority. Separately, in March 2026, the Supreme Court also declined to hear a related copyright case (Thaler v. Perlmutter, Case No. 25-449) involving the same DABUS AI system, reinforcing that AI-generated works require human authorship for IP protection across both patent and copyright law.

The practical implication is not that AI-generated innovations are worthless. It is that human inventors who directed the AI system must be identifiable, documentable natural persons, not AI systems or legal entities, who made significant contributions to each claimed element. Courts have not yet ruled on every edge case, and additional Federal Circuit decisions are expected as AI capabilities and the ways engineers use them continue to evolve.

Can AI Be Named as an Inventor in Other Countries?

No major patent office currently accepts an AI system as an inventor. Stephen Thaler's parallel DABUS applications failed in the major jurisdictions. The European Patent Office's Legal Board of Appeal held in case J 8/20 (decided December 2021) that the inventor named on a European application must be a natural person. The UK Supreme Court reached the same conclusion in Thaler v Comptroller-General of Patents [2023] UKSC 49, and Australia's Full Federal Court reversed an earlier ruling in DABUS's favor in 2022. The one outlier is South Africa, which granted a DABUS patent in 2021 under a system that checks formalities without substantive examination. If you plan PCT filings, plan on naming human inventors in every jurisdiction.

Can You Patent an AI Algorithm and What Makes an Algorithm Patentable

The widespread belief that algorithms cannot be patented is wrong. When founders ask can you patent AI algorithms specifically, the correct legal rule is narrower: an algorithm expressed only as a mathematical concept or abstract idea is not patent-eligible, but an algorithm implemented in a specific technical context to produce a specific technical result can be. Understanding this distinction is essential before pursuing patenting AI algorithms.

Why Algorithms Are Not Automatically Excluded From Patent Protection

The Alice framework does not categorically exclude algorithms from patent protection. It requires that an algorithm claim more than abstract ideas; it must have a concrete, real-world technical application. A machine learning algorithm that detects manufacturing defects in real time using computer vision is not just one of the excluded abstract ideas: it is a specific process, a practical application of AI technology, applied to a specific technical problem in a defined industrial context.

The Federal Circuit's decision in Enfish, LLC v. Microsoft Corp., 822 F.3d 1327 (Fed. Cir. 2016), opened a pathway specifically for software and AI patent applicants: if the claims are directed to an improvement in the functioning of the computer itself, not merely an abstract result achieved using a computer, they can survive § 101 scrutiny. This "improvement to the computer itself" standard is the most reliable route for AI algorithm patents that address technical performance problems such as processing speed, memory efficiency, or inference accuracy. Draft algorithm claims around the specific technical improvement the algorithm produces, not around the mathematical steps it executes.

For a comprehensive guide on what separates patentable algorithms from excluded abstract ideas, see Can You Patent an Algorithm? Essential Facts You Need to Know.

The Alice Two-Step Test Applied to AI and Machine Learning Inventions

Patent examiners at the United States Patent Office apply the Alice two-step test to every AI patent application, and understanding this test is central to answering can you patent AI in any specific technical context. Step one: is the claimed invention directed to abstract ideas, a law of nature, or a natural phenomenon? Machine learning training processes and neural network architectures are frequently flagged at this step. Step two: if directed to abstract ideas, does the claim include an inventive concept that transforms it into something significantly more than the abstract idea itself?

The USPTO's January 2019 Revised Guidance on § 101 refined how examiners apply the Alice framework, clarifying that mathematical concepts, mental processes, and certain methods of organizing human activity are the primary abstract idea categories. The Federal Circuit's decision in Berkheimer v. HP Inc., 881 F.3d 1360 (Fed. Cir. 2018), further established that the "inventive concept" analysis at step two is partly a factual question. Meaning a well-supported specification can help defeat an abstract idea rejection by establishing that the specific implementation was not well-understood, routine, or conventional at the time of filing, and that the inventive process reflected genuine human ingenuity.

