Artificial neural networks are at the core of modern artificial intelligence. From large language models to image classifiers, from recommendation engines to medical diagnostics, the neural network architecture is the computational engine driving commercial AI deployment. The question of whether those architectures, and the inventions they enable, are patentable in the UK has been answered by the Supreme Court in Emotional Perception AI Limited v Comptroller General of Patents [2026] UKSC 3. Issues do, however, remain.

The ruling is good news for AI developers, and indeed to industries that develop computer implemented inventions. The decision establishes that ANN-based inventions are not automatically excluded from patent protection in the UK, that the correct analysis involves a properly structured multi-stage assessment, and that the threshold for clearing the statutory exclusion has been significantly lowered to the presence of “any hardware.” But the ruling also clarifies that an ANN is itself a “program for a computer” within the meaning of UK and EPC patent law, a finding that has implications for claim strategy, specification drafting and prosecution approach. This finding is not ideal and is technically questionable, so one cannot rule out it being challenged or corrected, possibly via a very specific piece of legislation.

This article delves into the Supreme Court decision and, particularly, what it means for players developing – and seeking to protect through patent rights – AI-based technologies.

Is an artificial neural network a computer program under UK patent law?

Yes. The Supreme Court held that an ANN, whether implemented in dedicated hardware or emulated in software on a conventional computer, is a “program for a computer” within the meaning of Art.52(2)(c) EPC [European Patent Convention] and section 1(2)(c) of the Patents Act 1977.

The court rejected the applicant’s argument that a “hardware ANN” is a physical machine rather than a computer program and that an emulated ANN, being functionally equivalent, should be treated in the same way; this is questionable logic and potentially legal abstraction. 

The court held that an ANN is properly described as an abstract entity or computational model, not a physical object. It can be implemented on various types of hardware (dedicated neuromorphic chips, field-programmable gate arrays, conventional digital computers), but it is not itself a type of hardware. It is the abstract arrangement of neurons, weights, biases and activation functions that constitutes the ANN, and that arrangement is, in essence, a set of instructions to manipulate data in a particular way. That, in the court’s judgment, makes it a program for a computer notwithstanding that these resulting neuron “weighting factors” – whether finally frozen or under training adaptation – define a new electronic configuration that is computer-implemented but which does not need of a processor [in the sense of interaction at the level of each neuron] to deliver a final output from the network.

The court’s reasoning was that, at the end of the training process, the weighting factors on aggregation constitute a set of instructions for execution of the implementing hardware, wherein the individual neurons process data to produce a particular output. The fact that those instructions are created through a machine learning process, rather than written by a human programmer, does not affect their character as a computer program. The court noted that it is commonplace for computer programs to generate other computer programs, and that there is no principled basis for treating machine-generated instructions differently from human-generated ones. That, unfortunately, appears to represent a disingenuous mischaracterisation of human ingenuity. Neural networks are trained based on “learning objectives” that sit at a conceptually higher level of generalisation, with this contrasting with the ability of existing Large Language Models “LLMs” to generate a program through selection and knitting together of core program modules from a library.

Does classifying ANNs as computer programs mean they cannot be patented?

No, and this is a critical point. The fact that an ANN is a “program for a computer” merely engages the computer program “as such” exclusion of Art.52(2)(c) as tempered by Art.52(3) EPC. This first hurdle requires further qualification. 

Under the “any hardware” approach now adopted in Emotional Perception, any claim that involves hardware, which any ANN implementation necessarily does, is not a claim to a computer program “as such.” The presence of hardware, however conventional, anywhere within the claim is sufficient to take the claim outside the “as such” statutory exclusion, so a “computer-implemented method or a processor-based ANN achieves this objective. 

This means that ANN-based claims are not excluded from patentability at the Art.52 EPC/section 1(2) UKPA stage. What then matters is whether the invention satisfies novelty and inventive step, with this requiring an understanding of what amounts to the invention. An “intermediate step” between the exclusion analysis and inventive step consideration investigates the features of a claim that interact with one another to contribute to its technical character or, more appropriately, the technical effect achieved by the invention. This is a forward thinking analysis. 

This intermediate step is therefore capable of differentiating true technical invention over sound engineering, with the latter (for example) manifest in the straightforward automation of a business process. Contrastingly, in the former respect, the intermediate step identifies the underlying technically beneficial realisation of the invention “as a whole;” this aligns with the general principles expressed by the Technical Board of Appeal “TBA” decisions in, at least, T1249/2022 (paragraph 10 to 12) and T1769/10 (section 4,2) 

What is the “intermediate step” for ANN patent claims?

