Rethinking the Legal Ontology of Machine-Learning Inventions in Global Patent Law

A Comparative Legal Analysis of AI Patent Eligibility Doctrine

United Kingdom • United States • European Patent Office • China

Independent Legal Research Paper

August 2026

Abstract

For half a century, patent systems have examined artificial intelligence inventions through legal frameworks built for conventional software — asking whether a claim is “more than” a computer program rather than whether AI constitutes a distinct technological category. This paper uses the United Kingdom Supreme Court’s 2026 judgment in Emotional Perception AI Ltd v Comptroller-General of Patents as a lens through which to examine that inherited paradigm. Contrary to a popular reading of the case, the Supreme Court did not hold that a trained artificial neural network escapes the statutory computer-program exclusion; it held the opposite, while simultaneously abandoning the twenty-year-old Aerotel test and re-aligning UK law with the European Patent Office’s more permissive technical-contribution methodology. Read carefully, the decision illustrates both the resilience and the limits of the software paradigm. This paper surveys the parallel doctrinal struggles underway in the United States, at the European Patent Office, and in China; asks whether the technical distinction between trained models and conventional programs is legally significant; proposes a purpose-built four-factor framework for assessing AI inventions; and draws lessons from how patent law previously adapted to biotechnology, semiconductor topography, and business methods. It concludes that AI does not yet require a wholesale new legal species, but that discrete doctrinal reforms — to enablement, obviousness, and the identification of technical contribution — are overdue.

Table of Contents

1. Introduction
2. The Inherited Paradigm: AI as Software

III. Emotional Perception: A Four-Year Doctrinal Journey

1. A Comparative Survey of AI Patent Eligibility
2. Is Artificial Intelligence Ontologically Different from Software?
3. Toward a Purpose-Built Analytical Framework for AI Inventions

VII. Learning from History: How Patent Law Has Adapted Before

VIII. Institutional and Policy Considerations

1. Recommendations
2. Conclusion

References

I. Introduction

One of the most consequential developments in artificial intelligence patent law of the past several years is not a new model architecture or a new capability benchmark. It is a slow, uneven, and still-unfinished judicial reckoning with a question that patent offices have mostly avoided asking directly: is a trained machine-learning system properly understood as “just software,” or does it belong to a different technological category altogether?

For decades, the answer has been assumed rather than argued. Patent examiners, courts, and commentators have treated artificial intelligence inventions as a subspecies of computer-implemented invention, to be tested under legal frameworks that predate the deep-learning era by twenty, thirty, or in some cases fifty years. The dominant inquiry — in the United Kingdom and at the European Patent Office, whether an invention makes a “technical contribution” beyond the mere running of a program; in the United States, whether a claim is “directed to” an abstract idea and, if so, whether it contains an “inventive concept” — was developed for spreadsheets, databases, word processors, and business-transaction systems. Artificial intelligence inventions have simply been poured into that mould.

The United Kingdom Supreme Court’s judgment in Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks, handed down on 11 February 2026, offered the first apex-court opportunity in a major jurisdiction to examine that mould directly. The case asked, in terms a court had never previously had to answer squarely, whether a trained artificial neural network (“ANN”) is a “program for a computer” within the meaning of section 1(2)(c) of the Patents Act 1977. The judgment has been described by practitioners as a landmark and, in press coverage, has sometimes been characterised as recognising that AI is not “just” conventional software. That characterisation requires close scrutiny, because — as this paper explains in Part III — it does not accurately capture what the Supreme Court actually decided on the categorisation question, even though the decision is undoubtedly significant for AI patenting in the United Kingdom for other reasons.

This paper takes Emotional Perception as its point of departure, but its purpose is broader. It asks whether the software paradigm that patent law has applied to AI — across the United Kingdom, the European Patent Office, the United States, and China — remains fit for purpose, and if not, what should replace it. The argument proceeds in three movements. First, it examines the doctrinal history that led patent systems to treat AI as software by default, and traces, in detail, the four-year journey of the Emotional Perception litigation from the UK Intellectual Property Office through the High Court, the Court of Appeal, and finally the Supreme Court, correcting several points on which public commentary has diverged from the judgment’s actual reasoning. Second, it places that UK development alongside parallel doctrinal struggles in the United States (culminating in the Federal Circuit’s first precedential machine-learning eligibility decision, Recentive Analytics, Inc. v. Fox Corp.), at the European Patent Office (whose G 1/19 “potential technical effect” doctrine now anchors the UK’s own approach), and in China (whose patent office has gone furthest of any major jurisdiction in promulgating AI-specific examination guidelines). Third, having surveyed the comparative landscape, the paper asks the more fundamental question the Supreme Court chose not to resolve: is there a legally meaningful, technically grounded distinction between a trained model and a conventionally programmed system, and if so, what would a purpose-built patent-eligibility framework for AI look like?

The paper’s ultimate position is a qualified one. It does not conclude that artificial intelligence requires an entirely new title of intellectual property, in the way that semiconductor mask works eventually received sui generis protection. It concludes instead that the persistent, low-grade doctrinal confusion surrounding AI patent eligibility — visible in the fact that four different tribunals reached three different outcomes in a single UK case over four years — is a symptom of applying an outdated diagnostic question. The better question is not whether AI is “more than” software, but where, specifically, the inventive and technical content of a machine-learning system actually resides, and whether existing doctrine is capable of finding it. Part VI proposes a four-factor framework designed to answer that question directly, and Part VII draws on the history of biotechnology, semiconductor, and business-method patenting to caution against both excessive rigidity and excessive permissiveness in adopting it.

II. The Inherited Paradigm: AI as Software

A. Software’s Own Difficult Path to Eligibility

To understand why artificial intelligence was absorbed into software doctrine, it is necessary to recall that software itself did not have an easy path to patent protection. In the United States, the Supreme Court’s early encounters with computer-implemented inventions were deeply skeptical. In Gottschalk v. Benson, the Court held that a method for converting binary-coded decimal numerals into pure binary form, implemented on a general-purpose computer, was not patentable because it amounted to a claim on an abstract mathematical algorithm untethered to any particular application. That skepticism softened somewhat in Diamond v. Diehr, where a computer-implemented process for curing rubber was held eligible because the claim, taken as a whole, applied a mathematical formula to achieve a specific, useful, tangible result in an industrial process. The tension between Benson and Diehr — between treating an algorithm as inherently abstract and treating its practical application as patent-eligible — has never fully been resolved, and it is the direct ancestor of the two-step framework the Supreme Court later articulated in Mayo Collaborative Services v. Prometheus Laboratories and Alice Corp. v. CLS Bank International, discussed further in Part IV.

In the United Kingdom and before the European Patent Office, the path was different but the destination similarly uneasy. Article 52(2)(c) and (3) of the European Patent Convention excludes “programs for computers” from patentability “as such,” language mirrored in section 1(2)(c) of the UK Patents Act 1977. The qualifying phrase “as such” has generated decades of interpretive litigation, because it implies that some computer programs are patentable and others are not, without specifying the dividing line. The EPO’s answer, refined through the COMVIK line of Board of Appeal decisions, is to treat any mixture of technical and non-technical features by asking whether the non-technical features contribute to the solution of a technical problem; only technical contributions count toward inventive step. UK courts, for their part, developed the four-step Aerotel test in 2006, asking a court to (i) properly construe the claim, (ii) identify the actual contribution, (iii) ask whether it falls solely within excluded subject matter, and (iv) check whether the contribution is technical in nature.

The result, in both jurisdictions, was a body of doctrine calibrated to conventional software: word processors, spreadsheets, database indexing schemes, e-commerce checkout flows, and telecommunications protocols. The recurring examiner question — does this claim do anything beyond what any programmer would recognise as “just running code” — was designed to filter out attempts to dress up abstract business methods or mathematical algorithms in computer-implemented clothing. It was not designed with any particular attention to a technology in which the “code” itself is not authored by a human being in the traditional sense.

