Artificial Intelligence and Machine Learning Patents Under US Law

Artificial intelligence and machine learning inventions occupy one of the most contested doctrinal spaces in US patent law, where foundational eligibility doctrine collides with rapid technological development. The United States Patent and Trademark Office (USPTO) has issued specific guidance on how examiners evaluate AI-related claims under 35 U.S.C. § 101, yet the boundaries remain actively litigated. This page covers the definition and scope of AI/ML patents, the mechanics of examination, the classification boundaries between eligible and ineligible subject matter, and the tradeoffs that shape prosecution strategy.

• Definition and Scope
• Core Mechanics or Structure
• Causal Relationships or Drivers
• Classification Boundaries
• Tradeoffs and Tensions
• Common Misconceptions
• Checklist or Steps
• Reference Table or Matrix
• References

Definition and scope

AI and machine learning patents are utility patents, issued under Title 35 of the United States Code, that claim inventions involving the design, training, deployment, or application of computational systems capable of pattern recognition, inference, prediction, or autonomous decision-making. The subject matter spans a wide range: novel neural network architectures, training methodologies, inference optimization techniques, hardware accelerators designed for tensor operations, and domain-specific applications of AI to fields such as medical imaging, natural language processing, or autonomous navigation.

The USPTO administers AI patent examination through its standard utility patent framework, but the agency has issued dedicated guidance documents — notably the 2019 Revised Guidance on Patent Subject Matter Eligibility and the 2020 USPTO Artificial Intelligence and Emerging Technology Partnership materials — that address how AI-specific claim language should be evaluated. These documents do not create new law but direct examiner practice under Alice Corp. v. CLS Bank International (573 U.S. 208, 2014) and its progeny.

The scope of protectable AI inventions under US law is fundamentally governed by the patent eligibility requirements framework, which distinguishes between a concrete technical implementation — eligible — and an abstract mathematical relationship or mental process — ineligible without more. Understanding the regulatory context for patent law in which AI claims are evaluated is essential before drafting or prosecuting these applications.

Core mechanics or structure

AI patent examination at the USPTO follows the same formal procedural path as any utility patent: filing, examination, office action response, and allowance or appeal. However, the substantive analysis diverges at the § 101 eligibility stage, which is applied before novelty (§ 102) and non-obviousness (§ 103) are considered.

The Two-Step Alice/Mayo Framework

The controlling framework for evaluating AI patent claims derives from two Supreme Court decisions: Mayo Collaborative Services v. Prometheus Laboratories (566 U.S. 66, 2012) and Alice Corp. v. CLS Bank International (573 U.S. 208, 2014). The USPTO's 2019 Revised Guidance restructured the application of this framework into a three-prong, two-step analysis:

• Step 2A, Prong 1: Determine whether the claim recites a judicial exception — an abstract idea (including mathematical concepts, mental processes, or certain methods of organizing human activity), a law of nature, or a natural phenomenon. Most ML algorithms and mathematical optimization functions fall within "mathematical concepts."
• Step 2A, Prong 2: Determine whether the claim recites additional elements that integrate the judicial exception into a practical application. This is the critical threshold for AI claims. Examples include applying the algorithm to a specific technical field (e.g., improving the compression of medical image data), using the trained model to control a physical process, or achieving a measurable improvement in computer functionality.
• Step 2B: If the claim does not integrate the exception into a practical application, determine whether the additional elements amount to significantly more than the exception alone — such as a specific, unconventional hardware configuration or an inventive combination of steps.

Claim drafting for AI patents typically involves system claims (covering the hardware-software combination), method claims (covering the training or inference process), and computer-readable medium claims (covering the stored instructions). Independent claims must be written with sufficient specificity to survive Prong 2 analysis without being so narrow that they are easily designed around.

Causal relationships or drivers

The concentration of § 101 rejections in the AI field stems from the mathematical nature of machine learning itself. A neural network is, at its core, a parameterized function — a composition of linear transformations and nonlinear activation functions — and the USPTO's 2019 guidance explicitly lists "mathematical relationships" and "mathematical formulas or equations" as categories of abstract ideas. This creates a structural tension: the technical contribution of most ML inventions lies precisely in the mathematical innovation, yet claiming that innovation directly invites rejection.

Three primary factors drive the patent eligibility calculus for AI inventions:

1. Specificity of technical application. Claims tied to a specific technical domain — for example, a convolutional neural network architecture optimized for real-time anomaly detection in industrial sensor arrays — fare better than claims that generalize the algorithm across all possible uses.

2. Claimed improvement to computer functionality. The Federal Circuit's Enfish, LLC v. Microsoft Corp. (822 F.3d 1327, Fed. Cir. 2016) line of cases established that improvements to the operation of a computer itself — not merely using a computer to perform an abstract process faster — can survive § 101 scrutiny. AI claims that demonstrate a reduction in computational load, memory usage, or inference latency on defined hardware have leveraged this pathway.

3. Hardware integration. Claims that tie a trained model to specific processing architecture — application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or neuromorphic chips — introduce tangible, machine-specific elements that move the claim away from pure abstraction.

The America Invents Act (Pub. L. 112-29, 2011) shifted the US patent system to a first-inventor-to-file framework, which amplifies the strategic pressure to file AI patent applications early, before public disclosure of any underlying research.

