Emotional Perception sought to patent an artificial neural network system recommending media files by emotional similarity. The Supreme Court held that an ANN is a computer program, but abandoned the Aerotel test in favour of the EPO’s G1/19 ‘any hardware’ approach. The claims were therefore not excluded, and the application returns to the Hearing Officer.
Background
Emotional Perception AI Ltd applied for a patent for a system (claim 1) and a method (claim 4) that use an artificial neural network (ANN) to recommend media files, such as music, video or text, that are likely to produce a similar emotional response in a human user. The lower courts concentrated on claim 4, and the Supreme Court took the same approach.
Law Teacher writing support
Writing a law essay?
We choose a writer who is an expert in your area of law and qualified to at least the academic level you select. Most of our writers hold a master’s degree or PhD, have teaching experience, or both.
Send us your question and, if you wish, any marking criteria. We’ll write an essay to your instructions.
The applicant did not claim to have invented ANNs or machine learning. Its claimed contribution lay in the way the ANN is trained. Pairs of training files are compared in two ways:
- by their semantic similarity, derived from natural language processing of human descriptions and expressed as a separation distance in “semantic space”; and
- by their physical properties (for music, for example, tempo, tone, volume and rhythm), expressed as a separation distance in “property space”.
Through backpropagation, the ANN is trained so that distances in property space converge towards the corresponding semantic distances. The lower courts called this the “trick” of the invention. Once trained, its weights and biases are frozen. It then makes recommendations solely from the physical properties of files, without recourse to any semantic comparison.
Procedural history
- The UKIPO Hearing Officer refused the application under the exclusion of “programs for computers … as such” in section 1(2)(c) of the Patents Act 1977, which corresponds to article 52(2)(c) and (3) of the EPC.
- Sir Anthony Mann in the High Court allowed the appeal. He held that an ANN is not a computer program at all and that, even if it were, the invention was not excluded.
- The Court of Appeal (Birss LJ, with Arnold and Nicola Davies LJJ agreeing) restored the refusal.
Every decision below applied the four-step test in Aerotel Ltd v Telco Holdings Ltd [2006] EWCA Civ 1371.
The issues
- Issue 1: whether, in the light of the EPO Enlarged Board’s decision in G1/19, the Aerotel approach should still be followed.
- Issue 2: whether an ANN is, or contains, a “program for a computer”.
- Issue 3: whether the entire subject matter of the claims is excluded as a program for a computer “as such”.
The Court’s reasoning
Lord Briggs and Lord Leggatt gave the judgment, with which Lord Hamblen, Lord Stephens and Lord Kitchin agreed.
Issue 1: departure from Aerotel
The Aerotel test. Aerotel adopted the “technical effect approach with the rider”. It set out four steps:
- construe the claim;
- identify the actual contribution;
- ask whether that contribution falls solely within excluded subject matter;
- check whether the contribution is actually technical in nature.
Aerotel had expressly rejected the EPO’s “any hardware” approach. In G1/19, however, the Enlarged Board endorsed that approach without qualification, stating that:
“A method involving technical means is an invention within the meaning of Art 52(1) EPC.”
The Enlarged Board also adopted the Duns principles. By implication it accepted the criticism in Duns that Aerotel was not a good-faith interpretation of the EPC and confused the legal concept of “invention” with novelty and inventive step.
The status of EPO case law. UK courts are not bound by EPO decisions. They should, however, follow Enlarged Board decisions unless convinced that they are wrong or beyond the ambit of reasonable difference of opinion. The Court cited Lord Neuberger in Human Genome Sciences:
“It is plainly appropriate in principle, and highly desirable in practice, that all these tribunals interpret the provisions of the EPC in the same way.”
The Court also recalled Jacob LJ’s remark in Actavis v Merck:
“In the unlikely event that we are convinced that the commodore is steering the convoy towards the rocks we can steer our ship away.”
The Comptroller General’s arguments. The Comptroller General made three main submissions, all of which were rejected:
- No proven difference in outcomes. The Court said this approached the problem “from the wrong end of the telescope”. It is not for the party seeking alignment with the EPO to prove that alignment would change results.
- A “Frankenstein monster”. It was said that adopting G1/19 would force the EPO’s problem-and-solution approach onto UK courts and so overturn Pozzoli. The Court disagreed. Problem-and-solution is a matter of EPO practice, not something dictated by the EPC. Pozzoli, as approved in Actavis v ICOS, remains legitimate.
- Same exercise, different order. It was said that Aerotel does the same work as Comvik in a different sequence. The Court rejected this as well.
Why Aerotel was rejected. The decisive reason was that the Enlarged Board’s rejection of Aerotel rested on a determination that the Court of Appeal had misinterpreted the EPC. It was not merely a preference for a different methodology. The Court also found “real force” in the EPO’s criticism on its own terms. On a plain reading, article 52(2) goes solely to whether the subject matter is an invention. Aerotel instead asked whether the “contribution”, meaning what the inventor “has really added to human knowledge”, was excluded. That brings novelty and inventive step into the analysis. The Court said:
“The Aerotel approach at best jumbles up the test of an invention with the other requirements for patentability (if indeed it mentions that first question at all) and reverses the logical order of analysis”
Step 1 of Aerotel (construing the claim) survives, but steps 2 to 4 “cannot be maintained”.
