The implications from the EPAI case are extensive and much broader than for just artificial neural network “ANN” related technologies, or what is a computer program. It relates fundamentally or what is non-patent eligible subject-matter. EPAI has implications for patentability of quantum computers, business methods, and more esoteric generative subjects both in the UK and wider afield, including across Europe and other countries. The decision affects both commercial opportunity and asset value. The EPAI application is also globally significant because it is the first hard-hitting AI /ANN case to be considered by any Supreme Court. Prior to the U.K. High Court decision in late 2023, it is fair to say that EPAI’s application was not even on the radar. Since then, it has been monitored extensively from across the world by both legal practitioners and businesses alike. Hundreds of commentaries – both right and wrong – have been published on the development of the issues now under final consideration before the Supreme Court [UKSC 2024/0131]. I perhaps should say that it was apparent to me that this case was important not just because
of its implications for the client’s portfolio but rather because of wider public impact in the context of general patentability. I predicted and appropriately advised EPAI [and its parent company Time Machine Capital 2 ] that its application would likely end up in the highly rarified atmosphere of the UK Supreme Court, and so it did. They understood the reasons why and for the need to tread this difficult path. As a patent attorney and conseil, many legal precedent decisions arise from bad cases which should never have been argued at higher judicial levels, and some of these are outdated as, arguably, is the legislation. The UK Patents Act and European Patent Convention “EPC” have both, save for some changes made in 2008, remained larger unaltered from the texts agreed in the 1970s! Bad cases make bad case law, especially since their decisions frequently generalize the issues and their reasoning is then applied to patent cases in which the facts are never the same. The EPAI case was and is the right technology and poses the right legal questions at the time, i.e., when AI is being applied in everyday environments. The patent application was drafted and then consistently prosecuted, from Day 1 until the end of the UKSC hearing in July 2025, (a) to set up the technical issues and (b) to highlight both the pitfalls of the jurisprudence and the illogical way that it was being applied, whilst (c) arguing for a reasoned and sensible approach. Is the specification perfect? I might have made some minor tweaks with 6-years of hindsight, but it is solid and makes the right noises.

The objective of the EPAI invention was to “close the semantic gap” between the output of an ANN [which output is meant to be reality when it is not, but you are told that it is] and how people actually interpret data and think. The application of the EPAI technology is, essentially, an improved recommendation engine within a better system, and so it has implications for internet searching.  All ANNs are approximation machines, but their strength lies in their ability to interpret data. ANNs self-learn, based on a training data set, and are capable of addressing intractable problems, i.e., those problems which have indeterminate variables, and which therefore cannot be addressed by coding within a conventional computer programming environment. The invention is not a “classifier,” so it does not rely on the limitations associated with taxonomy. Rather, the inventors appreciated that semantic information has a higher value, so if it was possible to map highly variable semantic information to physically extractable information then the resulting neural network could be trained more closely to reflect human perception.  The applications for this form or training and the resulting ANN are quite significant. Whilst the invention was presented in the general context of the identification of a listing of semantically close songs, the principles can be extended to allow for the identification of semantically close pictures or texts. In the latter respect, a doctor’s report with CT imaging and other data could be fed into the system to identify treatment plans [unknown to the physician] that were not identical, but which had different results with varying patency.

In training ANNs, there are some systems which simply automate a known process and take out some of the mundane leg work. For example, if you want to obtain a bank loan, then the underwriter or loan officer, in the limit, makes a ‘yes or no’ decision on an amount to be loaned, with this decision based on variable factors falling within known response ranges, e.g., does the applicant have a mortgage or own their home, how much does the applicant earn, does the applicant have dependent children, etc.? In other words, a computer programmer [different to a computer scientist] more or less knows the nature of the inputs and they know, from prior business processes and business decisions, the resulting output for a final loan decision. This is essentially an automation using AI. In contrast, the EPAI invention, in the context of music, could have any piece of music as an input, as defined by selectable musical qualities such as timbre or tonality, whilst its output is also fundamentally unknowable. The relationship between the input and output, unlike the bank loan example, has no simple mapping. This requires technical considerations and an answer to the question, how can this be achieved?’

