The UK Supreme Court’s decision in Emotional Perception AI Limited v Comptroller General of Patents marks a doctrinal shift in how subject-matter eligibility questions regarding computer programmes are evaluated. Irrespective of its impact, Pragati argues that the judgment is marked by a test that the Supreme Court itself chose not to apply and further contains three errors of reasoning. Pragati Upadhyay is a final-year student at the Faculty of Law, Banaras Hindu University, with a research focus on the intersection of emerging technologies and IPR.
The UKSC’s “Any Hardware” Shortcut: Why Emotional Perception AI Raises More Questions Than It Answers — And Why India’s Framework Does Better
Pragati Upadhyay
In September 2024, Bharathwaj Ramakrishnan wrote on this blog examining how the UK Court of Appeal in Emotional Perception AI and the EPO’s Board of Appeal in Mitsubishi (T0702/20) (see here) had diverged on the legal nature of Artificial Neural Networks, one classifying them as computer programmes, the other as mathematical methods implemented on a computer. He noted that these diverging positions matter for India, given the pervasive influence of UK and EPO jurisprudence on Section 3(k) of the Patents Act, 1970 (see Ferid Allani v. Union of India for the Delhi High Court’s express acknowledgment of the similarity between Article 52 EPC and Section 3(k) Indian Patent Act, 1970).
On 11 February 2026, the UK Supreme Court closed that chapter and opened a more troubling one. In Emotional Perception AI Limited v Comptroller General of Patents [2026] UKSC 3 (the summary can be accessed here), a unanimous bench formally overruled Aerotel and adopted the EPO’s “any hardware” approach. The judgment has been widely welcomed as a step toward harmonisation, with major IP practices including Linklaters, DLA Piper, and specialist outlet Juve Patent praising it for delivering certainty and commercial clarity for AI innovators. I argue the opposite: that in its rush to harmonise, the UKSC has committed three compounding doctrinal errors — and that India’s approach, far from needing to follow suit, is structurally more honest about what the law is trying to do.
What the Court Did (and What It Admitted It Could Not Do)
A brief summary for context. Emotional Perception AI’s (EPAI) invention used an ANN to recommend media files based on their physical properties rather than human-assigned genre tags. The UKIPO rejected it. The High Court allowed it. The Court of Appeal rejected it again. The UKSC allowed the appeal and in doing so, formally overruled the Aerotel test and introduced a new two-stage framework: first, a low “any hardware” threshold to clear the eligibility gate, followed by an “intermediate step” requiring a feature-by-feature analysis of which claim elements contribute to the invention’s technical character before novelty and inventive step are assessed. Then did something striking, it remitted the case back to the UKIPO without itself applying the very framework it had just created, because, as the judgment candidly states, the Court felt it would be “unwise to embark on carrying out the analysis required by the intermediate step … without the benefit of both in-depth legal argument and the views of the UKIPO” (see para 117 of the judgment here).
Read that again. The Court introduced a new framework, applied it to the very case that prompted the framework, and then found itself unable to demonstrate how the framework actually operates on those facts. The Court was not remitting because the facts were incomplete or because a lower court needed to make findings of credibility. It was remitting because it could not demonstrate, on the facts of the case that had travelled all the way to the UK’s highest court over four years of litigation, how its own new analytical framework actually operates in practice. A court that introduces a framework but cannot show it working is not delivering legal clarity — it is deferring the hard question while appearing to answer it. That sentence, buried near the end of the judgment, is the most legally significant thing the UKSC said, and it has received almost no commentary.
Error One: Converting a Categorical Exclusion Into a Formal Threshold
The UKSC’s adoption of the “any hardware” test rests on a specific reading of the phrase “as such” in Section 1(2) of the Patents Act 1977 (which lists categories of subject matter that do not qualify as “inventions” for patent purposes, provides that the exclusions operate “only to the extent that” the application relates to the excluded subject matter “as such”) and Article 52(3) EPC (the equivalent provision in the European Patent Convention, which qualifies each exclusion with the phrase “as such”). The Court accepts the EPO’s long-standing position that “as such” means that only purely abstract subject matter is excluded — once any technical means is referenced in the claim, the exclusion does not bite. The Court acknowledges this is a “very low hurdle” (see para 99 of the judgment here) but justifies it precisely by reference to “as such.”
The problem is that this reading empties the enumeration in Section 1(2) of independent meaning. Parliament specifically listed computer programmes alongside mathematical methods and mental acts as excluded categories. If “as such” means merely “purely abstract,” then a purely abstract computer programme is already excluded as a mathematical method or mental act, the separate enumeration of computer programmes becomes redundant surplusage. A court’s primary obligation is to give effect to every word of a statute; reading “as such” so narrowly that the computer programme exclusion can never independently operate in the real world does the opposite precisely. It converts what Parliament designed as a categorical exclusion, a class of subject matter that should not be monopolised regardless of novelty, into a formal threshold that no real application will ever fail to clear.