That guidance had a measurable effect: one year after it issued, the likelihood of a first-action § 101 rejection in Alice-affected technologies had fallen 25%, according to the same USPTO report. The USPTO's July 2024 subject matter eligibility update, which added AI-specific examples 47 to 49, remains in effect. In Ex parte Desjardins, decided in September 2025 and designated precedential, the USPTO's Appeals Review Panel found that a method of training a machine learning model to reduce "catastrophic forgetting" was a technological improvement, not an abstract idea. The Office wrote Desjardins into MPEP § 2106 in December 2025 and now invites applicants to file declarations under 37 C.F.R. § 1.132 with evidence of the technical improvement, with updated memos in April and September 2026. Courts are not bound by USPTO guidance, though. In Recentive Analytics, Inc. v. Fox Corp. (Fed. Cir. 2025), the Federal Circuit held that claims applying generic machine learning to a new data environment, without improving the models themselves, are not patent-eligible.

Before filing, map every claim to the Alice two-step. If you cannot articulate the inventive concept beyond "AI does the task," the claim needs redrafting.

Deep Learning, Neural Networks, and the Patentability of AI Architecture

Specific AI architectures, novel arrangements of artificial neural network layers and components, custom training methodologies, and specialized deep learning frameworks, can be patentable when they represent a genuine technical improvement over the prior art. Patenting the architecture is legally distinct from patenting the outputs the architecture produces.

A novel artificial neural network topology that achieves faster inference at lower computational cost addresses a technical problem in the computer system itself. A neural network that happens to generate creative content addresses no such technical problem and struggles under Alice. The distinction is whether the claimed arrangement improves the performance of the technical system or simply uses a technical system to perform a task. Given how rapidly the AI field moves, prior art searches in this space are especially time-sensitive, as what was novel in 2022 may already have been published, preprinted, or patented by a competitor. If your team developed a custom AI architecture to solve a specific technical problem, treat that architecture as a separate patentable invention from the application it powers.

The same logic applies to no-code and low-code AI automation layers: generic automation usually draws a § 101 rejection unless the claims recite the concrete technical process the layer enables.

Can You Patent an AI Model and How to Protect What Your Model Actually Does

The question of can you patent AI when the invention is a trained model requires separating the components of that model, because different components attract different forms of intellectual property protection.

What Parts of an AI Model Can Be Covered by a Patent

A trained AI model has multiple protectable components, but not all of them are patentable. The trained model weights themselves are data, not a process or machine in the patent law sense, and are not patentable. What is patentable is the training methodology that produced those weights, if it is novel and non-obvious, claimed as a process. The reduction to practice of a training methodology, the concrete implementation that demonstrates how the method actually works in a practical application, is central to establishing that the claimed process is more than an abstract idea.

The inference pipeline, the system that takes inputs and produces model outputs, is patentable as a machine or system claim if it is configured in a non-obvious way that produces a defined technical result. The application-layer system in which the model is deployed is often the strongest patent target: a specific AI-powered system for medical imaging analysis, real-time language translation with defined latency constraints, or anomaly detection in industrial sensor data is a concrete technical system with identifiable prior art boundaries. Filing separate claims targeting the training methodology (process claims) and the deployed system (system claims) protects different facets of your AI model and together creates a stronger IP position.

In practice, patentable targets in AI software usually fall into one or more of these groups:

1. Novel pre-processing of model training data. This may include generating unique datasets prior to training an AI model.
2. Novel model training processes. This may include which algorithms are used in a process or improvements to a machine learning model or neural network algorithm.
3. Novel use of trained models. This may include implementing AI to control specific machines or to provide unique results from known products or methods.
4. Improvement to the underlying hardware implementing a trained AI model or algorithm or the production of a tangible result or object.

Why Trade Secrets May Protect AI Models Better Than Patents in Some Scenarios

When founders weigh can you patent AI models versus keeping them secret, the core tension is that patent protection requires public disclosure: when a patent issues, the full technical details of the invention are published and competitors can read exactly how the system works. For AI models with proprietary training datasets or fine-tuning methodologies, this disclosure requirement can hand competitors a roadmap.