The inventors realized that, in training an ANN, it was possible to associate/embed semantic information with physical properties. Normally, this semantic information is lost. The insight was that highly salient and important semantic information was present in a vector separation distance between two files in semantic embedding space, and that there should be a corresponding separation distance for the same pair of files as measured in different property embedding space. By converging the measured property separation distance to the measured semantic separation distance, a trained “frozen” ANN could ultimately operate simply to extract physical properties (in the form of a query vector) to allow that vector to be cross-referenced into a database to identify target files with similar vectors arising from extracted physical properties. The degree of vectorial closeness in property space between the query file and target files, determined in the resulting continuous property embedding space, thus could be used to identify, because of the earlier mapping, relative semantic closeness between the query and target files. These concepts are termed “learning objectives,” and are subsequently converted into an effective “loss function” that permits the ANN to be trained for a specific purpose. What the EPAI invention demonstrates is that ANN inventions, a core component of AI technology, have a practical application in the real-world, but importantly that the thought leadership is conceptual and at a different level of perception or “generality.” For sure, a software emulated ANN makes use of some computer coding to support its initial training [to improve accuracy and realism of its output], but such coding is simply a computer implementation, i.e., form over substance, of the important higher level conceptual and inventive thinking behind the learning objectives.

The UK Patents Act and European Patent Convention “EPC” both exclude certain forms of invention if that they relate to certain things “as such.” These things can be considered, in the sense of US patent statute, to be “abstract ideas” that do not provide a “new and useful process, machine, manufacture, or composition of matter, or any new and useful improvement thereof.” Returning to the UK/EPC, non-patent eligible subject-matter “as such” relates to inventions which are considered technically deficient, e.g., programs for computers [to the extent that they have no real-world impact and so are essentially technically trivial or a mere automation], discoveries, scientific theories mathematical methods, the presentation of information, schemes, rules and methods for performing mental acts, playing games or doing business, and aesthetic creations. A practical application for these exclusions within a bigger system is – or at least can be – another matter entirely, hence the qualifier “as such.” Unsurprisingly, there is equivalence in national German and French legislation to the “as such” rider, while other countries have logically similar considerations for exclusions, e.g., Japan, Australia, India, South Korea, and China to name but a few. One needs also to bear in mind that Art.27(1) Trade Related Aspects of IPR, a multi-national agreement of the World Trade Organisation to which the USA, the UK and another 164 countries are signatory, requires that patent rights been granted “in all fields of technology.” This was enshrined in the change to the EPC in December 2007. The UKIPO rejected the EPAI invention under a UK Court of Appeal decision called ‘Aerotel’ and its “technical effect approach (with the rider).” Aerotel is a subtractive test in that it omits from consideration any “contribution” construed to be broadly excluded, with this ultimately encouraging a distortion of the invention by subsequently not considering the claimed invention in its entirety.  The Aerotel approach has been criticised extensively, not only the EPO for being “irreconcilable with the European Patent Convention” but also by a later decision of the UK CoA [in a different composition] that described Aerotel as “a singular unhelpful test” that looked at the “gloss” rather than to concentrate on the language of the legislation. One last point to make, the judgment in Aerotel said that the UK courts should consider following the EPO approach on patentable exclusion but, at that time [October 2006], it could not do so because the case law at the EPO had not been finally settled.

Specifically, the UKIPO rejected the EPAI invention under Aerotel for being a computer program “as such” because the invention was allegedly just subjective and cognitive in nature, and the result consequently aesthetic with no technical contribution. The UKIPO just ignored the higher level conceptual pairwise approach to training the ANN to reflect semantic importance and further ignored the other parts of the claim which required selection of a close target file, based on extracted physical properties output from the trained ANN, and the communication of the identified target file(s) over a network. The UKIPO Hearing Officer did not consider the fact that the claim included a new product by a new process. Interesting, the UKIPO did not reject the EPAI invention for being a mathematical method “as such” because mathematical methods are highly specific and an application of mathematic process can bring about invention, as demonstrated in a decision T208/84-Vicom. The Vicom decision essentially related an invention that processed data in a digital filter to bring about a better image, with the decision saying that it also makes no difference if an invention is computer-implemented or realised as hardware; it is a question about the underlying technical process. The limited prior art, also identified internationally, was briefly considered by the Hearing Officer, but he identified that its teachings were considerably different. This Apple citation has been pulped by an expert report which identified fundamental differences in the respective technical approach and outcomes. What the treatment of the Apple document shows is that there is no excuse for a lack of understanding, and that poorly articulated prior art can be misapplied and misinterpreted especially when the prior art disclosure is not sufficiently enabled. Finally, the hearing officer ignored that the system communicated, over a network, a message containing a file identified as being a semantically close file to the enquiry and refused to consider the settled law of the EPO decided in March 2021 in the Enlarged Board of Appeal decision G1/19. The reason… ‘I am bound by current UK precedent.’ As an aside, the Aerotel decision has long been felt to be wrong by the profession and a mechanism that harms inventors and the protection of technology. However, there is more on this to follow…