B. Machine Learning’s Silent Absorption into Software Doctrine

When machine-learning inventions began reaching patent offices in meaningful volume during the 2010s, they were not greeted with a new doctrinal apparatus. They were simply treated as computer-implemented inventions like any other, and examined under the existing tests. A patent claim reciting “training a neural network on a dataset of labelled examples” was analysed, structurally, in the same way as a claim reciting “querying a relational database”: identify the technical and non-technical features, ask whether the non-technical features (here, the abstract mathematical operations of training) contribute to a technical effect, and grant or refuse accordingly. Examiners in the United States applied the same abstract-idea categories — mathematical concepts, mental processes, and methods of organising human activity — to machine-learning claims that had been developed for older forms of automation.

This absorption was, in one sense, entirely defensible as an interim measure: the alternative would have required patent offices to draft an entirely new examination apparatus for a technology whose contours were, in the early 2010s, still unsettled, based on incomplete information about how the technology would mature and what its characteristic forms of innovation would look like. But the absorption also carried a latent assumption that has rarely been examined on its own terms: that whatever distinguishes a trained model from a hand-written program is legally irrelevant. That assumption is the target of this paper’s inquiry.

C. Why the Fit Is Imperfect

Three features of modern machine learning sit uneasily within doctrine built for conventional software. First, a trained neural network’s operative behaviour is encoded in a large set of numerical parameters — weights and biases — that are not directly authored by a human programmer. They emerge from an iterative optimisation process operating over training data, guided by a loss function and an architecture chosen by the developer, but not dictated line-by-line. A developer who designs a transformer architecture and a training regime cannot, in general, predict the specific numerical values the model will converge to, nor fully explain after the fact why the model produces a particular output for a particular input. This is qualitatively different from the relationship between a programmer and a conventional if-then-else routine, even a highly complex one.

Second, the inventive labour in a machine-learning project is frequently concentrated in places that classical software-patent doctrine is not built to recognise: the curation and labelling of a training dataset, the design of a training curriculum or data-augmentation strategy, the choice of a loss function that induces a particular emergent behaviour, or an architectural innovation that improves sample efficiency. None of these map cleanly onto the “algorithm implementing a business process” template that motivated the exclusionary doctrines discussed above.

Third, and most consequentially for patent drafting, the opacity of trained models creates a genuine tension with the enablement and sufficiency requirements that anchor the patent bargain. A patent specification is supposed to teach a person skilled in the art how to reproduce the invention. For a conventional algorithm, this is achieved by disclosing the algorithm’s steps. For a trained model, exhaustively disclosing the final parameter values would satisfy a narrow, formalistic reading of sufficiency while teaching almost nothing about how to obtain similar results through a different but equally valid training run; conversely, disclosing only the training methodology may not guarantee that a skilled person following the same recipe would reproduce the same technical effect, given the stochastic nature of training. The European Patent Office’s Boards of Appeal have already flagged this tension in decisions such as T161/18 and T1191/19, discussed in Part IV, without yet resolving it.

III. Emotional Perception: A Four-Year Doctrinal Journey

No single case illustrates the strain between inherited software doctrine and modern machine learning better than Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks. Between 2021 and 2026, the case passed through the UK Intellectual Property Office, the High Court, the Court of Appeal, and finally the Supreme Court, with the outcome reversing at almost every stage. Its doctrinal odyssey is worth recounting in some detail, both because the final judgment reshapes UK law on computer-implemented inventions generally, and because the popular narrative surrounding the decision diverges from its actual holding on the question this paper cares about most.

A. The Invention and the UKIPO’s Refusal

Emotional Perception AI Limited (“EPAI”) sought to patent a media-recommendation system built around a pair of artificial neural networks. One network was trained to analyse the semantic properties of a media file — descriptors such as mood or genre, derived from human-generated tags or metadata. A second network was trained to analyse the file’s physical or musical properties directly from the underlying signal, without reference to human-assigned categories. The two networks were trained, through a contrastive process, so that files perceived as semantically similar would also sit close together in the physical-property network’s output space, and vice versa. The trained system could then take a “seed” file supplied by a user and recommend another file located nearby in that jointly learned property space — without relying on the pre-existing genre labels that conventional recommender systems use.

The UK Intellectual Property Office refused the application. Its Hearing Officer concluded that the invention’s contribution amounted to no more than a computer program, producing no technical effect beyond the ordinary running of a program on a computer, and was accordingly excluded under section 1(2)(c) of the Patents Act 1977 as a “program for a computer… as such.” EPAI appealed.

B. The High Court: A Categorical Distinction

In a judgment delivered in November 2023, Sir Anthony Mann, sitting as a judge of the High Court, allowed EPAI’s appeal. Central to his reasoning was a distinction between a conventional computer program — a predetermined sequence of instructions written by a human programmer — and a trained artificial neural network, whose weights are arrived at through an automated training process rather than direct human authorship. On this reasoning, an ANN, once trained, is arguably not executing a “program” in the sense contemplated by the 1977 Act’s exclusion at all. The High Court also found, independently, that the invention produced a technical effect: the recommended file was, in a qualitative sense, a better or more apt recommendation than a human-curated genre-based system could produce, and that improvement existed outside the computer in the sense the exclusionary case law requires.

The High Court’s judgment was widely read, at the time, as the first serious judicial recognition that machine learning might occupy a different doctrinal space than conventional software — precisely the intuition reflected in the opening premise of the debate this paper addresses. That reading proved short-lived.

C. The Court of Appeal: A Sharp Correction

The Comptroller-General appealed, and in a judgment delivered by Lord Justice Birss on 19 July 2024, the Court of Appeal reversed. The court’s reasoning proceeded from a deliberately formalist starting point: a computer is a machine that processes information; a computer program is a set of instructions that causes a computer to process information in a particular way; and an artificial neural network, however it is physically or logically implemented — whether in dedicated hardware or as software executing on general-purpose hardware — is itself properly characterised as a computer running a program. On this view, the fact that an ANN’s weights are learned through training rather than directly written by a human programmer does not remove it from the statutory exclusion, because the exclusion is concerned with what the machine does when it runs, not with the provenance of the instructions it executes.

On the separate question of technical effect, the Court of Appeal held that a media recommendation which a user is more likely to enjoy is, at bottom, a cognitive or aesthetic improvement rather than a technical one in the sense the exclusionary case law requires. The invention’s output was information about which the user might make a better subjective choice, not a technical effect operating on or through the computer system itself. On both grounds, the Court of Appeal restored the UKIPO’s original refusal.

EPAI sought and was granted permission to appeal further, to the Supreme Court — the first occasion on which the UK’s apex court would rule directly on the patentability of a machine-learning invention.

D. The Supreme Court: Realignment, Not Revolution

The Supreme Court handed down judgment on 11 February 2026, in a decision reported as [2026] UKSC 3. The Chartered Institute of Patent Attorneys and the IP Federation intervened, and the Court used the appeal as an opportunity to reconsider not only the narrow ANN question but the entire methodology UK courts had applied to computer-implemented inventions for two decades.

On the narrow point — whether a trained artificial neural network is a “program for a computer” — the Supreme Court agreed with the Court of Appeal, not the High Court. It held that an ANN, whatever its underlying implementation, is a computer, and that its trained, operative state is properly characterised as a program for the purposes of section 1(2)(c). The distinction the High Court drew between learned weights and human-authored instructions was not accepted as legally significant to the threshold classification question.