Timing matters too. U.S. patent applications publish 18 months after the earliest filing date by default, often before any patent issues, so competitors can study your disclosure and design around your claims. Trade secret law has its own limit: it offers no protection against independent discovery, so if another team develops the same technique on its own, you cannot stop them.

Trade secret protection under the Defend Trade Secrets Act (DTSA) protects model weights, training data, and methodology indefinitely, provided the company takes reasonable steps to maintain their confidentiality. Patent protection lasts 20 years from the filing date; trade secret protection is indefinite, contingent on secrecy. Many leading AI companies combine both approaches: patent the application-layer systems and novel architectures, while protecting trained models and datasets as trade secrets, a strategy that recognizes how ai technologies evolve faster than traditional IP cycles. The right answer depends on the competitive landscape, the company's ability to maintain secrecy operationally, and whether the technology will be detectable in deployed products. If competitors can reverse-engineer it, trade secret protection is fragile.

Reasonable steps means NDAs, access controls, and documented confidentiality procedures before you share any model architecture or data.

Before filing a patent application on your AI model, assess whether the technical disclosure required by patent law outweighs the competitive advantage of keeping the methodology confidential.

What Copyright and Other IP Protection Cover That Patents Do Not

A full intellectual property strategy for an AI product layers multiple forms of protection. Copyright protects the source code, including any computer program that implements the AI model, automatically without registration — though registration is required to sue for infringement in U.S. courts. Copyright does not protect the trained weights, the model's outputs, or the underlying algorithm. The U.S. Copyright Office's 2023 guidance on AI-generated content clarified that outputs generated autonomously by AI without human creative authorship are not eligible for copyright protection, though code written by human inventors that implements AI systems retains full copyright coverage.

Trademark protects the brand name and product identity of the AI system, administered through the trademark office alongside patent prosecution. Together, the optimal IP protection stack for most AI companies looks like this: patent the novel technical process and system, copyright-register the source code, protect confidential methodology and training data as trade secrets, and trademark the product name. Each layer protects something the others do not.

How to Structure a Winning AI Patent Application Before You File

Understanding what is theoretically patentable is only useful if the patent application is built correctly. Founders who have confirmed that can you patent AI resolves to yes for their invention still face a second challenge: many AI patent rejections are not caused by unpatentable technology. They are caused by poorly drafted claims and underspecified specifications.

How to Write AI Patent Claims That Survive Examiner Scrutiny

The patent claims are the legal boundary of what the patent protects, and poorly drafted claims are the most common reason AI patent applications receive § 101 rejections. Three principles govern strong AI claim drafting. First, specificity: name the specific technical components rather than generic placeholders like "a processor" and "a neural network." Second, functional connection: show how the components interact to produce the specific technical result. Third, concrete outcome: articulate the measurable improvement in the technical process or computer system.

Avoid claiming the abstract goal; claim the specific implementation. Every independent claim in an AI patent application should pass a simple internal test: remove the AI element. Does anything technical remain that a court could evaluate? If yes, the claim is likely well-anchored in a specific technical application. If no, redraft to anchor it more concretely.

Beyond Enfish, the Federal Circuit's decision in McRO, Inc. v. Bandai Namco Games America Inc., 837 F.3d 1299 (Fed. Cir. 2016), shows what survives: an automated lip-sync animation method was eligible because the claims used specific, bounded rules and did not preempt every way of solving the problem. Convert functional AI descriptions into structural claims that describe how the system reaches its technical result, and avoid purely functional language that states a result without the means. Founders who want a working template can start with our software patents checklist.

Understanding how to construct strong claims is the foundation of effective patenting your AI wrapper and any application-layer AI innovation.

Building the Patent Specification to Support Broad AI Claims

The patent specification must support every element of every claim. For AI inventions, this means describing the specific training data characteristics, the model architecture, the inference process, and the technical problem being solved with enough detail that a person skilled in the relevant field could replicate the invention. Under 35 U.S.C. § 112, the written description and enablement requirements are not formalities. A vague specification that describes AI technology as a general capability without explaining the specific implementation creates a written description problem that can invalidate claims even after a patent issues.