In July 2022, EPAI appealed the UKIPO decision to the UK High Court. The decision was handed down in November 2023, at which point things positively changed for patent applicants in the UK. Suddenly, ANNs were being granted as standalone machines, and arguably the door was opened to provide protection for business methods as well as generative composition using advanced AI tech. v The High Court decided that the “as such” rider was not engaged because an artificial neural network was not deemed to be a computer, so if there is no computer there cannot be a program for a computer. It also appeared reflect an appreciation that patentability can be established by how one implements technically an aim to be achieved in a non-technical field. At the High Court, the UKIPO made no argument that the mathematical exclusion was engaged by an ANN. That is interesting because the EPO says that a neural network is a mathematical method “as such.” 

At the  original hearing, the UKIPO had already decided that the EPAI invention was not a mathematical method because “its specific application here as part of a file recommendation engine is, in my opinion, enough to dispense with the mathematical method as such objection,” so the point was seemingly moot or resolved in the eyes of the UKIPO! We will need to visit this in Question 7, but for now it is suffice to suggest that some EPO Examining divisions do not seem to fully understand the nature or the training of ANNs and, particularly, how objectively one applies EPO established jurisprudence. Let’s just say that Examiner sometimes get in wrong before seeing the light, and that the EPO approach is different to the Aerotel approach described in the answer to Q.3.

The High Court decided, commented, or observed several salient things:

1. A computer is a machine intended to operate a program, so no surprises there!
2. An ANN can be realized in equivalent hardware or software forms, but that’s also not surprising not least because of the expert report from a leading AI professor… and it’s pretty darn obvious!
3. A hardware ANN is just a “physical box with electronics in it.”
4. An ANN consists of layers of neurons which, anthropomorphising somewhat, are akin to the neurons in the brain… I feel a Homer Simpson moment coming on, “D’oh!”
5. Once trained, the structure or the topology of an ANN is “frozen” allowing the ANN to be used for real, as opposed to training, data to deliver a functional effect.
6. Once frozen, no activity which might be called programming activity takes place in relation to the ANN or the data. The data which is passed through the ANN is subjected only to the ‘processing’ provided by the ANN via its network of fixed value neurons/nodes. The term ‘processing’ is somewhat strange because the nature of a neuron’s weights and biases is no different to the value for the resistance of a resistor or the capacitance of a fixed capacitor. The weights and biases are fixed properties defining a particular unique physical characteristic. An emulated ANN is, in substance, operating at a different level (albeit metaphorically) from the underlying software on the computer [it there is any within an ANN!], and the emulated ANN is operating in the same way as the hardware ANN. If the latter is not operating a program, then neither is the emulation. Yes!
7. The physically fixed values, i.e., the thousands of weights and biases within the ANN, attributed to each of the nodes are not determined by any human programming those nodes, but are determined by the ANN itself which learns via the learning process. [Referring to the explanation of the EPAI invention (see responses to Q.1 to Q.3), the pairwise convergence of separation distances between file pairs determined in related but independent embedding spaces are the conceptual “learning objective” by which the software emulation is trained. These principles sit above their implementation in program coding, and they support the production of a loss function by which efficacy of the output can be assessed relative to a desired “ground truth,” i.e., the ultimate objective that the output should satisfy. By evaluating the loss function against the ground truth [in a relatively ‘simple’ mathematical comparison – a minor but applied math process in the context of overall ANN training] and adapting the weights and biases of neurons in a process called “backpropagation,” the error in the ANN is successively reduced until the output reaches an acceptable level of performance. Again, this is just a process, albeit a computer-implemented process realised by some coding but all subordinate, i.e., at a different level of generality, to the conceptual learning objectives.]
8. There is a distinction between the underlying software on the computer which “creates” the emulated ANN, and the emulated ANN itself. I will have to return to this point later. The abstract machine in ML/AI technology is not processing data on a step-by-step instructional basis but instead uses training data to learn the logic to solve a specific problem and thereby reconfigures the machine. Machine learning does not therefore follow an ‘if-then’ statement approach.

