Patenting AI after Emotional Perception AI [2026] UKSC 3

In Emotional Perception AI Limited v Comptroller General of Patents, Designs and Trade Marks (UKSC/2024/0131) the Supreme Court has reset the UK’s approach to the “computer program … as such” exclusion, overruling the long‑used Aerotel methodology and aligning UK interpretation with the European Patent Office’s reading of article 52 EPC following G1/19.

The court held that artificial neural networks are “programs for computers”, but that the claims here were not for a program “as such” because they involved technical means and therefore cleared the (very low) invention hurdle; the refusal was set aside and the case returned to the UKIPO for novelty and inventive step. The practical message is twofold: eligibility objections should be framed with the EPO’s staged analysis in mind, while the real fight moves downstream to which features “contribute to the technical character” and can support inventive step.

Facts and procedural history

The appeal concerned a patent application for a recommendation system using an artificial neural network (“ANN”). The claimed method trains an ANN using pairs of media files assessed in two different ways: (i) semantic similarity based on human textual descriptions (converted into vectors), and (ii) similarity based on measurable physical properties (also converted into vectors). The “trick” is to train the ANN so that distances in “property space” align more closely with distances in “semantic space”, then use the trained model to recommend items by proximity in property vectors without semantic comparisons at runtime. The judgment provides a clear, practitioner‑friendly explanation of ANN training, freezing of weights and biases, and subsequent use.

The application failed at the UK Intellectual Property Office on the basis that it fell within section 1(2)(c) of the Patents Act 1977 (the exclusion for “programs for computers … as such”). On appeal, the High Court of Justice allowed the application to proceed, holding (in substance) that an ANN is not a computer program at all and that, even if it were, the claims were not for a program “as such”. The Court of Appeal of England and Wales reversed that, restoring the refusal. The Supreme Court allowed the applicant’s appeal, set aside the hearing officer’s decision, and left the remaining patentability questions (novelty and inventive step) to be considered by the UKIPO.

Two interveners appeared: Chartered Institute of Patent Attorneys and IP Federation. Their participation reflects the wider significance recognised by the court: this was not treated as a niche point about one AI claim set, but as an issue of “fundamental importance” for UK treatment of computer‑implemented inventions, “particularly in the rapidly expanding field of AI”.

Legal framework

The statutory starting point is section 1 of the Patents Act 1977, whose structure is intended to track article 52 of the European Patent Convention (“EPC”). The court placed emphasis on section 130(7), which states that key provisions of the UK statute “are so framed as to have, as nearly as practicable, the same effects” as the corresponding EPC provisions. In consequence, the Supreme Court said it would “work directly from the underlying source”, namely article 52 EPC, rather than focus on minor textual differences between the two instruments.

Article 52(2) lists excluded matter including “programs for computers”, while article 52(3) limits that exclusion to the extent a claim relates to such subject matter “as such”. The historic UK approach to this territory has been dominated by the four‑step test in Aerotel, built around identifying the “contribution” and then asking whether it falls solely within excluded matter. The Supreme Court’s key move was to treat that “contribution” focus as having been decisively undermined by the European Patent Office’s post‑Aerotel jurisprudence, culminating in the Enlarged Board decision G1/19.

The court also rehearsed the constitutional and practical context: UK courts are not formally bound by EPO decisions, but policy and comity usually call for alignment so far as possible. The Supreme Court summarised the “authoritative dicta” as supporting a general rule that UK tribunals should respect and follow Enlarged Board decisions and uniform EPO jurisprudence unless convinced those decisions are wrong or beyond reasonable disagreement. At the same time, the judgment drew a sharp line between aligning on interpretation and copying EPO practice: there may be more than one method of implementing the Convention once meaning is fixed, and (on the court’s view) adoption of the EPO’s interpretation of article 52 did not automatically force the UK to adopt the EPO’s “problem and solution” approach to inventive step.

Court’s reasoning and holdings

Aerotel is overruled, but not replaced with a blank cheque

The court’s first issue was whether Aerotel should continue to be followed. The strongest reason for rejecting it was straightforward: the Enlarged Board in G1/19 had “emphatically rejected” the methodology as incompatible with the EPC, including endorsing the “any hardware” approach as the necessary first stage. The Supreme Court rejected the argument that disruption at the UKIPO justified retaining an “idiosyncratic approach” inconsistent with established EPO jurisprudence.