What makes the judgment significant, however, is what the Court did next. Rather than stopping at that classification and reinstating an automatic exclusion, the Supreme Court held that classifying an invention as involving a computer program does not, by itself, resolve patentability — a claim is not excluded merely because it refers to the use of a computer, a computer-readable storage medium, or other technical means. The operative inquiry, the Court held, must instead ask whether the invention makes a technical contribution once features that do not contribute technically are filtered out. In reaching that conclusion, the Court expressly abandoned the twenty-year-old four-step Aerotel framework and replaced it with a methodology aligned with the European Patent Office’s COMVIK approach, including the Enlarged Board of Appeal’s “potential technical effect” doctrine from G 1/19 — the notion that a technical effect can be recognised even where the claim does not itself specify the downstream technical use to which the AI system’s output will be put, provided such a use is a plausible, intended application of the underlying calculation.

The Court also addressed, without disturbing, the established approach to inventive step: the familiar Pozzoli framework for assessing obviousness remains applicable, though the Court left open whether its methodology might require future adjustment in light of the eligibility realignment. The practical upshot, as commentators have noted, is that AI-related inventions will now pass the threshold “is this an invention at all” stage more easily than under the old Aerotel approach, but will face correspondingly closer scrutiny at the technical-contribution, novelty, and inventive-step stages that follow.

E. What the Judgment Actually Decided — Correcting the Popular Narrative

It is worth pausing on a point of some importance for the broader debate this paper engages with. Commentary on Emotional Perception has sometimes framed the Supreme Court’s decision as vindicating the intuition that a trained neural network should not be regarded as “just” a conventional computer program — an intuition that, as noted in the Introduction, animates much of the current public conversation about AI patent law. On a careful reading of the judgment, that framing inverts the actual holding on the categorisation question. The Supreme Court did not accept that an ANN escapes classification as a computer program; it affirmed, against the High Court and in line with the Court of Appeal, that an ANN is a computer running a program, regardless of whether its parameters were learned or hand-written.

What changed is not the ontological status of AI relative to software, but the permissiveness of the downstream technical-contribution filter and the abandonment of a UK-specific test that had, over two decades, become stricter than its European counterpart. The judgment is properly understood as a doctrinal realignment — bringing UK law back into step with the EPO’s COMVIK/technical-effect methodology — rather than as a declaration that machine learning is a categorically distinct subject matter deserving its own exclusion-free treatment. This distinction matters, because it means the underlying question this paper poses — whether AI inventions warrant a purpose-built analytical framework distinct from software doctrine — remains genuinely open even after the UK’s highest court has spoken. The Supreme Court sidestepped that deeper question by resolving the case on narrower, more traditional grounds: harmonising the technical-contribution methodology rather than reclassifying the technology. Parts V and VI of this paper take up the question the Supreme Court left unresolved.

IV. A Comparative Survey of AI Patent Eligibility

The United Kingdom is not alone in wrestling with how to fit machine learning into pre-existing eligibility doctrine. This Part surveys parallel developments in the United States, at the European Patent Office, and in China, in order to situate Emotional Perception within a broader, genuinely global pattern: every major patent system is, in its own idiom, discovering that the software-era tests it inherited were not built with training-based systems in mind.

A. United States

Patent eligibility in the United States is governed by 35 U.S.C. § 101, as glossed by the Supreme Court’s two-step framework from Mayo Collaborative Services v. Prometheus Laboratories and Alice Corp. v. CLS Bank International. Step one asks whether a claim is “directed to” a patent-ineligible concept — a law of nature, a natural phenomenon, or an abstract idea. Step two asks whether the claim, considered as an ordered combination, contains an “inventive concept” sufficient to transform the abstract idea into a patent-eligible application. The United States Patent and Trademark Office’s 2019 Revised Patent Subject Matter Eligibility Guidance organised abstract ideas into three groupings — mathematical concepts, certain methods of organising human activity, and mental processes — and machine-learning claims are routinely tested against all three, since a trained model can plausibly be characterised as a mathematical concept (the underlying statistical function), an automation of a mental process (a task a human analyst previously performed), or both.

The Federal Circuit’s first precedential encounter with machine-learning eligibility came in Recentive Analytics, Inc. v. Fox Corp., decided on 18 April 2025 and reported at 134 F.4th 1205. Recentive held patents directed to using machine-learning models to generate optimised event schedules and television “network maps.” The district court had dismissed Recentive’s infringement suit on the ground that the patents were directed to the abstract ideas of producing schedules and network maps using known, generic mathematical techniques, and that the recited machine-learning steps — iterative training, dynamic updating — did not supply the missing inventive concept. The Federal Circuit affirmed, framing the appeal as a question of first impression: whether claims that do no more than apply established machine-learning methods to a new data environment are patent-eligible. The court answered no, holding that patents which apply generic machine learning to a new data environment, without disclosing improvements to the machine-learning models themselves, do not satisfy § 101.

Recentive is doctrinally important for this paper’s purposes because of what it does not do. The Federal Circuit did not hold that machine learning is categorically excluded, nor did it treat trained models as inherently more or less patentable than conventional software. Instead, it drew the operative line exactly where Part VI of this paper argues it should be drawn: between claims that recite a genuine technical improvement to the training methodology or model architecture, and claims that merely substitute a machine-learning “black box” for a process a human analyst, or a simpler algorithm, previously performed. The court observed that event planners had long used historical data, ticket sales, and weather forecasts to schedule events, and that broadcasters had long created network maps manually; performing the same tasks faster and with greater apparent sophistication by invoking machine learning, without more, does not transform an abstract idea into a patent-eligible invention. The decision generated substantial commentary and, according to court filings, at least one petition for certiorari, reflecting continued uncertainty about how far its reasoning extends to machine-learning patents more generally; the USPTO has since issued internal guidance to its examining corps addressing the eligibility of AI-related claims in light of the decision.

B. European Patent Office

The European Patent Office excludes “programs for computers” as such under Article 52(2)(c) and (3) EPC, but its Boards of Appeal have, since the COMVIK line of decisions, applied a comparatively permissive methodology: any claim reciting technical means (a computer, a processor, a network) is treated as having technical character sufficient to be an “invention” at the threshold stage, with the real filtering work deferred to the assessment of inventive step, where only features that contribute to solving a technical problem are given weight.

The Enlarged Board of Appeal’s 2021 decision in G 1/19 addressed computer-implemented simulations, but its reasoning has proved directly relevant to machine learning, given the structural similarity between simulating a physical system and training a model to approximate one. G 1/19 rejected the proposition that a technical effect requires a direct link to physical reality, and introduced the concept of a “potential technical effect”: a technical effect that would arise if the simulation’s (or, by extension, the model’s) output were put to its intended technical use, even where the claim does not itself recite that final application step. Following G 1/19, the EPO’s Guidelines for Examination, most recently revised through 2025, set out in section G-II, 3.3.1 two routes by which an AI or machine-learning claim can establish technical character: by application to a field of technology (for example, controlling a physical process or diagnosing a technical condition), or by a specific technical implementation motivated by considerations relating to the internal functioning of the computer itself — for instance, an architectural choice that reduces memory consumption or improves training efficiency on particular hardware. Both the concrete examples in the Guidelines and successive rounds of revision reflect an EPO practice that treats AI and machine learning as continuous with, rather than distinct from, its existing computer-implemented-invention doctrine — but one that has, in the process, developed a set of examples and reasoning patterns increasingly tailored to how AI systems are actually built and described.

The Boards of Appeal have also begun to confront the disclosure difficulties described in Part II. Decisions such as T161/18 and T1191/19 flag that a claim may fail the sufficiency requirement of Article 83 EPC where the mathematical methods and training datasets are not disclosed in sufficient detail for a skilled person to reproduce the claimed technical effect without undue burden, using common general knowledge, across the full scope of the claim. This is, in substance, an early judicial recognition of the enablement tension identified in Part II — though the EPO has so far addressed it through case-specific sufficiency objections rather than through a general doctrinal accommodation of how trained models can and cannot be reproduced from a specification.