The High Court judge also decided that the sending, over a network, of a message containing the recommended file moved the claim outside of a computer program “as such.” What this boils down to is that a final “frozen” ANN is nothing more than a network of components, and that the network is analogous to any conventional circuit, such as a resonator. I think that term “artificial neural network” possibly gives the game away! The software realization cannot therefore be easily divorced from, say, a digital filter, such as a finite impulse response filter which, incidentally, can also be considered to be an approximation machine and which is refined by a process that is analogous to that employed in ANN training.  The real kicker and why the High Court decision changed the game to a pro-patentee position and extended the ability to acquire patent rights to other technical applications. The judge said that the result of the invention had an external effect in the transmission of a chosen file, and that it was incorrect to consider that a subjective appreciation of the output of the system was just that, subjective and in the user. The EPAI system operated on physical data, and it is noted that many inventions have a subjective output on an end user. For example, does a massage gun produce a subjective benefit on the recipient of the massage gun’s oscillations? Is an image, produced by a digital filter, on a screen viewed as better or worse by two people with differing eyesight? Who makes the decision? The justification for there being invention was explained in terms that, in EPAI’s invention, a file has been identified, and then moved, because it fulfilled certain criteria. It might be true that those criteria, e.g. semantic appreciation of tonality, may not be technical criteria in the sense that they can be described in purely technical terms, but they are criteria nonetheless. The resulting trained ANN was considered to have gone about its analysis and selection in a technical way because it had not selected just any old file; it was a file identified as being semantically similar by the application of technical criteria which the system has worked out for itself. The output was therefore a file that would not otherwise have been selected. One cannot ignore the fact that a technical thing is actually produced. It would not matter if the user never listened to the file. The file, with its similarity characteristics, is still produced via the system which has set up the identification system and then implemented it. 

Is it not really a question of how one achieves/implements technically a desired result irrespective of whether the invention resides clearly inside a “field of technology” or the invention is set up technically to address an aim to be achieved in a non-technical field? Perhaps the real question is, ‘what is technology?’ Once one starts to appreciate that the nature of the final output is not important and can be subjective, the nature of inventions that can be protected grows… and industry and investors alike are happy! One final observation, the UKIPO tried to suggest that the application lacked sufficiency and required disclosure of the training dataset. My comment, “poppycock.” Such training data has no relationship to the computer scientist’s – not a computer programmer’s – ability to generate the conceptual “learning objectives.” And I would also like to point out that the disclosure of the training data would probably run to another 20,000 pages for the patent application.

Now the rollercoaster really begins…

So, July 2024 arrives, and the UK CoA “vacated” the High Court decision. I hear John McEnroe scream, “You cannot be serious!” The CoA agreed that the form of implementation of a neural network is irrelevant. Consequently, a software emulation equals a hardware realisation. They also said that a hardware ANN is nothing more than a collection of resistors and transistors interconnected by wires! It’s a circuit or network of interacting components, as I have said previously. But that was not all the CoA said… It seems worthwhile briefly to point out that the UKIPO’s verbal and written submissions included statements that (i) it had “no skin in the game,” notwithstanding that it receives considerable income from fees charged for patent applications and renewal fees charged for maintaining patents in force, and (ii) the UKIPO’s “interest in this appeal is simply to determine the law and its proper application in this area,” to achieve clarity for all. I would like to quote part of an interview given to my editor friend by Steve Jobs, “Horse apples,” and I don’t mean the fruit! More on that a bit later. At the CoA, the UKIPO also introduced the issue of the mathematical exclusion “as such” even though, strictly speaking, this was not an avenue open to them because of the previous lack of argument before the High Court. Why would the UKIPO, at the CoA, possibly want to criticise its own considered position from a senior hearing officer? Win and all costs? Hubris? Or a desire to resolve how the exclusions to patentability apply? However, it really did not or does not matter.

