Our Blog

AI and Machine Learning Patent Applications Under CIPO’s Physicality Standard

April 29, 2026

Part 2 of a 4-part series on subject matter eligibility after CIPO’s March 2026 Practice Notice.

What CIPO Says

Computer-implemented Example 3 in the March 2026 Practice Notice describes a system that uses a layered neural network, trained on historical data, to generate crop irrigation schedules. The specification discloses the neural network architecture, the training methodology, and the resulting outputs.

CIPO concludes that a claim directed to the system, without physical irrigation equipment, is not patentable on the basis that nothing other than an algorithm has been discovered. The analysis treats the neural network as an abstract algorithm implemented on a computer.

How the Analysis Operates in Practice

Neural networks are treated as algorithms unless tied to computational effects

In the examples, a trained neural network is treated as an abstract algorithm where the claimed invention is framed in terms of generating a recommendation or prediction. The presence of a detailed architecture and training process does not, on its own, prevent this characterization.

In practice, this means that a neural network may be viewed as an implementation of an algorithm unless the specification and claims connect the model to identifiable computational effects.

Computational improvements define the eligibility path

The Practice Notice indicates that improvements to the operation of the computer, including reduced memory usage, fewer computations, or faster processing, can satisfy the physicality requirement.

For machine learning systems, these improvements often arise from the trained model itself. A model that achieves a given level of performance with fewer parameters, reduced computation, or improved inference efficiency affects how the computer performs the task. These types of effects align with the Notice’s description of improvements to the computer’s operation for a specific task.

Specification detail affects CGK characterization

The examples show that elements described with limited technical detail may be treated as common general knowledge. This applies to programming, model implementation, and system components.

For machine learning inventions, this creates a direct link between how the architecture and training process are described and how those elements are treated in the eligibility analysis.

Prosecution Takeaways

Describe the model architecture as a technical contribution

The specification should present the neural network architecture as part of the invention, not as background implementation. Where the architecture provides computational advantages, those advantages should be described in technical terms.

This includes, where applicable, parameter efficiency, reduced computational complexity, improved throughput, or other measurable effects on system performance.

Characterize training in computational terms

The training process should be described as a sequence of computational operations that result in a defined configuration of the system. Where the training process produces models with improved efficiency or performance characteristics, those outcomes should be identified and described.

Claims directed to trained models or training methods should reflect those characteristics where they form part of the technical contribution.

Identify and support computational efficiency explicitly

Where the invention achieves improvements in inference speed, memory usage, or computational load, those improvements should be clearly identified in the specification. These effects provide a basis for distinguishing the invention from an abstract algorithm implemented on a computer.

Align claim structure with the technical contribution

Independent claims should incorporate the features that provide the computational improvement where possible. Claims that focus only on generating outputs, such as predictions or recommendations, without reciting how the computation is improved, are more likely to be characterized as abstract.

Dependent claims can provide additional fallback positions, but the primary eligibility position should not depend on downstream physical actions.

Address CGK characterization in prosecution

Where an examiner treats elements of the model or implementation as common general knowledge, responses should refer to the specification’s technical disclosure to support their characterization as part of the invention. The level of detail in the specification can be used to distinguish those elements from conventional techniques.

Under the March 2026 Practice Notice, eligibility for machine learning inventions depends on how the invention is framed in computational terms. Where the specification and claims identify measurable improvements in how the computer performs the task, those improvements provide a basis for satisfying the physicality requirement.

The next post in this series addresses quantum computing inventions and how the physicality requirement applies in that context.