The specification is also where human inventors document the inventive concept: what makes this technical approach different from prior art methods, and why the reduction to practice of this specific approach was non-obvious at the time of filing. Courts, including in Berkheimer v. HP Inc., have confirmed that evidence in the specification about what was unconventional at the time of filing can defeat a § 101 challenge. Draft the specification before finalizing claims. The depth of the technical description should determine how broad the claims can legitimately be, not the other way around.

Provisional Patent Applications as the First Step for AI Inventors

A provisional patent application establishes a priority date without the full cost or formality of a non-provisional application. For AI founders, this matters urgently: the AI patent landscape is moving fast, prior art accumulates rapidly, and a provisional filing locks in the filing date while giving the inventor 12 months, under 35 U.S.C. § 111(b), to refine claims and conduct a freedom-to-operate analysis.

A provisional application must still disclose the invention with enough technical detail to support the later non-provisional, including a clear description of the reduction to practice. How the invention actually works in a concrete implementation that demonstrates a practical application of the claimed technology. A one-paragraph summary is insufficient. A well-drafted provisional can serve as the basis for multiple related applications and continuation applications as the AI product evolves. File a detailed provisional patent application before publicly launching or demoing your AI product. Public disclosure before filing can permanently bar patent protection under U.S. law.

The SaaS Patent Guide 2.0 covers provisional patent strategy in depth for founders building AI-powered software products.

A Seven-Question Pre-Filing Checklist for AI Inventors

Run through these seven questions with a registered patent attorney before drafting a single claim. Each one maps to a failure mode that sinks AI patent applications.

(1) Can every named inventor articulate a specific, significant contribution to conceiving the claimed invention, independent of what the AI system produced? A "no" here means an inventorship defect that can invalidate the patent later.

(2) Have you documented the R&D process in dated engineering logs, commit histories, or lab notebooks that show human decision-making throughout? These records can be critical evidence if inventorship is challenged.

(3) If an AI tool was used, have you identified which outputs the human inventors adapted, applied, or modified versus adopted wholesale? Under the USPTO's revised inventorship guidance, conception is the test, and a person who simply pressed "run" did not conceive anything. If you cannot answer yes to all three, pause the filing and complete a formal inventorship review.

(4) Do your core claims describe a specific technical improvement to a computer system or a concrete application of an AI method, rather than the generic use of AI to perform a function? If the claim reads as an abstract idea, it will draw a §101 rejection.

(5) Has a professional prior art search been run against not only existing patents but also academic literature, GitHub, and preprint servers like arXiv? AI research is commonly published in open venues before any patent is filed, and AI-related projects on GitHub reached roughly 1.8 million by 2023. A patent-only prior art search is insufficient for AI inventions.

(6) Have you mapped each component of the AI invention, algorithm, trained model, dataset, and application layer, to the right protection vehicle, patent, trade secret, or copyright? Different components deserve different protection, and a computer program layer may qualify for copyright protection in addition to other vehicles.

(7) Have you evaluated whether trade secret protection fits any component better than patent protection, given the disclosure requirements and the nature of your competitive advantage? If the novelty lives in your training data or model weights, disclosing them in a patent may cost more than it protects. Answering these two questions early prevents the most expensive strategic mistakes in AI IP planning.

What the USPTO Examination Process Looks Like for AI Patent Applications

The patenting process for an AI invention runs in four stages: identify the patentable innovation, search the prior art, prepare and file the application, and prosecute it through examination. The sections above cover the first three. Total pendency for U.S. applications often runs two years or more (the USPTO Patents Dashboard shows current figures), so an application built to withstand examination saves real time.

How Patent Examiners Evaluate AI Patent Applications

Most AI applications are examined in Technology Center 2100 (computer architecture and software), though AI applied in a field such as biotechnology may go to that field's technology center. Examiners evaluate each application for § 101 eligibility, § 102 novelty, § 103 obviousness, and § 112 written description and enablement.

AI patent applications face heavier prior art scrutiny than most fields because AI technologies and machine learning techniques are extensively documented in academic literature, which qualifies as prior art. The scale is enormous. According to Stanford's 2024 AI Index Report, spanning computer science and adjacent disciplines, annual AI publications nearly tripled from about 88,000 in 2010 to more than 240,000 by 2022. Examiners routinely cite conference papers, arXiv preprints, and open-source repositories in AI cases. Commission a prior art search before drafting your patent application to map the landscape your examiner will be searching.