The judgment’s doctrinal core is that Aerotel wrongly entangles the threshold “invention” inquiry with novelty and inventive step. The Supreme Court accepted the EPO critique that a “contribution” test risks treating “invention” as if it must already be new or inventive. By contrast, the EPO approach asks first whether there is an invention at all (a technical character threshold), leaving novelty and inventiveness for later stages. The court distilled its conclusion sharply: “steps 2 to 4 cannot be maintained” (para 64).

What replaces Aerotel is not a fully‑fleshed domestic code. The Supreme Court adopted the Enlarged Board’s interpretation of article 52 (including the “any hardware” first hurdle), but repeatedly stressed that the court had not received submissions on how downstream novelty/inventive step analyses should be re‑tooled in the UK, and it was not prepared to create binding precedent on unargued detail. The judgment therefore gives structure and signposts, rather than a complete practitioner manual.

An ANN is a computer program

On issue 2, the court rejected the applicant’s attempt to exclude ANNs from the “program for computers” category by drawing a line between a “hardware ANN” (physical circuitry) and an “emulated” ANN implemented through software on a conventional computer. The court endorsed the Court of Appeal’s functional definitions of a computer and a program, and treated the ANN’s adjustable parameters (and not only those parameters) as instructions for processing information. In language likely to be quoted back in future eligibility disputes, the Supreme Court concluded: “Viewed as a whole, an ANN is a program for a computer” (para 88).

The court also made clear what will not, in its view, support carve‑outs for modern AI systems. It held that it is irrelevant that some ANNs can be “baked” into hardware or fixed into permanent form; the program/processor distinction remains meaningful whether instructions are changeable or not. It was also irrelevant that ANN operations are not reducible to an “if‑then” series, because computers can process in parallel; and irrelevant that training uses machine learning or produces weights and biases not foreseeable to a human, because law does not hinge on whether the program was written directly by a human or generated through computation.

The claims are not “as such” excluded: the refusal was wrong

Having held that ANNs are “programs for computers”, the court turned to whether the claims nonetheless fell within the “as such” exclusion. This is where the switch away from Aerotel mattered. Applying the EPO’s “any hardware” approach, the court held it was “beyond dispute” that the claims were to an invention and were not excluded by article 52(2)(c), because the claimed method necessarily involves computer hardware and the claims also refer to a database, a communications network and a user device. The court said: “That is sufficient to show that the subject matter of the claims has technical character and is not to a computer program ‘as such’” (para 98).

The formal holding followed: “the UKIPO was wrong to decide, for the reason it did, that the application for a patent should be refused” (para 98). The court accepted that it could simply return the case for novelty and inventive step, but considered complete silence would be unhelpful and costly. It therefore offered limited guidance on what comes next in the EPO framework: an “intermediate step” that filters which features can contribute to inventive step by asking whether they “contribute to the technical character of the invention as a whole”, with “interaction” between non‑technical and technical subject matter as the key discriminator.

Disposal: remit for novelty and inventive step

The court ultimately refused to decide, itself, how the intermediate step should apply to these particular claims or whether the contributing technical features involve an inventive step. The reason was institutional as much as analytical: the matter had been argued through the Aerotel lens at all levels, and the court did not have the benefit of focused submissions on the intermediate step or on its interaction with domestic inventive step methodology. In those circumstances, the Supreme Court thought it “unwise” to create binding precedent without full argument and specialist input.

The dispositive paragraph is explicit: “the right course is to allow the appeal” (para 118). The Supreme Court set aside the hearing officer’s decision and confined its ruling to the issues it had resolved, leaving the remaining questions “to be considered by the Hearing Officer” (para 118).

Practical implications for practitioners

The judgment is a genuine shift in the advice solicitors give on computer‑implemented inventions, including AI‑enabled systems. It changes the first‑line eligibility analysis, but it does not guarantee grant: it re‑positions where the difficult arguments will be fought.

First, eligibility objections should now be framed in the EPO’s staged terms. The invention hurdle is deliberately low (the court itself describes it as a “very low hurdle” (para 99)). In practice, many refusals previously justified as “as such” exclusions may need to be reformulated (or defended) by reference to whether, after the intermediate step, any distinguishing features that contribute to technical character remain to support inventive step. Solicitors advising on prospects should therefore separate (i) article 52 threshold from (ii) inventive step risk, and explain to clients that the “hard work” may simply have moved downstream.

