Patenting AI and Machine Learning Inventions

Artificial intelligence and machine learning inventions are patentable in every major jurisdiction — but the rules for how to claim them differ more between countries than for any other technology type. An AI patent claim that sails through the JPO may be rejected at the USPTO under Alice. A claim that the EPO allows for its "technical effect" may be refused at the Indian Patent Office under Section 3(k). Navigating these differences is the central challenge of AI patent strategy.

This guide covers what is claimable, how to draft claims that survive examination across jurisdictions, and the specific AI patentability rules at each major office.

What Is Patentable in AI/ML

Patentable

AI-implemented inventions that achieve a specific technical effect. An ML model that improves the accuracy of a medical imaging system, optimises the energy efficiency of an HVAC system, reduces defects in a manufacturing process, or enhances the signal-to-noise ratio of a sensor — these tie the algorithm to a measurable physical or technical outcome and are patentable in most jurisdictions.

Novel training methods. A new technique for training neural networks — a specific loss function, a novel regularisation method, a particular data augmentation strategy that produces measurably better results for a defined task — can be patentable when tied to a specific application and demonstrated effect.

AI hardware. Novel chip architectures, accelerators, memory configurations, or processing pipelines designed specifically for AI workloads. These are mechanical/electronic inventions and face fewer eligibility issues.

Data processing pipelines. Specific sequences of data preprocessing, feature extraction, model inference, and post-processing that together produce a technical result. The pipeline — as a system — is the invention, not any individual algorithmic step.

Not Patentable (in most jurisdictions)

Pure mathematical models. A neural network architecture described only in mathematical terms, without a specific application, hardware context, or measurable technical effect.

Abstract algorithms. An optimisation algorithm with no defined input data, no specified hardware, and no measurable output improvement.

Business methods implemented by AI. "Using machine learning to recommend products to customers" is a business method — not patentable at the EPO, extremely difficult at the USPTO post-Alice, and excluded in India, China, and most other jurisdictions.

Jurisdiction Comparison

Claim Drafting Strategy for AI Inventions

The Universal Rule: Tie the Algorithm to a Technical Application

Across all jurisdictions, the single most effective claim drafting strategy for AI inventions is to anchor the algorithm in a specific technical application with a measurable outcome. The claim should recite:

1. Specific input data — not "data" generically, but "sensor measurements from an array of temperature sensors positioned at defined locations in a turbine engine"
2. Specific processing — not "applying machine learning" but "processing the sensor measurements using a trained convolutional neural network having [specific architecture features] to generate a prediction of remaining bearing life"
3. Specific output/action — not "producing an output" but "generating a maintenance alert when the predicted remaining bearing life falls below a threshold, the alert transmitted to a maintenance scheduling system"
4. Specific technical improvement — "wherein the prediction accuracy exceeds 95% for remaining life predictions within a 500-hour window, compared to 78% for vibration-analysis-only methods"

This structure satisfies Alice Step 2 (US), the EPO's technical effect requirement, CNIPA's technical solution test, and the JPO's "creation of technical ideas" standard simultaneously. It also provides concrete, measurable evidence for non-obviousness arguments.

US-Specific: Surviving Alice

The most common AI patent rejection in the US is § 101 under Alice. To survive:

Frame the claim around the technical improvement to a computer or technical system — not around the business outcome. "Improving bearing failure prediction" is better than "reducing maintenance costs." Include hardware elements in the independent claim — specific sensors, processors, memory, communication interfaces. Cite specific performance metrics in the specification that demonstrate a concrete technical improvement. Consider filing both system claims (strongest for § 101) and method claims (broader coverage if they survive).

EPO-Specific: The COMVIK Approach

The EPO uses the COMVIK approach for mixed inventions (inventions combining technical and non-technical features). Only features that contribute to the technical character of the invention are considered in the inventive step assessment. AI features that merely automate a business process contribute nothing to technical character and are ignored in the analysis.

The specification must clearly articulate the technical problem solved and the technical effect achieved by the AI. A trained model that produces a measurable improvement in a technical process has technical character; a trained model that produces a business recommendation does not.

China-Specific: Hardware Integration Required

CNIPA's 2021 examination guidelines for AI inventions require that the claim describe a "technical solution" — which in practice means the claim must include hardware elements. A claim describing only algorithmic steps without hardware context faces rejection. Include specific hardware (processors, sensors, actuators, memory), data flow between hardware components, and the relationship between the AI processing and the physical system being controlled or monitored.

Trade Secret vs Patent for AI

AI presents a genuine patent-vs-trade-secret dilemma:

Model weights and training data are often better protected as trade secrets — they cannot be reverse-engineered from the model's outputs (in most cases), they are not disclosed in a patent, and they can be maintained indefinitely. A patent publishes the architecture and method for the world to read.