How to Respond When Your AI Patent Application Is Rejected

A rejection is not the end of an AI patent application. Examiners issue a rejection when they find a problem with the merits of a claim, such as eligibility, novelty, or obviousness, and an objection when something is formally wrong, such as a typographical error. Objections can usually be fixed by amendment. Rejections are answered with amendments, legal arguments, or both, and once claims have been rejected twice you can appeal to the Patent Trial and Appeal Board (PTAB).

It is not unusual for an Office Action on an AI application to combine several grounds, for example § 101 for an abstract idea, § 103 for obviousness over prior art, and § 112 for not describing how the AI system works. Respond strategically rather than narrowing claims to escape every rejection: every amendment and argument becomes part of the prosecution history that courts read when interpreting claim scope. For § 101 rejections, an eligibility declaration is now an option alongside argument and amendment. The surest way to avoid these holdups is drafting before filing: independent claims at broad, medium, and narrow scope, plus a specification that explains what a probabilistic AI system does in terms a skilled engineer could replicate.

The Inventorship Question That Could Invalidate Your AI Patent After It Issues

Inventorship is the most dangerous single issue in AI patent prosecution, and it is the one most founders treat as a formality, even after the question of can you patent AI has been answered in their favor.

Why Inventorship Errors Are the Most Dangerous Mistake in AI Patent Prosecution

Inventorship is a legal requirement: if human inventors who made a significant contribution are omitted, or if someone is listed who did not contribute significantly to the conception of the claims, the patent can be challenged and invalidated. This is not a technical error that can be quietly corrected. It is a legal defect that opposing counsel will raise in inter partes review (IPR) or litigation. For AI-assisted inventions, the challenge is sharper: if the AI system made decisions that are reflected in the patent claims, listing human operators as inventors without proper documentation invites the argument that the named natural persons never actually conceived the claimed subject matter.

Treat inventorship determination as a legal analysis conducted before the application is filed — not a checkbox completed at filing time. For a detailed review of how rights flow from properly documented inventorship, see Rights of Patent in the Modern Era: Blocking Competitors, Funding Growth, Monetizing IP.

How Joint Inventorship Works When Multiple Humans and AI Systems Collaborate

Under 35 U.S.C. § 116 of the Patent Act, joint inventors do not need to have worked together simultaneously or made equal contributions, but each named inventor must have made a significant contribution to the conception of at least one claim. The Federal Circuit's standard from Pannu v. Iolab Corp., 155 F.3d 1344 (Fed. Cir. 1998), requires that each joint inventor: (1) contribute to the conception of the invention, (2) make a contribution that is not insignificant in quality relative to the full invention, and (3) do more than merely explain well-known concepts or the current state of the art to the real inventors. The USPTO's revised guidance confirms that these factors apply among the human contributors even when AI tools were used.

For AI startups with distributed development teams, one person trained the model, another designed the application layer, a third integrated the inference pipeline, the right approach is to map each team member's contribution to specific claims in the application before filing. AI systems cannot be listed as joint inventors under current United States patent law, and each human contribution must be distinguished from AI-generated outputs in the documentation record. Document each contribution at the time it occurs. Retroactive documentation is harder to defend and easier to challenge.

What Happens When You Use Third-Party AI Tools in Your Inventive Process

Many founders build AI innovations on top of third-party foundation models: large language models, generative AI platforms, or open-source frameworks. The human inventors at those foundation model companies are not inventors of what you build with their tools, just as a database vendor is not an inventor of the application you build using their software. These distinctions are especially important for related inventions that share underlying AI components across multiple patent applications. What matters is whether the founder conceived the claimed invention, including problem definition, architectural decisions, training approach, and validation criteria. Each human contribution must be documented separately from what the AI tool produced.