C. United Kingdom After Emotional Perception

As explained in Part III, the Supreme Court’s judgment brings UK law into closer alignment with the EPO’s COMVIK/technical-effect methodology, including its “potential technical effect” concept, while discarding the domestically developed Aerotel test. This is best understood as a convergence outcome: rather than the UK developing an AI-specific doctrine of its own, it has adopted — for AI inventions and computer-implemented inventions generally — the more permissive continental methodology, which had for two decades produced a materially different (and, on balance, more generous) body of case law than the UK’s own approach. The near-term practical effect, as several commentators have observed, is that fewer AI patent applications will be refused outright at the threshold “invention” stage in the UK; more of the substantive contest will shift to whether the claimed technical contribution is novel and non-obvious.

D. China

China’s National Intellectual Property Administration (“CNIPA”) has gone further than any other major patent office in developing AI-specific examination doctrine, doing so through a series of guideline revisions rather than through litigation. Revisions to the Guidelines for Patent Examination that took effect on 20 January 2024 introduced, under Article 2.2 of the Patent Law, two express eligibility routes for AI and big-data algorithm claims: where the algorithm has a specific technical relationship with the internal structure of a computer system, and where it solves a technical problem that improves the system’s internal performance — defined broadly enough to include not only hardware architecture but also data storage and data-scheduling efficiency.

CNIPA went further still on 31 December 2024, issuing dedicated Guidelines for Patent Applications for AI-Related Inventions (Trial Implementation), following a short public consultation. These guidelines classify AI-related patent applications into four categories — AI algorithms or models considered in themselves, hardware-related AI inventions, field-specific applications of AI to particular industries such as transportation, telecommunications, medicine, finance, and entertainment, and AI-generated inventions — and address, category by category, subject-matter eligibility, disclosure sufficiency, and inventive-step assessment. The guidelines also resolve an inventorship question that has proved contentious elsewhere: an artificial intelligence system cannot itself be named as an inventor, and the name of an organisation or a collective cannot be substituted for an individual inventor’s name.

A further amendment, adopted on 13 November 2025 and effective from 1 January 2026, added an explicit legality and ethics screen: applications built on algorithm or business-method features — including AI and big-data inventions — will not be granted where the underlying data collection, tagging, rule-setting, or recommendation methodology violates law, social ethics, or public interest. The same round of amendments added further worked examples addressing sufficiency of disclosure for algorithm- and data-related inventions and clarified inventive-step assessment where an invention’s contribution lies in improved data sharing between components of a machine-learning system, such as convolutional neural networks exchanging feature information to reduce memory consumption and improve output accuracy.

China’s approach is instructive for the comparative purposes of this paper for two reasons. First, it demonstrates that a major patent office considers the software-derived technical-effect test insufficiently precise for AI inventions as currently drafted, and has been willing to promulgate detailed, AI-specific administrative guidance rather than wait for case law to accrete gradually, as has happened in the UK and at the EPO. Second, its explicit ethics and legality screen for training-data provenance addresses a policy dimension — the propriety of the data used to train a claimed model — that has no clear analogue in conventional software-patent doctrine, and that Western jurisdictions have not yet begun to incorporate into eligibility analysis, even though the underlying concern (patents built on unlawfully obtained training data) is not unique to China.

E. Synthesis: Convergence on Method, Divergence on Detail

Read together, these four bodies of law display a striking pattern. None has adopted, or even seriously entertained, a formal doctrine that AI inventions constitute a distinct statutory category exempt from ordinary eligibility analysis. Each continues to route AI claims through the general computer-implemented-invention framework it applies to software generally: the abstract-idea/inventive-concept inquiry in the United States, the technical-contribution/COMVIK methodology now shared by the UK and the EPO, and the technical-problem/technical-relationship inquiry in China. Yet each has, in its own idiom and at its own pace, begun to develop AI-specific glosses on that general framework — Recentive’s distinction between generic and improved machine-learning methodology in the US; the EPO’s “potential technical effect” and its two AI-specific technical-character routes; China’s four-category taxonomy and its dedicated disclosure and ethics rules. The overall picture is one of doctrinal convergence at the level of general method, but growing divergence at the level of AI-specific detail — precisely the pattern one would expect if the underlying general framework is serviceable but incomplete, rather than either wholly adequate or wholly obsolete. That observation motivates the analysis in Parts V and VI.

V. Is Artificial Intelligence Ontologically Different from Software?

Having seen that no major patent system has accepted a categorical distinction between trained models and conventional programs — and that the UK Supreme Court expressly rejected the High Court’s attempt to draw one — it is worth examining the distinction on its own technical and doctrinal merits, independent of what any court has so far concluded. This Part sets out the strongest version of the argument for a categorical distinction, the strongest version of the argument against it, and then stakes out an intermediate position that informs the framework proposed in Part VI.

A. The Case for a Categorical Distinction

The strongest argument for treating trained models as a distinct category rests on three technical observations, each with a doctrinal analogue. First, authorship. A conventional computer program is a sequence of instructions a human being wrote, line by line, in order to cause specific, intended behaviour. A trained neural network’s operative parameters are not written; they are the output of an optimisation process — typically stochastic gradient descent or a variant — applied to a loss function over a training dataset. The developer chooses the architecture, the loss function, the optimiser, and the data, but does not choose the resulting weights directly, any more than a plant breeder chooses the precise genome of a new cultivar by writing out its DNA sequence. This is a genuine, technically accurate distinction, and it is the distinction Sir Anthony Mann’s High Court judgment sought, unsuccessfully, to give legal weight.

Second, opacity. Even the developer of a trained model frequently cannot state, in the way a programmer can for conventional code, why the model produces a particular output for a particular input. This “black box” property is not a peripheral inconvenience; it goes to the heart of the enablement and sufficiency doctrines discussed in Part II, and the EPO’s Boards of Appeal have already had to grapple with it in the sufficiency case law discussed in Part IV. A legal framework built on the assumption that an invention’s operation can be explained step-by-step in a specification sits uneasily with a technology whose central characteristic is that it cannot be so explained, even by its own creator.

Third, the locus of inventive labour. As Part II observed, the genuinely inventive contribution in many machine-learning projects lies not in an algorithm’s explicit logical structure but in choices about data curation, training curricula, loss-function design, and architecture — choices whose value is demonstrated empirically, through evaluation on held-out data, rather than proven analytically, through logical derivation. Patent doctrine built around describing what a program does, step by step, is not naturally suited to protecting or even accurately describing this kind of contribution.

B. The Case Against

The case against a categorical distinction is, on close examination, at least as strong, and it is essentially the case the Court of Appeal and the Supreme Court accepted in Emotional Perception. At the level of physical execution, a trained neural network is nothing more than a fixed sequence of matrix multiplications and non-linear transformations, executed by a processor — general-purpose or specialised — operating within the same computational paradigm as any other software. Nothing about a trained model escapes the Church-Turing framework that underlies all digital computation. The fact that the weights were arrived at through training rather than direct authorship changes how the instructions came to exist, but not what the machine does when it executes them, and the Court of Appeal’s formalist point — that a computer is a machine that processes information, and a program is a set of instructions that causes it to do so — applies with equal force whether the instructions were typed by a programmer or learned by an optimiser.