The intermediate step is the analytical bridge between the “any hardware” exclusion threshold and the full novelty and inventive step assessment. It was described by the EPO’s Enlarged Board of Appeal in G1/19 and endorsed by the Supreme Court in Emotional Perception as a necessary part of the analysis for computer-implemented inventions; it acts as an objective filter.

For ANN-based claims, the intermediate step requires a feature-by-feature examination of the claim to determine which elements contribute to the technical character of the invention when the invention is considered “as a whole”. This is particularly important for “mixed” inventions, those that combine ANN components with other system features, some of which may be technical and some of which may serve non-technical purposes.

The intermediate step essentially asks two key questions about each feature of the claim:

  1. Does the feature contribute to the technical character of the invention when considered contextually as a whole? This potentially eliminates technical features that are irrelevant to the invention.
  2. Does the feature – whether technical or not – interact with other technical subject matter of the claim in a way that supports delivery of an overall technical effect?

Features that pass this analysis, either as inherently technical contributing features or as non-technical features that interact with the technical subject matter, are included in the inventive step assessment. Features that do not contribute to the technical character are filtered out.

For AI patent applicants, this means that the specification and claims must clearly articulate the technical interactions between the ANN architecture – and indeed how the ANN is trained and assessed relative to some form of cost function and ground truth – and other system components. An ANN training method that produces a specific technical effect, improved signal quality, greater compression efficiency, more accurate anomaly detection, through its interaction with hardware, sensors, networks or databases has a stronger claim to technical character than one whose described effects are purely functional or abstract.

How should AI developers structure ANN patent claims?

The practical lessons from the Emotional Perception ruling and the intermediate step analysis point toward several claim drafting principles for ANN-based inventions.

Claim the system, not just the model. A claim directed solely to a trained ANN, its weights, biases and topology, without reference to the hardware environment in which it operates, may struggle to clear the “as such” exclusion. A claim directed to a system that includes the ANN, hardware components (computing hardware, sensors, databases, communications networks) and a defined method of operation is on stronger ground.

Describe the training method in technical terms. ANN training is a technical process. The choice of training objective, loss function, dataset structure, and backpropagation configuration are technical decisions that can contribute to the technical character of the invention. The specification should explain why those technical choices were made and what technical effect they achieve, not simply describe them as steps in a generic training loop. This supports protection of the ANN per se as a “product-by-process.”

Identify what the trained model does differently. The inventive step assessment will focus on what is different about the claimed invention compared to the prior art. For ANN inventions, this often means explaining clearly what the trained model can do that prior art systems could not, with precision about the technical mechanism, not just the commercial outcome. The objective is the identification of a new product produced by a new process to deliver a new result.

Preserve dependent claim fallbacks on architecture. ANN architecture choices, the number of layers, the type of activation functions, the connectivity pattern, the embedding dimension, can all be technically significant, but they need to be expressed in the context of applied technology rather than an abstract idea. Dependent claims that specify these architectural choices may provide fallback positions for prosecution and can support an inventive step argument. They should be expressed in terms of the effects that they achieve and, potentially, why the structure was chosen to support a particularly beneficial technical realisation.

How are different types of ANN treated under UK patent law?

The Supreme Court’s analysis treated ANNs as a category without distinguishing between different architectures or training paradigms. The reasoning applies broadly to:

Feedforward networks and deep learning models. Standard supervised learning models trained by backpropagation fall squarely within the court’s description of an ANN. The weights and biases of a trained feedforward network are computer programs; the system implementing the trained network involves hardware; the claim is not excluded “as such.”

Recurrent networks and transformers. These architectures involve more complex connectivity patterns and, in the case of transformers, attention mechanisms rather than traditional weight layers. The same analysis applies: the architecture is an abstract entity, its implementation requires hardware, and the claim to the system is not excluded “as such.”

Generative models. GANs, VAEs, diffusion models and large language models are all ANN-based systems. Their patentability under UK law depends on the same analysis: technical character and technical effect, feature interaction in the intermediate step, and novelty and inventive step of the specific architecture, training approach or application.

Embedded and hardware-implemented ANNs. Where an ANN is implemented in dedicated hardware, a neuromorphic chip, an ASIC specifically designed for neural network inference, the hardware implementation does not change the legal character of the ANN. It remains a program for a computer. The hardware implementation does, however, make it even clearer that the claim involves hardware and is therefore not directed to a program “as such.”