However, using third-party AI tools raises a separate question from inventorship: freedom to operate. If the foundation model or its outputs are covered by existing patents held by the tool's developers, the system you build on top of those outputs may infringe those patents regardless of who invented it. Inventorship may be clear, but the freedom to commercialize the invention requires a separate legal analysis.

The AI Patent Mastery resource covers third-party AI tool risks and freedom-to-operate strategy in depth for AI founders navigating this issue.

Building an AI Patent Portfolio Strategy That Grows With Your Company

A single patent is a legal document. A patent portfolio is a competitive strategy. For AI founders, the difference matters enormously at the Series A, in acquisition due diligence, and when a well-funded competitor enters your market.

An issued U.S. patent gives its owner the right to exclude others from making, using, selling, offering to sell, or importing the claimed invention in the United States, generally for 20 years from the filing date. That is what lets you stop unauthorized use of your AI invention, license it for revenue, and show investors a defensible position. "Fast followers" can carve out market position too, with improvement patents on what they build next.

Leading Agentic AI Patent Filers in the U.S. (Feb 2024 to Apr 2025): Google led, followed by Nvidia, IBM, and Intel. Source: Axios, May 2025

Why a Single Patent Is Rarely Enough to Protect an AI Innovation

A single patent protects the specific claims as filed, and u.s. patents in particular are limited to domestic enforcement. As an AI product evolves, with new features, new use cases, and updated architecture, the original patent may not cover the current product at all. The right approach is to treat the first patent application as the foundation of a portfolio. Continuation applications capture new claim directions from the same original disclosure without requiring a new filing date. Divisional applications are filed when the U.S. Patent Office determines that a single application covers more than one distinct invention. A rolling provisional filing practice lets the team lock in priority dates for new technical developments as they occur.

A well-managed patent portfolio creates meaningful barriers to entry and strengthens the company's negotiating position with investors, acquirers, and potential licensees. Continuation practice is how leading AI companies stay ahead of competitors without starting from scratch with each product update. For a structured approach to building a portfolio that scales, see Planning A Rock Solid Patent Portfolio Strategy.

When to File International AI Patent Protection Through PCT Applications

A U.S. patent protects the invention only within the United States. If your AI product will be commercialized in Europe, Asia, or other markets, international protection requires separate filings. The Patent Cooperation Treaty (PCT) provides a single application mechanism that preserves the right to pursue patents in more than 150 countries, with 12 months from the priority date for initial PCT filing and 30 months for national phase entry. This timeline allows AI companies to assess market strategy before committing to the cost of individual national filings.

Key international jurisdictions for AI companies include the European Patent Office (EPO), which requires that claimed software and AI inventions have "technical character," a standard that shares ground with the USPTO's § 101 framework but is applied differently, and China, which has rapidly expanded its AI patent examination infrastructure to keep pace with surging AI technologies. WIPO confirmed that between 2014 and 2023, more than 38,000 generative AI inventions originated from China, six times more than from the second-place United States, and its 2026 update counted over 43,000 more patent families from China-based inventors published in 2024 and 2025 alone. The strategic filing destinations depend on where the AI product will be deployed commercially, where competitors are active, and where the most valuable licensing opportunities exist. Missing the PCT deadline eliminates international patent rights permanently.

How a Strong Patent Portfolio Supports VC Fundraising and Acquisition Conversations

Investors conduct IP due diligence, and an AI startup that cannot demonstrate defensible patent protection, or that has visible inventorship vulnerabilities, faces valuation discounts and deal delays. Patent applications signal that the technical innovation has been formally claimed and that the team has taken concrete legal steps to protect the competitive advantage. Issued patents go further: they are assignable assets, licensable revenue streams, and demonstrable moats in acquisition conversations.

Start building your portfolio before your Series A. Investors want to see that the IP is owned, documented, and defensible, not a plan to file something eventually. For real-world examples of how patent portfolios create enterprise value in AI-driven industries, see AI Patents for FinTech Explained: What Bank of America, FICO, and Lucinity Got Right.

How AI Is Changing the Patent Process Itself

AI is not only the subject of patents. It is increasingly a tool inside the patent process, on both sides of the examiner's desk.