This objection also has force as a matter of legal history and administrability. Complex rule-based expert systems, genetic algorithms, and even classical numerical solvers have long produced behaviour their authors could not fully predict in advance, without anyone suggesting they required a distinct doctrinal category. If unpredictability alone were sufficient to justify a new legal species, the line would have to be drawn somewhere within a continuum of algorithmic complexity that has existed since the earliest heuristic search programs of the 1960s, not at the introduction of gradient-based training in the 2010s. A rule that turned patentability on whether a claim’s author can fully explain the resulting behaviour would also invite drafting arbitrage: an applicant motivated to escape an exclusionary doctrine would have every incentive to characterise a hand-coded system as an opaque “learned” one, or vice versa, and patent offices have limited practical means of verifying such characterisations at the examination stage.

Finally, the “just software” framing itself is largely an artefact of the specific statutory language used in the UK and at the EPO — the “program for a computer … as such” exclusion under Article 52(2)(c) EPC and its UK equivalent. Neither the United States nor China excludes software as a category at all; both instead apply general subject-matter or technical-character tests under which software, including machine learning, is eligible in principle and is refused only where a specific claim is drawn too abstractly. For roughly two of the four jurisdictions surveyed in this paper, the entire premise of the debate — whether AI escapes a software-specific exclusion — does not even arise in the same form. That should give pause to any proposal that treats the European exclusionary architecture as the natural starting point for a global AI-specific framework.

C. An Intermediate Position

The better view, and the one this paper adopts, is that the categorical question — is an AI system a computer program, yes or no — is the wrong question, not because both sides lack force, but because the classification does not do the doctrinal work either side wants it to do. The Supreme Court’s approach in Emotional Perception is, on this point, more sophisticated than either the High Court’s or the Court of Appeal’s: by holding that classification as a computer program does not itself resolve patentability, and that the real inquiry concerns technical contribution, the Court effectively concluded that the fight over labels was a distraction from the fight that actually matters.

The more productive question, pursued in Part VI, is not whether AI systems are computer programs, but whether the type of technical contribution the patent system is designed to identify and reward has, for AI inventions specifically, migrated to places that existing tests are not well calibrated to find — training methodology, dataset design, architectural innovation, and the interaction between a learned model and a physical system — and whether the enablement doctrine can be adapted to accommodate the genuine, technology-driven opacity of trained models without becoming a loophole for functional claiming untethered to any disclosed technical means.

VI. Toward a Purpose-Built Analytical Framework for AI Inventions

If the categorical question is the wrong question, what should replace it? This Part develops the three diagnostic questions sketched in the debate that motivates this paper into a more complete, four-factor framework, designed to be workable within existing statutory architecture — that is, as an interpretive gloss on the existing technical-contribution and abstract-idea doctrines, rather than as a proposal for new legislation.

A. Design Criteria

A workable framework for AI patent eligibility should satisfy four design criteria drawn from the comparative survey in Part IV and the ontological analysis in Part V. It should be technology-neutral in form, so that it can be adopted as interpretive guidance within the United States’ abstract-idea doctrine, the UK/EPO technical-contribution methodology, and China’s technical-relationship inquiry, without requiring any of the three architectures to be abandoned. It should discriminate between genuine technical innovation in machine learning and the mere relabelling of a known technique, in the manner Recentive already does in the United States. It should accommodate the genuine opacity of trained models without collapsing into a rule that any claim reciting “train a model” automatically satisfies enablement. And it should be administrable by patent examiners working under ordinary time constraints, without requiring specialised machine-learning expertise beyond what many examining corps already possess in electronics and computer-science art units.

B. A Proposed Four-Question Test

1. The Technical-Problem/Trained-Model Nexus

Does the claimed invention solve a technical problem — in the ordinary patent-law sense of a reproducible, objective problem falling outside the excluded categories of mental acts, business methods, and purely aesthetic judgment — by means specifically enabled by a trained model, as opposed to merely automating a pre-existing human or business process behind an unspecified machine-learning “black box”? This question operationalises the Recentive distinction and should be asked regardless of jurisdiction: a claim that recites training a model to perform a task humans have long performed manually, without identifying any technical improvement in how the training or inference is carried out, should fail this threshold question.

2. The Locus of Inventive Contribution

Where, specifically, does the claimed invention’s contribution lie? A framework should require applicants and examiners to locate that contribution in one of three places: the training methodology itself (a novel loss function, curriculum, regularisation technique, or data-augmentation strategy that produces a demonstrably better or more efficient result than known methods); the resulting learned representation or architecture (a structural innovation not obvious from prior architectures, analogous to a novel device structure in conventional engineering patents); or a novel technical interaction between the trained model and a physical system, such as sensors, actuators, or a real-time control loop. A claim that identifies none of these — that recites only the application of an off-the-shelf model to a new domain — should not pass, whatever domain-specific language surrounds it.

3. Enablement Through Reproducibility, Not Explicability

Because a trained model’s decision path cannot, in general, be exhaustively explained even by its creator, sufficiency of disclosure should be assessed by reference to whether a skilled person, following the disclosed training methodology, dataset description, and evaluation protocol, could reproduce the claimed technical effect — not by reference to whether the specification explains the causal chain by which any individual output is produced. This tracks the direction the EPO’s Boards of Appeal have already taken in T161/18 and T1191/19, and this paper argues that direction should be made explicit, rather than left to accrete case-by-case as an unstated sufficiency risk factor. A specification that discloses training data characteristics, architecture, hyperparameters, and evaluation metrics sufficient for reproduction should satisfy this requirement even though it cannot, and need not, explain why the resulting model behaves as it does on any given input.

4. Learned Parameters as a Technical Artifact, Not Merely Data

A claim to a trained model itself — its architecture and parameters, or a computer-readable medium storing them — should be assessed under criteria analogous to those applied to product-by-process claims in chemistry and biotechnology, where a claimed product’s patentability can rest on a novel and non-obvious process used to obtain it, even though the product’s composition might not be distinguishable from a hypothetical product obtained differently. Treating learned parameters purely as “data,” categorically outside the scope of technical subject matter, undervalues the genuine technical achievement that a novel training process can represent; treating them purely as a conventional manufactured article, without regard to the process by which they were obtained, risks over-rewarding trivial applications of known training techniques to new datasets — exactly the pattern Recentive rejected.

C. Applying the Framework: Two Worked Examples

Applied to Emotional Perception’s own facts, the framework yields a more nuanced answer than either the High Court’s or the Court of Appeal’s binary conclusion. The invention’s claimed contribution — a specific contrastive training methodology that jointly aligns two independently trained networks within a shared property space, without relying on human-assigned genre labels — plausibly satisfies factor two (a specific training-methodology contribution) and arguably factor one, if the technical problem is characterised as improving the objective structure of a learned representation space rather than merely producing a subjectively preferable recommendation. Whether it in fact does so is an empirical and legal question the framework does not answer in the abstract; what the framework offers is a more precise vocabulary for asking it, in place of the blunter question of whether the output is a “technical” or merely “cognitive” improvement, which proved highly sensitive to how each court chose to characterise the invention’s output.

Applied to Recentive’s claims, the framework reaches the same result the Federal Circuit reached, but by an explicit route: the claims fail factor one, because they automate a task — event scheduling and network-map generation — that humans had long performed using the same categories of input data, and they fail factor two, because the specification disclosed the application of generic, previously known machine-learning training techniques to that new data environment, without identifying any innovation in the training methodology, architecture, or model-to-system interaction itself. The convergence between the framework’s analysis and the Federal Circuit’s actual holding, reached under conventional § 101 doctrine without any explicit AI-specific framework, is itself evidence that the framework is not proposing a wholesale departure from current outcomes, but rather making explicit the distinction current doctrine is already, if inconsistently, reaching for.

VII. Learning from History: How Patent Law Has Adapted Before

Patent law has confronted genuinely novel technological categories before, and its responses — sometimes through case-by-case adjudication, sometimes through bespoke legislation, and at least once through a permissive overcorrection later reversed — offer instructive guidance for how the framework proposed in Part VI should be implemented, and how it should not be.