The real concern from EPAI and the legal team was that the CoA was preconceived and just believed that some form of AI surely could not be anything other than a form of excluded mental act to realise a “subjective and cognitive” result and a matter of “aesthetic.” That potentially prejudiced – and dare I suggest technically naïve – belief is not unreasonable given the statements made during the May ’24 hearing and the decision itself. Any preconception was always going to support a way to overturn the original High Court decision if the concept about the protection of core AI did not sit well with the judges of the CoA. Basis for these concerns is set out below with reference to the events as they unfolded. The UKIPO’s counsel spent the best part of two hours trying to explain the technical nature of artificial neural networks “ANNs.” I still fail to see why it took so long but can say that the lengthy presentation was obfuscation and a tactic or it was just confused. It thoroughly confused me, so I can only imagine what the judges thought. Anyway, my suggestion that the CoA may have harboured some preconceptions is far from unreasonable since they decided to ignore many technical considerations of the invention and, particularly, the features and language of the claim. For a reason known only to one judge, the court suggested that the sending of a message, ignoring its content, was also excluded because the invention was “aesthetic in terms of a recommendation, it is a mental act, and it is presented to you in the form of a list, which is presentation of information.” ‘Bonkers,’ do I hear people shout? No complaints from me if I do, but I might also be a bit biased but hopefully not. Most of these issues had never crossed anyone else’s objective thinking, so why?

The main arguments that resonated with the CoA in its decision were that:

(a) the weights and biases of neurons, upon aggregation, amounted to a computer program and thus were excluded from patentability, and 

(b) an ANN was a computer because (i) a computer was “a machine for processing information,” and (ii) a computer program was just “a set of instructions for a computer to do something.” 

In understanding the exclusion relating to a program for a computer, it is necessary ‘not to put the cart before the horse.’ One firstly needs to understand what amounts to a computer. The definition of “computer” by the CoA was and has been troubling to many because a computer, under this decision, would be defined by any electrical or electronic circuit, and the values attributed to conventional resistors and capacitors would amount to a computer program from the logic identically applied to an ANN in the decision. This could and does potentially place existing patent rights in the UK under scrutiny and threat of invalidity even though historically there was never a problem with this form of circuit in a patent claim. To further expose the lunacy, the computer definition could also apply to a sextant, a mechanical device that takes a determined inclination of the sun or landmark and converts that into an estimated latitude or a measure of distance off. Did the CoA decision arbitrarily differentiate neural network technologies from conventional circuits, thereby applying inconsistent practice and prejudice?

In January 2025, this computer definition saw the UKIPO revise its patent examination guidelines to say that ASICS and quantum computers would need to be assessed as computers and their functions considered (or not considered) potentially to be excluded programs for computers “as such”. Alarm bells were now ringing in industry! This goes to the point made above in the answer to Q.1 and problems of generalisation in legal decisions. Industry’s concerns were drawn to the attention of the U.K, Supreme Court, as discussed later.

One now needs to understand what is a computer and what was reasonably – if not actually – the legislative intent of the exclusion for programs for computers “as such”? 

The CoA judgment described the function of programmers, not computer scientists, to produce programs written in a form “which is sometimes called a high-level programming language… [and] ordinary computers work by running machine code, which is different and hard for humans to understand… [and] derived by a computer system (normally what is called a compiler program) under the direction of a human programmer. There is no justification for drawing a distinction in law between instructions created by a computer and those created by a human.”  Patent law is there to protect technology and technical innovation, so technical terms require technical understanding. From multiple dictionaries, technology is “the branch of knowledge dealing with the mechanical arts and applied sciences;… [t]he application of such knowledge for practical purposes”, “the practical, especially industrial, use of scientific discoveries”, “the practical application of knowledge especially in [engineering]”, and from T0154/04-Duns Licensing at [8] “creations in engineering and technology… practical scientific applications.” Technology, at its core, is the practical application of scientific or engineering knowledge. 