Can the Emotional Perception ruling help with previously rejected ANN applications?

Potentially, yes. But it will depend on the timing/date of the rejection and therefore whether the application can be revived. Applications that were refused – or subject to an ongoing refusal – primarily on the basis of the Aerotel ‘contribution with rider’ analysis, may therefore be recoverable in light of the changed legal framework. The contribution with rider idea, now rejected as “unsound,” put the cart before the horse, essentially asking a backwards facing question of what the invention had contributed to the font of human knowledge.

The change in law does not automatically entitle any refused application to be passed to allowance and ultimately to grant. Each case depends on its specific facts: the claim language, the way the technical contribution was presented, the procedural status of the application and the strength of the underlying invention. But where an application was refused under Aerotel in circumstances that would not have produced a refusal under the “any hardware” approach and, particularly, the “intermediate step” that properly considers the invention “as a whole,” there may be grounds to revisit the position.

For AI developers who have previously been advised that UK patent protection for an ANN-based invention was not viable, it is worth reviewing that position. With Hepworth Browne recognised as having world-leading experts in this area of national and international patent law, we are well-placed to advise on securing patent rights in this important area of technical commerce.

For innovations – whether pre-filing or in active prosecution – Hepworth Browne provides guidance on strategy, specification drafting and prosecution approaches for ANN-based and computer-implemented “CII” inventions that reflect, particularly, Bruce Dearling’s work before the UK Supreme Court in our Emotional Perception case. You can also explore the broader AI patentability framework on the firm’s main AI practice page.

Frequently asked questions: ANNs and UK patents

Is a trained neural network model a patentable invention? The trained model itself, viewed in isolation, is a computer program and therefore potentially excluded. A system that includes the trained model, hardware and other technical components may be patentable depending on the technical character of the claimed invention “as a whole” and, fundamentally, whether it is differentiated from the prior art.

Can the training process for an ANN be patented? A training method can potentially be patented if it produces a technically inventive outcome, e.g., a trained model with particular technical properties achieved through a technically non-obvious training approach. The training method claim must therefore include technical features that deliver, in the final network, a tangible benefit.

Does it matter whether the ANN runs on a GPU, CPU or dedicated chip? For the purpose of the “as such” exclusion analysis, no. The Supreme Court confirmed that the choice of hardware implementation is irrelevant to the legal character of the ANN. For the intermediate step and inventive step analysis, the specific hardware is generally irrelevant unless the choice is specific in that the choice also contributes to an overall technical character of the invention.

How does the Emotional Perception ruling affect ANN patent strategy at the EPO? The EPO already applied the “any hardware” approach. The ruling primarily aligns UK practice with the existing EPO framework and the use of the intermediate step. 

The EPO, however, views ANNs as mathematical methods and not computer programs. Both require a positive application of the underlying technology to overcome potential assertions that the claimed invention is abstract.

For dual UK/EPO applicants, the convergence to “any hardware” means that prosecution strategies developed for the EPO are more directly applicable to UK prosecution than they were before. However, the UK approach to assessing inventive step – the Pozzoli questions and particularly the question ‘are the differences obvious’ – is different to the problem-solution approach. Pozzoli arguably is not inhibited with hindsight reasoning that can arise [to distort the actual invention]  from formulating a problem based on identified technical differences between the claimed invention and the prior art.

How does the Emotional Perception ruling affect ANN patentability in other jurisdictions?

The approach taken by the Unified Patent Court “UPC” – following the UPC’s headnote 1 in its Court of Appeal decision UPC-CoA-0000901/2025 – now also aligns with the UKSC decision, albeit that both the UPC and EPO favour the “problem-solution” approach to the assessment of inventive step. 

Whether countries, like Singapore, follow the UKSC Emotional Perception decision appears likely, at least to the extent that “any hardware” and the “intermediate step” are objectively sensible assessment mechanisms. Whether these countries adopt the restrictive definition of an ANN being a computer is less evident. More liberal views about what amounts both to a computer program and the nature of ANNs appears technically and commercially more likely, especially given the response of the Association of Singapore Patent Attorneys “ASPA” to the IPOS consultation on the relevance of the Emotional Perception decision to sovereign law in Singapore.

In light of the effective codification of the Aerotel decision into New Zealand patent law (see section 11), the Emotional Perception decision holding Aerotel to be “unsound” may have influence on the continuing appropriateness of this test in NZ. Other national states influenced by the UKSC may similarly be subject to this influence and line of argument.