A. Biotechnology: Case-by-Case Line-Drawing

When genetic engineering first produced patent applications for living organisms, the United States Supreme Court confronted the question directly in Diamond v. Chakrabarty. The Court held that a human-made, genetically modified bacterium capable of breaking down crude oil was eligible subject matter as a “manufacture” or “composition of matter” under 35 U.S.C. § 101, rejecting the argument that living things were categorically excluded from patentability, and emphasising instead that the relevant distinction was between products of nature and human-made inventions, “irrespective of whether they are living or inanimate.” That distinction has continued to generate difficult line-drawing decades later: in Association for Molecular Pathology v. Myriad Genetics, the Court held that a naturally occurring, merely isolated DNA sequence remains an unpatentable product of nature, while synthetically created complementary DNA, from which non-coding sequences have been removed, is patent-eligible because it does not occur in nature in that form.

Biotechnology’s lesson for AI patent law is that courts are capable of drawing durable, technically sensitive distinctions within a single statutory framework — “manufacture or composition of matter,” unchanged since 1952 — without requiring a new statutory category, provided the distinction tracks a genuine difference in human contribution. Chakrabarty and Myriad did not create a new title of intellectual property for biological inventions; they refined, through litigation, what counts as sufficient human intervention to convert an unpatentable natural product into a patentable human-made one. The four-factor framework proposed in Part VI pursues an analogous strategy: rather than creating a new exclusion or a new subject-matter category for AI, it refines, within the existing technical-contribution and abstract-idea doctrines, what counts as a sufficient and identifiable human (or human-directed) technical contribution to a trained system.

B. Semiconductor Chip Topography: Bespoke Legislation

Not every mismatch between existing IP doctrine and a new technology has been resolved through case law. When integrated-circuit manufacturers sought protection for the three-dimensional layout, or topography, of semiconductor chips in the early 1980s, neither patent law (which offered no protection for a layout not itself meeting the novelty and non-obviousness thresholds applicable to a functional circuit design) nor copyright law (which was a poor fit for a largely functional, industrially mass-produced layout) provided an adequate remedy. Congress responded with the Semiconductor Chip Protection Act of 1984, codified at 17 U.S.C. §§ 901–914, creating a genuinely new, sui generis form of intellectual property — protection for “mask works” — with its own term, its own registration system, and its own infringement standard tailored to the technology’s characteristic form of copying (photographic reproduction of a chip layout) and characteristic form of legitimate reuse (reverse engineering for the purpose of designing a compatible but non-identical chip).

The semiconductor precedent shows that legislatures are capable of creating bespoke intellectual-property regimes when neither of the existing boxes fits, and that doing so can be a clean, comparatively fast solution once the technology’s characteristic features are well enough understood to be codified. It also illustrates the cost of that approach: mask-work protection has remained a narrow, jurisdiction-specific regime, adopted unevenly around the world, and its bespoke infringement and reverse-engineering rules required their own decades of interpretive litigation. A legislative sui generis regime for AI-generated technical content — model weights, architectures, or training methodologies — is not inconceivable, and some commentators have floated the idea in other contexts (for example, database protection in the EU), but the comparative survey in Part IV suggests the current international appetite for such a step, in patent law specifically, remains low; no major jurisdiction has proposed anything resembling a mask-work-style regime for trained models, and the more realistic near-term path, discussed in Part VIII, runs through administrative guidance and interpretive doctrine rather than new legislation.

C. Business Methods: The Cautionary Tale of Overcorrection

The most cautionary historical analogue is the United States’ experience with business-method patents. In State Street Bank & Trust Co. v. Signature Financial Group, the Federal Circuit held, in 1998, that a data-processing system for administering a mutual-fund investment structure was patent-eligible because it produced a “useful, concrete, and tangible result,” a standard widely understood at the time as opening the door broadly to business-method and financial-software patents. The following decade saw a substantial increase in low-quality business-method patent filings and litigation, much of it associated with so-called patent-assertion entities asserting broadly drafted claims against operating companies. The Supreme Court corrected course, first in Bilski v. Kappos, rejecting the “useful, concrete, and tangible result” test as the sole measure of eligibility and reaffirming the centrality of the abstract-idea exclusion, and decisively in Alice Corp. v. CLS Bank International, which extended the Mayo two-step framework to abstract ideas generally and precipitated a sharp increase in patents invalidated at the pleading stage.

The lesson for AI patent policy is not that permissiveness is always wrong, but that a doctrinal shift toward greater eligibility, adopted without corresponding rigour at the obviousness and disclosure stages, can generate a period of low-quality patenting that later requires a disruptive judicial correction — imposing real costs, in the interim, on the operating companies and follow-on innovators who must litigate against overbroad claims. This is directly relevant to the framework proposed in Part VI, and to the Supreme Court’s own move in Emotional Perception toward a more permissive threshold eligibility test: any relaxation of the threshold “is this an invention” inquiry for AI, whether accomplished through the UK/EPO’s technical-contribution realignment or through a purpose-built framework, must be paired with correspondingly rigorous scrutiny of novelty, obviousness, and enablement at the examination stage, precisely because the threshold gate is being made easier to pass. Recentive itself models this discipline reasonably well, resolving a permissive-sounding technology (machine learning) restrictively at the threshold stage by insisting on a genuine, disclosed technical improvement — the opposite sequencing from State Street, and one this paper argues other jurisdictions should study.

D. Synthesis

Three lessons emerge. First, patent systems are institutionally capable of adapting to genuinely novel technology, whether through incremental case-by-case line-drawing, as in biotechnology, or through deliberate legislative design, as with semiconductor mask works; the absence of either response to date for AI reflects a policy choice, or a lack of consensus, rather than an institutional incapacity. Second, case-by-case adaptation, while slower and messier than legislation, has proved durable and has not required abandoning the underlying statutory categories — “manufacture or composition of matter” in the United States has absorbed biotechnology without amendment for over seventy years. Third, and most importantly for the framework proposed in this paper, the direction of doctrinal error matters: the business-method episode demonstrates that a framework calibrated to be too permissive can do as much damage, through patent-quality erosion and litigation cost, as one calibrated to be too restrictive. Any AI-specific framework, including the one proposed in Part VI, should therefore be judged not only on whether it correctly admits genuinely inventive AI systems, but on whether it is rigorous enough to exclude the merely generic application of known techniques that Recentive shows courts are already, correctly, refusing to reward.

VIII. Institutional and Policy Considerations

A. The Case for Change

Several considerations support moving toward the more explicit, purpose-built framework proposed in Part VI, rather than continuing to rely on ad hoc accretion of case law and examiner practice. Patent offices are already, in effect, improvising AI-specific doctrine — the EPO’s G-II 3.3.1 guidelines, the Federal Circuit’s Recentive analysis, and CNIPA’s four-category taxonomy are each, in their own way, informal recognitions that the general computer-implemented-invention framework needs AI-specific glosses to function predictably. Making that recognition explicit, through a shared analytical vocabulary of the kind proposed in Part VI, would reduce the inconsistency visible in the fact that a single set of facts — Emotional Perception’s recommendation system — produced three different outcomes across three UK tribunals over roughly three years, imposing substantial cost and delay on the applicant and creating years of uncertainty for the wider industry watching the case. Greater predictability would disproportionately benefit smaller AI developers and research-derived startups, which typically lack the in-house patent expertise and litigation budgets that larger, incumbent developers can deploy to navigate uncertain doctrine or to prevail in multi-year appeals of the kind Emotional Perception required.