Can AI Write a Patent Application?

AI can draft parts of one, but it cannot supply the judgment that decides whether the patent holds up. Generative AI produces first-draft claims and specification language quickly, and AI search tools scan patents and papers far faster than manual review. The USPTO accepts applications prepared with AI assistance, and there is no general duty to disclose that you used an AI tool unless the Office asks. Its April 2024 guidance on AI tools reminds practitioners that existing rules "apply regardless of how a submission is generated," including signature certifications and the duty of candor. A named human signer, not the tool, answers for every statement.

The risk in AI drafting is quality: claims that are too broad invite rejections, claims that are too narrow give away scope, and tools can invent authority: in Mata v. Avianca, Inc. (S.D.N.Y. 2023), a federal court sanctioned lawyers who filed a brief citing cases ChatGPT had made up. Treat every AI output as a starting draft that a registered patent attorney reviews line by line, and keep confidential invention details out of tools that may retain or train on your inputs.

Can AI-assisted drafting improve your odds of a grant? Only indirectly. What moves outcomes is covered above: claims tied to a specific technical improvement, a specification that supports them, and a prior art search that covers papers and code. AI helps a practitioner get there faster, and the learning curve is less about operating the software than about knowing where its output fails.

The USPTO uses AI on its side too. Its Automated Search Pilot, launched in October 2025 and extended into 2026, gave participating applicants an AI-generated list of up to ten prior art references before examination began, so they could amend their claims early.

What Generative AI Means for Patent Portfolio Management

Cheaper drafting means more AI filings and more crowded prior art. Stanford's AI Index reports that AI patent grants worldwide rose 62.7% from 2021 to 2022 alone. For founders managing a portfolio, that means running freedom-to-operate and patent landscape analyses earlier in the product cycle, before committing significant R&D budget to a feature. AI landscape tools make white space easier to spot, and our analysis of patents by Apple shows what one looks like.

The Future of AI and Patent Law and Key Developments to Watch

The rules are still moving. Within about a year, the USPTO replaced its AI inventorship guidance (November 2025), wrote Desjardins into the MPEP (December 2025), and updated its eligibility declaration practice (April and September 2026). Congress has also considered the Patent Eligibility Restoration Act, which would rewrite the § 101 exclusions, and the Federal Circuit continues to decide AI eligibility cases such as Recentive. Revisit your AI patent strategy at least once a year.

Frequently Asked Questions

Can an AI system be patented?

An AI system can be patented when the patent application focuses on the specific technical problem the system solves and the specific implementation used to solve it. The USPTO does not grant patents on the abstract concept of artificial intelligence, but it does grant patents on specific AI-powered systems and processes that meet the novelty, non-obviousness, and subject matter eligibility requirements under 35 U.S.C. § 101. The critical requirement is that a human inventor must have conceived each claimed invention, as confirmed by the USPTO's revised inventorship guidance of November 2025 (90 FR 54636), which replaced the 2024 guidance published at 89 FR 10043. A registered patent attorney can assess whether your AI system's architecture and functional design meet these thresholds.

What is the difference between patenting an AI model versus patenting an AI application?

Patenting an AI model typically means claiming the training methodology, the model architecture, or the inference pipeline as a process or system. The technical components that satisfy patent eligibility requirements. Patenting an AI application means claiming the deployed system in which a model operates to solve a specific real-world problem, such as a fraud detection engine or a medical imaging platform. The strongest intellectual property positions layer both: model-level patents protect the underlying technology, while application-level patents protect the commercial product. Together, both types of claims in a coordinated patent portfolio create coverage that neither achieves alone. Understanding the reduction to practice requirements for each type of claim is essential to drafting applications that survive examination.

Can you patent an AI algorithm if it was generated using another AI tool?

Yes, if human inventors conceived the claimed algorithm, including defining the problem, selecting the approach, and directing how the AI tool was applied. The USPTO's revised inventorship guidance (November 2025) makes clear that using AI tools in the inventive process does not disqualify an invention from patent protection: AI tools are treated like other instruments, and the same conception standard applies, provided a human inventor can be identified as the natural person who conceived the claimed subject matter. What current U.S. law does not permit is listing an AI system as an inventor, or filing claims where no human decision-maker can be identified as having conceived the core technical approach. The distinction between human contributions and AI-generated outputs must be documented clearly before filing.