Secondly, do not assume that “it’s AI” takes an invention outside the program category. The Supreme Court’s treatment of machine learning is intentionally normalising: training regimes, opaque weights and parallel processing are not legally distinctive. Advisers should expect the UKIPO and litigants to cite the court’s statement that “an ANN is a program for a computer” and to resist attempts to characterise trained models as hardware artefacts rather than executable instructions. For drafting, the message is to focus less on metaphors of “brain‑like” behaviour and more on concrete technical interaction and effect.

Thirdly, claim drafting should be undertaken with the intermediate step in mind, even though the Supreme Court did not decide how it plays out here. The court is clear that the relevant filter asks whether a feature contributes to the technical character “of the invention as a whole”, not whether it is technical in isolation. That suggests drafting should spotlight how specific claimed features interact with technical means (data structures, signal processing, device constraints, performance improvements, system architecture) rather than resting on an abstract objective (better recommendations). Where the judgment is silent on how the UKIPO will operationalise the intermediate step in examination practice, that is currently unspecified; solicitors should warn clients that implementation details are likely to emerge case‑by‑case and may evolve as UKIPO guidance and appellate decisions develop.

Fourthly, litigation risk and strategy will change. For contentious work, the overruling of Aerotel invites renewed scrutiny of existing refusals and pending appeals that were decided on steps 2–4. The Supreme Court’s reasoning provides a strong platform for appellants to argue that decisions mistakenly collapsed “invention” into novelty/obviousness. Conversely, respondents may shift focus to whether any purportedly technical features are, in the EPO sense, contributing within the overall claim context. The court also flags its caution about making unargued precedent; that creates space for early test cases to shape how UK tribunals articulate the intermediate step domestically.

Fifthly, on “clearance” (in the patentability sense), practitioners should update internal checklists. The question is no longer: “Will this fail Aerotel step 4 as non‑technical?” It is: “Does the claim clearly have technical character (usually yes), and if so, what features will survive the ‘contributes to technical character’ filter when assessing inventive step?” That requires closer collaboration with inventors and technical experts at an earlier stage, because the dividing line between (i) a non‑technical aim and (ii) a technical implementation that contributes may be determinative later. The Supreme Court did not evaluate those features here; those assessment criteria remain, for now, unspecified at Supreme Court level beyond the general framework described.

Checklist for solicitors: six takeaways

• Treat Aerotel steps 2–4 as superseded : Re‑frame advice and pleadings around the EPO/G1/19 structure; keep “invention” distinct from novelty/obviousness.
• Assume ANNs are “programs for computers” : Avoid arguing that AI models are outside the program concept; focus on “as such” and inventive step instead.
• Expect the invention hurdle to be low : If hardware/technical means are present, plan for the dispute to move to the intermediate step and inventive step.
• Draft for “contribution to technical character” : Make interactions with technical means explicit; link features to technical operation, not just business or cognitive goals.
• Manage client expectations on grant prospects : Explain that clearing article 52 does not secure a patent; novelty and inventive step remain untested and may be decisive.
• Watch for early post‑judgment decisions : Treat UKIPO practice and early appellate cases as the next source of operational detail; the judgment leaves some mechanics unspecified.

Concluding reflections

This decision does two things at once. It brings UK interpretation back into line with the EPC’s institutional centre of gravity, and it does so by dismantling a familiar domestic tool (Aerotel) rather than by incremental adjustment. Yet it is not a carte blanche for software, AI or data‑driven claims. The Supreme Court’s move is structural: it relocates the boundary‑policing from the “invention” stage to the way inventive step is assessed after filtering for features that contribute to technical character.

For practising solicitors, the immediate task is pragmatic: revise eligibility advice, revise drafting habits, and revise the way clients are told to measure “risk”. The longer‑term task is to track how the UKIPO and courts implement the intermediate step. The Supreme Court itself declined to go further without full argument and specialist input, and it expressly left open how the filter should be applied to this invention and how the contributing features should be tested for inventive step. The outcome of the underlying application after remittal, and any resulting UKIPO practice guidance, are therefore unspecified on the face of the judgment.

Articles and content on this site are for informational purposes only and do not constitute legal advice. Do not rely solely on this information.