Classification boundaries

AI patent claims are distributed across the USPTO's Cooperative Patent Classification (CPC) system, with the most concentrated activity in:

• G06N — Computing arrangements based on specific computational models, including machine learning (G06N 20/00), neural networks (G06N 3/00), and Bayesian approaches (G06N 7/00).
• G06F — General-purpose digital computing, covering AI-assisted data processing architectures.
• G16H — Healthcare informatics, for AI applied to medical diagnosis and treatment planning.
• B60W / G08G — Cooperative driving and traffic control, for autonomous vehicle AI systems.

The boundary between eligible and ineligible AI subject matter under § 101 is not determined by CPC classification but by claim language. The following structural distinctions define the doctrinal line:

The patentable subject matter doctrine governs where each AI claim type falls within this spectrum, and claims drafted without careful attention to practical application language routinely receive § 101 rejections in prosecution.

Tradeoffs and tensions

Breadth versus eligibility. Broad claims that capture the full scope of an algorithmic innovation are most likely to receive § 101 rejections, while narrow claims tied to specific applications preserve eligibility but limit the scope of exclusion. This creates a fundamental prosecution tradeoff: applicants who draft broadly protect more territory in theory but face a longer, more expensive prosecution.

Trade secret versus patent disclosure. AI training pipelines and model architectures may be more effectively protected as trade secrets under the Defend Trade Secrets Act (18 U.S.C. §§ 1831–1839) than through patent disclosure. Patent applications require full enablement under 35 U.S.C. § 112, which for AI inventions may require disclosing training datasets, hyperparameter settings, or architectural details that competitors can exploit even after the patent expires. The comparative analysis of trade secret vs. patent protection is particularly consequential for foundation model developers.

Inventorship and AI-generated contributions. The Federal Circuit's decision in Thaler v. Vidal (43 F.4th 1207, Fed. Cir. 2022) held that AI systems cannot be listed as inventors under 35 U.S.C. § 100(f), which defines an inventor as an "individual." Human inventorship attribution for AI-assisted invention remains an active doctrinal problem where the AI's output was the material creative contribution.

International harmonization. The European Patent Office (EPO) and the USPTO apply divergent standards: the EPO's "technical character" requirement under Article 52 of the European Patent Convention produces different eligibility outcomes than the Alice/Mayo framework, meaning a claim allowable in Europe may face rejection in the US, or vice versa.

Common misconceptions

Misconception: Describing an AI system on a computer automatically satisfies § 101.
The USPTO's 2019 Revised Guidance explicitly rejects the view that reciting a generic processor or "a computer" transforms an otherwise abstract AI claim into eligible subject matter. Step 2B requires that any additional elements amount to significantly more than the judicial exception — generic computer implementation alone does not meet this threshold.

Misconception: A published research paper on a neural network architecture cannot be patented.
Academic publication triggers a 1-year grace period under 35 U.S.C. § 102(b)(1)(A) for the inventor's own disclosures under the first-inventor-to-file system, but bars third-party filers immediately. The claim's eligibility under § 101 is independent of novelty — an architecture can be novel yet still fail eligibility if not integrated into a practical application.

Misconception: All machine learning patents are software patents.
AI patent claims that cover purpose-built hardware — neuromorphic processors, AI-specific memory architectures, or dedicated inference chips — are hardware patents, not software patents. These claims follow a different eligibility analysis that focuses on structural novelty rather than abstract idea doctrine.

Misconception: The USPTO does not grant AI patents.
The USPTO granted more than 60,000 AI-related patents in fiscal year 2020 alone, according to the agency's own reporting. The doctrinal complexity of § 101 does not preclude allowance — it shapes how claims must be drafted to achieve it.

Checklist or steps

The following sequence describes the structural stages of an AI patent application's lifecycle at the USPTO, framed as a reference for understanding the process:

1. Prior art search — Identify existing patents, published applications, and non-patent literature in the relevant CPC subclasses (G06N, G06F, and domain-specific classifications). The patent prior art search process establishes the field's state of the art.
2. Claim architecture design — Draft independent claims at the system, method, and computer-readable medium levels. Confirm that at least one independent claim integrates the AI innovation into a specific practical application rather than claiming the mathematical relationship in the abstract.
3. Specification drafting — Prepare a written description that enables a person of ordinary skill in the art to make and use the full scope of the claimed invention, including training procedures, datasets used, and architectural parameters, per 35 U.S.C. § 112(a).
4. Filing — File a non-provisional application or a provisional application to establish a priority date. Review provisional patent application mechanics if early priority date is the primary objective.
5. Examination — Respond to examiner office actions addressing § 101, § 102, and § 103 rejections. For § 101, provide a reasoned response under the 2019 Revised Guidance demonstrating practical application integration.
6. Interview — Request an examiner interview to clarify claim scope and identify acceptable claim language before final rejection issues.
7. Appeal or allowance — If claims are finally rejected, pursue appeal to the Patent Trial and Appeal Board (PTAB) under 37 C.F.R. § 41.31 or consider claim amendment to narrow scope.
8. Post-grant considerations — After grant, monitor for inter partes review petitions at PTAB, which can challenge issued claims on § 102 and § 103 grounds. Review inter partes review procedures for defensive posture planning.

Reference table or matrix

AI Patent Claim Type — Eligibility and Prosecution Characteristics

The software patent law framework provides the closest doctrinal analog to AI patent prosecution, and case law developed in that context — including Enfish, McRO, Inc. v. Bandai Namco Games America (837 F.3d 1299, Fed. Cir. 2016), and Core Wireless Licensing S.A.R.L. v. LG Electronics, Inc. (880 F.3d 1356, Fed. Cir. 2018) — remains directly applicable to AI claim eligibility analysis. A comprehensive overview of the US patent system's structure and scope is available on the site index, which maps the full range of patent law topics covered across this resource.