The architecture and training method may be better protected by patent — if competitors could independently discover the same approach (through published research, open-source experimentation, or hiring your researchers), trade secret protection fails the moment the secret is independently developed. A patent provides 20 years of protection regardless of independent discovery.

The practical approach: Patent the method and system (how the AI is applied to solve the specific technical problem). Keep the trained model weights, training data, hyperparameters, and proprietary datasets as trade secrets. This combination provides both published, enforceable patent rights and unpublished, perpetual trade secret protection.

AI Inventorship: Current USPTO Rules

The USPTO has issued revised inventorship guidance for AI-assisted inventions, superseding earlier interim guidance. This is the current governing framework for any inventor using AI in their development process.

The core rule: AI cannot be named as an inventor or joint inventor on any patent application — utility, design, or plant. AI systems are tools, analogous to laboratory equipment or software. Only natural persons can be inventors. An application listing an AI system as an inventor will be rejected under 35 USC §§101 and 115.

The inventorship test: The traditional "conception" standard applies. A human inventor must have formed a "definite and permanent idea of the complete and operative invention." Using AI to assist in the inventive process does not change this standard — the question is whether the human conceived the invention, not whether AI helped along the way.

Joint inventorship: The Pannu factors (requiring each joint inventor to make a significant contribution to conception) apply only among multiple human contributors. They do not apply between a human and an AI system. If a single human uses AI to develop an invention, the inventorship analysis is straightforward: did that human conceive the invention?

What this means for AI/ML inventors:

Document your human conception process. Before running AI experiments, document the problem you identified, the approach you conceived, and the hypotheses you formed. When AI generates multiple outputs, document why you selected, modified, or combined specific outputs. Use timestamps and version control to show the progression from human conception through AI-assisted development.

Do not list AI tools as inventors, co-inventors, or contributors in patent applications. Do not describe the AI as having "invented" or "discovered" anything in the specification — frame the AI as a tool used by the inventor.

For foreign filings: the USPTO will not accept priority claims to foreign applications listing an AI system as an inventor. Ensure international filings name only natural persons as inventors, even if foreign law is more permissive.

The Recentive Analytics Warning

The Federal Circuit's decision in Recentive Analytics, Inc. v. Fox Corp. significantly tightened the §101 eligibility bar for AI/ML claims. The court held that patents which do no more than claim the application of generic machine learning to new data environments, without disclosing improvements to the machine learning models themselves, are patent ineligible under §101 — even if the AI achieves improved accuracy or efficiency compared to prior methods.

What this means in practice: Claims that say "apply a neural network to [standard task] and get better results" are no longer sufficient. The claim must identify a specific, non-generic technical improvement — in the model architecture, in the data processing pipeline, in the hardware implementation, or in the interaction between the AI system and the physical system it controls. Generic application of known ML techniques to new data domains is increasingly vulnerable to §101 rejection.

How to protect yourself: The claim drafting strategy described earlier in this article (specific input data, specific processing, specific output, specific technical improvement) is designed to survive Recentive Analytics scrutiny. The key is specificity — the more concrete and technically detailed the claim, the stronger its §101 position.

Worked Example: Predictive Maintenance AI

An inventor develops a system that predicts bearing failure in industrial turbines using acoustic emission data processed by a trained neural network.

Independent system claim:

"A predictive maintenance system for rotating machinery, comprising: an array of acoustic emission sensors mounted on a turbine housing at defined monitoring points; a processor executing a trained neural network model, the model comprising a one-dimensional convolutional layer followed by a recurrent layer, configured to receive time-series acoustic emission data from the sensor array and to generate a remaining useful life prediction for each monitored bearing; and an alert module configured to transmit a maintenance notification to a scheduling system when the remaining useful life prediction falls below a configurable threshold; wherein the system achieves a prediction accuracy of at least 93% for remaining useful life predictions within a 500-operating-hour window."

This claim recites specific hardware (sensors, processor, alert module), specific AI architecture (1D convolutional + recurrent), specific input (acoustic emission time-series), specific output (RUL prediction + maintenance alert), and specific performance (93% accuracy within 500 hours). It survives Alice (technical improvement to machinery monitoring), satisfies the EPO's technical effect test (measurable improvement in maintenance prediction), and includes the hardware integration required by CNIPA.

Sources

1. USPTO Revised Inventorship Guidance for AI-Assisted Inventions (Nov 2025) — Current USPTO framework for AI-assisted inventorship
2. 35 U.S.C. § 101 — Patent Eligibility — US statutory basis for patent-eligible subject matter (Alice framework)
3. EPO Guidelines for Examination — AI/ML — European patentability standards for AI inventions (Part G, Chapter II)
4. CNIPA Patent Examination Guidelines — China's 2021 examination guidelines for AI-related inventions
5. JPO AI/ML Examination Guidelines — Japan's 2019 guidelines for AI and IoT patent examination