B. The Risks

The business-method history canvassed in Part VII counsels real caution. A framework that makes AI eligibility easier to establish, without disciplined application of factors two through four, risks reproducing the low-quality patenting problem of the late 1990s and 2000s, this time directed at foundational machine-learning techniques — attention mechanisms, diffusion processes, particular training objectives — that are, in current practice, widely shared as open architectures across the research community. Because much of the modern AI ecosystem depends on a comparatively small number of foundational techniques being freely available for downstream application, aggressive patenting at that foundational layer could concentrate defensive and offensive patent leverage in the hands of the large, well-resourced organisations best able to file broadly and early, echoing patent-thicket concerns that arose in software and, earlier still, in the semiconductor industry before cross-licensing norms matured.

A second risk concerns the interaction between the fourth framework factor — treating learned parameters as a technical artifact — and the separate, and largely unresolved, legal questions surrounding the data used to train a model, including copyright in training data and data-protection law governing personal information within a training set. Recognising trained parameters as a patentable technical artifact does not, and should not, resolve whether the process of creating them was itself lawful; CNIPA’s newly adopted legality and ethics screen is, in this respect, addressing a genuine gap that the UK, EPO, and US frameworks currently leave open, and other jurisdictions should consider whether an analogous screen, however administered, is warranted.

A third risk is more technical: relaxing enablement requirements to accommodate the genuine opacity of trained models, as factor three of the proposed framework does, could become a vehicle for impermissible functional claiming — claims that recite a desired result (“a model that predicts X with accuracy exceeding Y”) without disclosing any specific means of achieving it. Both the EPO’s Article 84/83 clarity and sufficiency requirements and the United States’ § 112(f) functional-claiming doctrine already police this risk in conventional software patenting, and the framework proposed here does not seek to displace them; the reproducibility standard set out in factor three is meant to operate alongside, not instead of, those existing safeguards, and patent offices adopting it should make that relationship explicit to avoid the risk being read as an invitation to functional claiming.

C. Who Should Decide?

Three institutional pathways are available, and they are not mutually exclusive. The first is continued incremental development through case law and administrative guidelines — the status quo, and the path each of the four jurisdictions surveyed in Part IV is currently on. Its principal advantage is flexibility: guidelines and Board of Appeal decisions can be revised more quickly than statutes, as the EPO’s repeated revisions to G-II 3.3.1 and CNIPA’s three rounds of guideline amendment between 2023 and 2025 demonstrate. Its principal disadvantage is the inconsistency and delay Emotional Perception itself illustrates, and the risk that different jurisdictions drift further apart rather than converging, absent any coordinating mechanism.

The second pathway is legislative reform, on the semiconductor model. This is more difficult in practice than the case-by-case alternative: proposals to reform § 101 in the United States have circulated in Congress for over a decade without enactment, reflecting genuine and unresolved disagreement among stakeholders about where the eligibility line should sit, not merely legislative inertia. A legislative approach also risks freezing a framework in statutory language at a moment when the underlying technology, and the community’s understanding of where its inventive content actually lies, may still be evolving rapidly — a concern that did not apply in the same way to the comparatively mature and stable semiconductor manufacturing processes Congress addressed in 1984.

The third pathway is international harmonisation through the World Intellectual Property Organization, which has convened an ongoing WIPO Conversation on Intellectual Property and Frontier Technologies to address exactly these questions on a multilateral basis. Given that AI research, AI patent filing, and AI litigation are all substantially transnational — as the comparative survey in Part IV itself demonstrates, with the same underlying technologies raising materially similar questions in London, Washington, Munich, and Beijing within the same few years — international coordination has an obvious appeal in principle. Its practical record so far, however, counsels modest expectations: the WIPO Conversation has not yet produced binding standards, and historical experience with multilateral IP harmonisation, including the decades required to negotiate and implement the TRIPS Agreement, suggests any binding international AI-specific patent standard remains a long-term prospect rather than a near-term solution to the problems identified in this paper.

On balance, this paper’s institutional recommendation, developed further in Part IX, favours the first pathway in the near term — adoption of an explicit, shared analytical framework through administrative guidance and interpretive case law, in the manner the EPO and CNIPA have already begun — while treating legislative and international harmonisation as longer-term complements rather than near-term substitutes.

IX. Recommendations

Drawing together the comparative analysis in Part IV, the ontological assessment in Part V, the proposed framework in Part VI, the historical lessons in Part VII, and the institutional analysis in Part VIII, this paper makes five concrete recommendations.

First, patent offices should formally adopt an explicit, shared analytical framework for AI inventions — along the lines of the four factors proposed in Part VI — as interpretive guidance within their existing statutory architecture, rather than waiting for further piecemeal case law to accumulate. The EPO’s G-II 3.3.1 revisions and CNIPA’s dedicated AI guidelines show this is administratively feasible without new legislation; the UK, having just undertaken a broader methodological realignment in Emotional Perception, is well placed to issue analogous AI-specific guidance building on that realignment; and the USPTO, having already issued examiner guidance following Recentive, should consider formalising the generic-versus-improved-technique distinction that decision establishes into guidance applicable across the machine-learning docket generally, not solely to scheduling and recommendation-system claims.

Second, patent offices should adopt explicit disclosure standards for machine-learning patents addressing training-data characteristics, training methodology, and evaluation protocol, calibrated to the reproducibility standard set out in framework factor three, rather than leaving sufficiency of disclosure for opaque models to be resolved through untethered, case-by-case objections of the kind seen so far only in isolated EPO Board of Appeal decisions.

Third, examiners and courts should apply heightened scrutiny, at the obviousness or inventive-step stage, to claims that recite only the application of an established machine-learning technique to a new data environment or industry vertical, following the model Recentive Analytics v. Fox Corp. already provides in the United States. This recommendation is deliberately restrictive in orientation, consistent with the caution drawn from the business-method history in Part VII: a more permissive eligibility threshold for AI, of the kind Emotional Perception introduces in the UK, should be paired with, not substituted for, rigorous downstream scrutiny.

Fourth, patent offices should monitor patent-thicket risk at the level of foundational AI architectures and training techniques, and should consider, as a first line of response, more aggressive use of prior-art search tools tailored to machine-learning literature and heightened obviousness scrutiny, before considering more structural interventions such as compulsory licensing or standard-essential-patent-style FRAND commitments for architectures that become de facto industry standards — tools that exist within current law and do not require new legislative categories.

Fifth, jurisdictions that have not yet done so should consider whether an explicit legality or data-provenance screen, of the kind CNIPA adopted in its November 2025 amendments, is warranted, recognising that the propriety of a model’s training data is a policy question distinct from, but increasingly entangled with, the technical-contribution inquiry this paper has otherwise focused on.

X. Conclusion

The question posed at the outset of this paper — is it time to stop treating artificial intelligence as “just software” — does not admit of a simple yes-or-no answer, and the UK Supreme Court’s judgment in Emotional Perception, read carefully rather than through the lens of its press coverage, illustrates why. The Court did not hold that a trained neural network escapes classification as a computer program; it held the opposite, while using the occasion to dismantle a stricter, UK-specific eligibility test and replace it with a more permissive, European-aligned methodology. The result is a significant development for AI patenting in the United Kingdom, but not the categorical reclassification of AI as a distinct technological species that some commentary has suggested.

That does not mean patent law’s inherited software paradigm is fit for purpose as currently applied. The comparative survey in this paper shows every major patent system quietly developing AI-specific glosses on general doctrine — the Federal Circuit’s generic-versus-improved-technique distinction in Recentive, the EPO’s potential-technical-effect doctrine and its two AI-specific technical-character routes, CNIPA’s four-category taxonomy and legality screen — without any of them yet assembling those glosses into a coherent, shared analytical framework. The four-factor framework proposed in Part VI is an attempt at that assembly: not a proposal to exempt AI from the ordinary discipline of patent-eligibility doctrine, but a proposal to make explicit, and consistently applicable, the distinction between genuine technical innovation in how a model is trained, structured, or connected to the physical world, and the merely generic application of known machine-learning techniques to a new problem — a distinction current doctrine is already reaching for, unevenly, across four jurisdictions.