Error Two: A Filter Without a Methodology
Aware that a low threshold needs a substantive downstream filter, the Court introduced the “intermediate step”: before assessing novelty and inventive step, the Patent Office must conduct a feature-by-feature analysis identifying which claim elements contribute to the invention’s technical character, assessed holistically. Non-technical features are filtered out.
This sounds workable. But the Court then explicitly declined to specify any methodology for performing this step, noting only that “any appropriate method” may be used, citing the Pozzoli inventive concept analysis (see judgment here) as one possible example. The EPO performs an equivalent analysis through its problem-solution approach, which the UKSC declined to import. The result is that the intermediate step is an instruction without a procedure: examiners are told what to identify, but given no tool for how to identify it. This concern is not merely theoretical — technical commentary has similarly noted that the judgment leaves examiners with a clear doctrinal framework but virtually no operative guidance on its application to ANN-based inventions (see Emanuilov’s Blog on Kluwer Patent Blog).
This is not merely an administrative inconvenience. For the period between now and whenever the UKIPO issues practice guidance, and a subsequent court defines the methodology, which could take years, the only operative rule at the UKIPO is: mention hardware, pass the gate. The filter exists on paper but not in practice. The Court’s own remittal in the very case that generated the framework is the clearest possible evidence of this: if the UKSC, with full adversarial argument, could not demonstrate the intermediate step on EPAI’s claims, it is difficult to see how UKIPO examiners will do so consistently and correctly in the interim.
Error Three: Uniform Application to a Non-Uniform Problem
This is the most technically precise of the three errors, and the one that has received the least attention. The intermediate step presupposes that an examiner can trace which claimed features produce which technical effects. For conventional software, explicit algorithms, if-then logic, and defined computational steps, this is a demanding but feasible analysis. The claim describes what the software does, and the examiner assesses whether what it does is technical.
ANN-based inventions are structurally different, not as a matter of legal category, but as a matter of how technical character is generated. An ANN claim describes an architecture and a training methodology. These are the code-level specifications. But the actual technical character of the trained network, the specific outputs it produces, the specific problems it solves, emerges from the trained weights: billions of numerical parameters generated by exposure to training data, which no human specified and which do not appear in any claim. The weights are the product of the code operationalising, not the code itself.
The intermediate step, applied to an ANN claim, therefore analyses the recipe while the technical character lives in the dish. An examiner performing a feature-by-feature technical character analysis of EPAI’s claims would examine the dual-ANN architecture and the training methodology, and would have no claim-level access to the weight configurations that determine whether the system actually produces the asserted technical effect of cross-modal emotional similarity inference. The uniform application of the intermediate step to ANN inventions is not merely methodologically underspecified, it is applied to the wrong layer of the invention by design.
The UKSC’s remittal, once more, speaks for itself. The Court could not perform the intermediate step on an ANN invention, not because of insufficient argument, but because the intermediate step, as designed, cannot be demonstrated on a class of invention whose technical character is post-training and unclaimed.
Why India’s Framework Is More Honest
India has not adopted the “any hardware” threshold, and the 2025 CRI Guidelines, issued by the Indian Patent Office on 29 July 2025, confirm that it should not. Under the Indian Patent Act, 1970, the Section 3(k) inquiry front-loads the technical effect analysis: the applicant must demonstrate, at the eligibility stage itself, that the invention produces a concrete technical advancement beyond what the underlying software or algorithm does in the abstract. As the Delhi High Court clarified in Raytheon Company v. Controller General of Patents (2023), novel hardware is not required, but a real technical effect is, and it must be shown at the threshold, not assumed and deferred.
This is, paradoxically, a better framework for ANN inventions precisely because it asks the right question at the right stage. Rather than performing a feature-by-feature analysis of claimed elements, which, for ANNs, is an analysis of the recipe rather than the dish, India asks holistically: does this invention, as it actually operates, produce a technical advancement? That question can be answered by examining outputs, by reviewing the training methodology’s effect on real-world performance, and by assessing the concrete problem solved. It does not require tracing technical character contribution through unclaimed weight configurations.
India’s framework is not more restrictive than the UKSC’s, it is rather more structurally honest about where the technical character of ANN inventions actually resides, and it builds its inquiry around that reality rather than around a claim-level analysis that the nature of the invention defeats.
Conclusion: Harmonisation Is Not Always Progress
The UKSC’s ruling in Emotional Perception AI [2026] UKSC 3 is significant, but significance is not the same as correctness. By converting a categorical Parliamentary exclusion into a formal threshold, introducing a filter without a methodology, and applying that filter uniformly to a class of invention it structurally cannot analyse, the Court has traded doctrinal integrity for the appearance of alignment. As Shivam Kaushik observed in a different article but resonant context on this blog, there is a risk in letting policy preferences lead and legal analysis follow, in treating hard doctrinal questions as inconvenient noumena while attending only to the observable phenomena of harmonisation.