What is the most important thing to do before filing an AI patent application?

Two things are equally critical: commission a prior art search to establish that the AI invention is novel and non-obvious relative to existing publications and patents, and work with a registered patent attorney to draft claims that satisfy the Alice two-step test before the application is submitted. Filing without both steps significantly increases the probability of a § 101 rejection. Subject matter eligibility rejections are a frequent obstacle for AI applications, and they are far easier and cheaper to prevent before filing than to fix after.

Can AI be granted a patent or be named as an inventor?

No AI system can be named as an inventor. Under current U.S. law, only natural persons can be inventors: the Federal Circuit held in Thaler v. Vidal that the Patent Act's "individual" means a human being, and the USPTO's revised inventorship guidance of November 2025 confirms that an application naming an AI system as inventor will be rejected. You can still patent an AI-driven invention, as long as a human conceived the claimed invention and is named as the inventor.

What is the relationship between AI and patent law, and how is AI changing it?

AI and patent law intersect in two ways. First, AI technologies are the subject of patent applications: AI systems, machine learning methods, neural network architectures, and AI-powered applications. Second, AI tools are now used inside the patent process for prior art searching, drafting, and landscape analysis, by applicants and by the USPTO itself. The core legal questions are § 101 eligibility, inventorship, and a prior art base that keeps growing as AI publications surge. Courts and the USPTO are still working through these rules, so expect continued change.

Will AI take over patent lawyers?

No, though AI tools are changing how patent lawyers work. Generative AI can speed up drafting and prior art research, but navigating § 101 rejections, preserving claim scope, advising on inventorship, and building a portfolio take legal judgment a tool cannot supply, and a human practitioner remains responsible for every filing. Attorneys who use AI well are likely to outpace those who do not.

What is the 30% rule in AI, and does it apply to patents?

The "30% rule" is not a legal doctrine. It is an informal rule of thumb from business writing about splitting work between AI and people, usually that AI handles most of a workflow while humans keep about 30% for judgment, creativity, and quality control. Patent law uses no percentage test. Inventorship turns on conception: whether a natural person formed the complete and operative idea of what is claimed. Treating an informal adoption guideline as the legal standard can lead to inventorship mistakes.

How do I choose an AI patent attorney?

Look for a registered patent attorney, admitted to practice before the USPTO, who understands machine learning well enough to describe your architecture and training method in technical detail, has experience answering § 101 rejections in software and AI applications, and can advise on inventorship and trade secret strategy as well as drafting. Our guide to choosing an AI patent attorney covers what to look for.

Your Next Steps to AI Patent Protection Success

Protecting an AI invention requires more than a good idea and for any founder still asking can you patent AI, the answer is yes, but only if the claims are drafted to survive § 101 scrutiny, human inventors' contributions are documented and defensible, and a portfolio strategy grows alongside the product.

The bottom line: a weak AI patent application, or one filed without a prior art search and proper inventorship analysis, is not just a wasted filing fee. It is a liability that competitors can exploit. A strong AI patent application, built on a well-drafted specification with concrete reduction to practice details and claims anchored in specific technical improvements, creates a real competitive moat.

To protect your AI innovation before a competitor does:

• Schedule a Free IP Strategy Call with Andrew Rapacke to assess whether your AI invention qualifies for patent protection and what claims strategy makes sense
• Download the SaaS Patent Guide 2.0 for a step-by-step framework on building an AI patent portfolio as a software founder
• Access the AI Patent Mastery resource for advanced guidance on AI-specific patent prosecution strategy

Rapacke Law Group offers fixed-fee patent prosecution and the RLG Guarantee: a full refund for patentability issues on your provisional patent application. Founders building on AI technology deserve IP counsel that is aligned with their outcomes, not billing by the hour against their runway.

To Your Success,

Andrew Rapacke
Managing Partner, Registered Patent Attorney
Rapacke Law Group