The replacement approach. UK courts should apply the Duns principles, except principle G and the second paragraph of the Comvik headnote, as explained in G1/19. This includes the “intermediate step” between the “any hardware” stage and the assessment of inventive step. The question whether the claim discloses an “invention” must be considered first and separately.
Issue 2: is an ANN a computer program?
The parties’ positions. The applicant drew a distinction between a “hardware ANN” and a software “emulated ANN”. It argued that a computer requires a CPU and that a program is a sequence of logical instructions, so an ANN contains no program. The interveners (CIPA and the IP Federation) argued that an ANN is a computational model and that its weights are data rather than instructions.
What counts as a computer. The Court agreed with the applicant and interveners that the Court of Appeal’s definition of a computer as “a machine which processes information” was too broad, since it would capture ovens and jukeboxes. What marks out a computer is that it performs computations. But the term should not be confined to conventional digital computers. The EPC should accommodate technological change, including analogue and quantum computers. Nicholls LJ’s CPU-based description in Gale was given in the context of a digital computer and should not be read as limiting.
The nature of an ANN. An ANN is an abstract model, not a physical object, so it is not itself a computer. The Court of Appeal was mistaken on that point. The distinction between hardware and emulated ANNs was “untenable” because the choice of implementation is merely practical. The machine on which an ANN runs functions as a computer.
The ANN as a program. The Court held:
“Whatever the specific form of the machine on which an ANN is implemented, the ANN constitutes, in essence, a set of instructions to manipulate data in a particular way so as to produce a desired result. In other words, an ANN is a program for a computer.”
The program is not limited to the weights and biases. It also includes the network’s topology and activation functions. The following features were held to be irrelevant:
- that the instructions may be fixed rather than changeable (applying Gale on ROM);
- that the network processes data in parallel rather than as “if-then” logic;
- that the instructions are generated by machine learning rather than written by a human;
- that the trained weights cannot be predicted in advance.
On the last point the Court endorsed the reasoning in Mitsubishi (T 702/20) that a neural network “still operates according to the programming of its structure and learning scheme”. The computer program exclusion was therefore engaged.
Issue 3: is the subject matter excluded “as such”?
Under the “any hardware” approach, the claims are to an invention. The ANN can only be implemented on computer hardware. The claims also refer to a database, a communications network and a user device. That was sufficient to give the claims technical character, so they are not to a program “as such”. The UKIPO was therefore wrong to refuse the application on the ground it relied on.
The intermediate step
The Court gave guidance on the intermediate step required by G1/19. Its main features are:
- It is implied by article 52 and is not merely EPO practice, because the “any hardware” threshold is too low to allow reliable testing for novelty and inventive step.
- It requires the technical character of the invention, viewed as a whole, to be identified first. Any appropriate method may be used, potentially including the Pozzoli “inventive concept”.
- It is a feature-by-feature filter. Its purpose is to exclude features that do not contribute to that technical character from the assessment of inventive step.
- The sole criterion is contribution to technical character. Technical features may be filtered out, and non-technical features that interact with the technical subject matter may be retained.
- It differs sharply from the holistic step 4 of Aerotel.
- It arises only once the claim has passed the invention test.
- Novelty and inventiveness are irrelevant at this stage.
Why the Court did not apply it. The Court declined to apply the intermediate step to these claims. The case had been argued throughout through the lens of Aerotel, and the Court had received no significant submissions on how the intermediate step should be conducted. It considered it would be unwise to create binding precedent in the rapidly expanding field of AI without in-depth argument and the views of the UKIPO and the specialist judges below.
Practical significance
The decision overrules the long-standing Aerotel methodology and aligns UK law with G1/19. The “invention” hurdle for computer-implemented inventions now operates as a very low, hardware-based threshold. The real filtering will take place at the intermediate step and at the stage of novelty and inventive step. Pozzoli nonetheless remains a legitimate approach to inventive step.
The Court confirmed that ANNs, including trained networks with frozen weights embodied in dedicated hardware, are “programs for computers”. This removes any argument that AI models fall outside the exclusion altogether.
How the intermediate step is to be applied to AI and other computer-implemented inventions, and whether earlier English case law on technical effect still assists, have been left for future cases. The UKIPO and practitioners will need to re-train and adapt accordingly.
Verdict: Appeal allowed. The decision of the Hearing Officer was set aside. The Supreme Court held that the Aerotel approach should no longer be followed and that UK courts should apply the G1/19 / Duns approach. It held that an ANN is a program for a computer, but that the claimed invention is not a program for a computer ‘as such’ because it involves hardware. The case returns to the Hearing Officer to apply the intermediate step and to consider whether the features contributing to the invention’s technical character involve an inventive step.