The deeper lesson of both the Emotional Perception litigation and the comparative material surveyed here is that patent law has adapted to genuinely novel technology before, through both incremental case-by-case development and, where necessary, bespoke legislation, and that the principal risk in doing so lies not in choosing the wrong institutional pathway, but in miscalibrating the balance between eligibility and rigour — as the business-method episode of the late 1990s and early 2000s demonstrates. Artificial intelligence may or may not eventually warrant a legal category as distinct as the one Congress created for semiconductor topography in 1984. What it warrants now, on the evidence assembled in this paper, is not a rhetorical declaration that AI is no longer “just software,” but the more modest and more difficult work of adjusting enablement, obviousness, and technical-contribution doctrine to track where AI’s genuine inventive content actually lives — work that four patent systems have already begun, unevenly and without coordination, and that this paper has attempted to describe as a single, coherent project.

References

[1]  Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks [2026] UKSC 3 (11 February 2026).

[2]  Emotional Perception AI Ltd v Comptroller-General of Patents, Designs and Trade Marks [2023] EWHC 2948 (Ch) (Sir Anthony Mann).

[3]  Comptroller-General of Patents, Designs and Trade Marks v Emotional Perception AI Ltd [2024] EWCA Civ 825 (Birss LJ).

[4]  UK Supreme Court, Case Summary UKSC-2024-0131, Emotional Perception AI Limited (Appellant) v Comptroller General of Patents, Designs and Trade Marks (Respondent).

[5]  CMS Law, “Emotional perception reaches the top: Supreme Court delivers landmark ruling on AI patentability” (2026).

[6]  Wilson Gunn, “Emotional Perception AI v Comptroller: The UK Supreme Court Reshapes UK Patent Law” (2026).

[7]  HLK, “UK Supreme Court hands down landmark Emotional Perception AI decision” (2026).

[8]  Linklaters Digilinks, “Emotional Perception AI: Supreme Court overhauls decades of UK case law on patentability of computer-implemented inventions” (2026).

[9]  Source Advisors, “Supreme Court ruling kickstarts a new era for software patentability in the UK” (2026).

[10]  Stephenson Harwood, “Computer implemented inventions: Comptroller General of Patents v Emotional Perception AI Limited” (2024).

[11]  Mathys & Squire LLP, “Emotional Perception High Court ruling overturned by Court of Appeal” (2024).

[12]  Recentive Analytics, Inc. v. Fox Corp., 134 F.4th 1205 (Fed. Cir. 2025).

[13]  Mintz, “Recentive Analytics v. Fox: The Federal Circuit Provides Analysis on the Patent Eligibility of Machine Learning Claims” (2025).

[14]  Baker Botts, “Federal Circuit Refines Section 101 Eligibility of Machine Learning Inventions” (2025).

[15]  Fish & Richardson, “Federal Circuit Clarifies Limits of Patent Eligibility for Machine Learning Claims” (2025).

[16]  Nixon Peabody LLP, “Federal Circuit limits patent eligibility of machine learning in Recentive Analytics v. Fox” (2025).

[17]  Greenberg Traurig LLP, “Federal Circuit: Machine Learning Patents Ineligible in Recentive Analytics, Inc. v. Fox Corp.” (2025).

[18]  Bracewell LLP, “Recentive v. Fox: Machine-Learning Claims Fail to Make the Grade” (2025).

[19]  Sterne Kessler, “2025 Federal Circuit IP Appeals: Recentive Analytics, Inc. v. Fox Corp.” (2026).

[20]  Ryan Abbott, “Reflections on Recentive v. Fox: To Do or not To Do it with AI?”, Patently-O (2025).

[21]  Petition for a Writ of Certiorari, Recentive Analytics, Inc. v. Fox Corp., No. 25-505 (U.S. Sup. Ct., filed Oct. 21, 2025).

[22]  European Patent Office, Enlarged Board of Appeal, G 1/19 (Mar. 2021).

[23]  IAM Media, “How revised EPO guidelines affect treatment of AI inventions” (2022).

[24]  Legal Patent (MD Legal European Patent Attorneys), “AI patent application: Current EPO requirements” (2026).

[25]  Reddie & Grose, “Changes to the European Patent Office Guidelines for Assessing the Patentability of Computer Implemented Inventions following G1/19” (2022).

[26]  Venner Shipley, “G1/19 – what does it mean for AI?” (2025).

[27]  HLK, “Revised EPO Guidelines: AI and Computer-Implemented Inventions – Emphasis on COMVIK and Technical Detail” (2023).

[28]  epi Information, “G 1/19 released: The Enlarged Board of Appeal decides on the Patenting of Computer-implemented Simulations and Designs” (2021).

[29]  Venner Shipley / Lexology, “Navigating AI Patents: key updates from EPO and UKIPO Guidelines” (2025).

[30]  European Patent Office, Guidelines for Examination, s. G-II, 3.3 & 3.3.1 (as revised).

[31]  CNIPA, Guidelines for Patent Examination (amended), effective 20 January 2024.

[32]  Spruson & Ferguson / Lexology, “China: Patent Guidelines for Examination – changes for AI, big data and software” (2024).

[33]  CNIPA, Guidelines for Patent Applications for AI-Related Inventions (Trial Implementation), issued 31 December 2024.

[34]  Hogan Lovells, “Navigating AI patent applications in China: Key insights from CNIPA’s new examination guidelines” (2025).

[35]  China IP Law Update, “CNIPA Releases Draft Guidelines for Patent Applications for Artificial Intelligence-Related Inventions” (2024).

[36]  China IP Law Update, “China’s National Intellectual Property Administration Issues Guidelines for Patent Applications for AI-Related Inventions” (2025).

[37]  China IP Law Update, “China’s National Intellectual Property Administration Releases Revised Guidelines for Patent Examination Effective January 1, 2026” (2025).

[38]  IAM, “CNIPA updates patent examination guidelines to address emerging technology, regulate dual filing and invalidations” (2025).

[39]  Japan Patent Office & CNIPA, Comparative Study on AI-Related Inventions Report 2023.

[40]  Diamond v. Chakrabarty, 447 U.S. 303 (1980).

[41]  Association for Molecular Pathology v. Myriad Genetics, Inc., 569 U.S. 576 (2013).

[42]  Gottschalk v. Benson, 409 U.S. 63 (1972).

[43]  Diamond v. Diehr, 450 U.S. 175 (1981).

[44]  Mayo Collaborative Services v. Prometheus Laboratories, Inc., 566 U.S. 66 (2012).

[45]  Alice Corp. v. CLS Bank International, 573 U.S. 208 (2014).

[46]  Bilski v. Kappos, 561 U.S. 593 (2010).

[47]  State Street Bank & Trust Co. v. Signature Financial Group, Inc., 149 F.3d 1368 (Fed. Cir. 1998).

[48]  Semiconductor Chip Protection Act of 1984, 17 U.S.C. §§ 901–914.

[49]  Aerotel Ltd v Telco Holdings Ltd [2006] EWCA Civ 1371.

[50]  Patents Act 1977 (UK), s. 1(2)(c); European Patent Convention, art. 52(2)(c), (3).

[51]  United States Patent and Trademark Office, 2019 Revised Patent Subject Matter Eligibility Guidance, 84 Fed. Reg. 50 (Jan. 7, 2019).

[52]  United States Patent and Trademark Office, Examiner Guidance Memorandum on Machine Learning Claims Under 35 U.S.C. § 101 (Aug. 4, 2025).

[53]  World Intellectual Property Organization, WIPO Conversation on Intellectual Property and Frontier Technologies (ongoing).