Technology also evolves. We should bear in mind the nature of computers in the 1970s and the nature of programs. At the point of drafting the legislation, these were machines for determining rocket re-entry angles and coordinates, and sometimes for playing “pong” tennis in which paddles moved up and down a screen to deflect, upon determined pixel contact, the movement of a simulated electronic ball. We are presently teaching our 16–18-year-olds that a computer has certain component requirements to permit execution of computer programs. The component was known at the time of drafting the legislation in the UK and for the EPC, and the component remains applicable today and for future quantum computing. Even the courts have referred to the component(s). It is a Central Processing Unit “CPU,” or its equivalent QPU for 21st century quantum computers. The definition is taken from work in 1939 [the Harvard computer] and then 1945 von Neumann architecture which both say that a CPU, in a computer, is required and that a CPU is actually more basic and requires at least some registers and an arithmetic logic unit to support program operation (not to be confused with even more fundamental instruction set operation). The CoA missed the point that an ANN is not a computer because it is simply a circuit without an active CPU function, so a trained ANN does not execute instructions at a level at which a computer is conventionally and technically defined. This is explored more in the response to Q.6. So why did the CoA decide to depart from what reasonably seems to be what the UK and European legislation intended to be a program for a computer? We might look at the “analogue/analog computer” argument discussed in answer to Q.6, understanding for now that analog computers are themselves just circuits that do not include a CPU function in the way a digital computer does. This is evident from just a basic online search query. Analog computers use physical quantities, like voltage or current, to represent data and manipulate these directly through the physical characteristics of components or via mechanical linkages to perform calculations.

The proceedings before the UKSC are interesting and impactful because fundamental patent principles are under scrutiny. The questions being resolved by the UKSC are much bigger than just applying to AI. The function of the UKSC is the final court of appeal in the U.K. There is no further appeal possible… anywhere. The UKSC only “hears cases of the greatest public or constitutional importance affecting the whole population.” In this respect I note that, in patent law and over the course of my 35-year career, there are only a handful of decision from the UKSC (or formerly the House of Lords) on patent matters.  There needs to be some context of the machinations going on behind the scenes. Returning to the UKIPO’s ‘objective’ to seek clarity in the law for the exclusions to patentability (see answer to Q.5), it is worth noting that the U.K. CoA denied leave to appeal to the UK Supreme Court. EPAI therefore was forced, in July 2024, to seek permission to appeal. The UKIPO objected to this application. However, the nature of the grounds raised resulted in the UK Supreme Court allowing a full and final appeal to proceed. EPAI had no issue with upfront agreeing that “interventions” – these are amicus curiae briefs – could be filed at the UK Supreme Court, yet the UKIPO were far from forthcoming and again objected. In granting the leave to appeal, the UKSC would already have been questioning the sense and reasoning of the CoA decision, even though the CoA decision was delivered by apparently two of the preeminent U.K. patent judges with significant patent experience, or was it that the UKSC wanted to put the issue finally to bed. More on the significance of any decision from the UKSC will be explained in the response to question 8.

The issues under consideration by the UKSC are more fundamental than just the invention. They go to the core of patentability in the UK and Europe.

The common and core axioms to international patent legislation are:

1. Exclusion to patentability, i.e., what is non-patent eligible subject-matter?
2. Novelty of the invention viz all that has been known before.
3. Obviousness or “inventive step,” i.e., is the claimed invention obvious over a combination of two or more documents and/or common general knowledge?
4. Does the invention have industrial application?

These fundamental requirements are not buried deep down in later sections of the respective patent legislation but are front and centre in the principal sections of all global patent legislation. The UKSC, in the Emotional Perception EPAI case, is now looking at (1) the exclusions and (3) the way in which obviousness is to be assessed.

The UKSC will further rule on or has been asked to rule on:

1. What amounts to a computer and whether and ANN is a computer?
2. What amounts to a program for a computer, and whether weight and biases in an ANN are instructions in the context of a program for a computer?

But these are subsets of much bigger questions, particularly:

1. Is the UK approach to assessing exclusions to patentability “as such” in line with the approach taken by the European Patent Office and, if not, does it regardless give the same result even though the tests are different?
2. Should the UK approach under the Aerotel statutory exclusion test (see the response to Question 2) be jettisoned as bad law and replaced by the EPO’s “any hardware” approach for assessing exclusions to patentability? [Explanation: The presence of ‘any hardware,’ even if this were simply a known cup onto which an image is surface printed, would see the test for exclusion passed and the claimed invention then considered on the merits of inventive step/obviousness. In this example, the cup is technical and ‘any hardware’, so the invention is not excluded.] There are nine (9) influencing and articulated factors for why the ‘any hardware’ approach could be adopted, namely:
   1. The EPO’s Enlarged Board of Appeal “EBA”, in decision G1/19, said that the ‘any hardware’ approach is settled law. 
   2. Various decisions of the UK Court of Appeal have simply said that, before the EPAI case, the issue of “settled” European law had not been present, had not been arguable and so could not be considered by the courts.
   3. The CoA has said, even in Aerotel, that harmonization to a decided settled position of the EPO would be appropriate, unless the change meant that the case law was being steered by the commodore onto the rocks.
   4. The Aerotel test has already been criticized as being “singular unhelpful” in that it looks at the “gloss” and not the substance of the legislation.
   5. The EPO has long said that the Aerotel test is “is not consistent with a good-faith interpretation of the European Patent Convention in accordance with Article 31 of the Vienna Convention on the Law of Treaties of 1969… [and] is irreconcilable with the European Patent Convention also for the further reason that it presupposes that “novel and inventive purely excluded matter does not count as a ‘technical contribution'” (Aerotel/Macrossan, e.g. paragraph No. 26(2)). This has no basis in the Convention and contravenes conventional patentability criteria.”
   6. The considered opinion of UK industry is that the UK and EPO approaches give inconsistent and different results detrimental to industry and technological innovation by representing a discouragement to business investment and the damage to the innate value of commercial enterprises.
   7. UK patent statute, at section 130(7), expressly states that the considerations of patent eligibility and obviousness must have, as near as practicable, the same effects as the corresponding provisions of the European Patent Convention.
   8. Are the requirements of Art.27(1) TRIPS, requiring grant of patent rights in “all fields of technology” satisfied by the UK approach?
   9. The EBA has indicated that the exclusions to patentability must be viewed narrowly, so how does the UK approach in this instance justify the scope and reach of the definitions provided for a computer and a program for a computer as decided by the CoA in EPAI?
3. If the ‘any hardware’ approach is adopted, then the question about what qualifies as a computer program or a computer becomes less significant because a computer implementation would satisfy the ‘any hardware’ requirement. The technical effects achieved by the invention, as a whole, would then be considered under the test for obviousness [per another settled position of the EPO under T641/00-Comvik]. This has broad and immediate ramifications for protection of technologies in the UK because claims to business systems or generative esoteric results would be more easily protected, as they are to an extent already in the EPO. 
4. If ‘any hardware’ is now appropriate, does this mean that the UK’s approach to assessing inventive step is also still in line with that of the EPO, or does the UK Pozzoli test require modification? In fact, does the UK now need to consider adopting the EPO’s problem-solution approach to considering inventive step/obviousness?

We can perhaps now see why EPAI is being described as a generationally important case. One decision potentially resets quite a lot or has major limiting effect. At the UKSC hearing, the UKIPO denied that there was a need to change to any hardware by arguing that the competing tests give the same result, and that a change would mean that the “roof would fall in.” The profession and industry said otherwise. The UKIPO argued that a change would mean re-training about four hundred patent examiners, and also revisiting pending cases rejected under the Aerotel test. They also said that a change would prevent immediate rejection of an application without conducting a search. So what! Law changes and bad law should not be retained for reasons of ease, and applicants pay fees for securing patent rights and are entitled to receive the search or examination for which money has been paid. Finally, the UPIPO alleged that a move to ‘any hardware’ would produce a “Frankenstein” test because the UK Pozzoli questions for inventive step analysis applies to both grant procedures and infringement actions before the courts, whilst the European Comvik approach is simply related to grant. Firstly, moving to ‘any hardware’ would be significant for applicants since it potentially supports acquiring technical business inventions if not supporting more certain acquisition of patent rights in wider fields of technology, as required by Art.27(1) TRSPS. 

It is necessary now to understand the Comvik approach to assessing inventive step/obviousness and how this potentially knits into the Pozzoli obviousness test/questions in the UK.

The EPO’s Comvik decision requires that, in a mixed invention containing both technical and non-technical features, when considering inventive step/obviousness, you exclude from consideration those features which do not contribute technically to the invention. It therefore requires that you look at the nature of feature interactions. The approach is described as a “two-hurdle” process requiring three steps. Following ’any hardware,’ an intermediate step simply establishes whether a feature contributes to the technical character of the invention, and that intermediate step serves as a filter for features contributing to a technical solution of a technical problem in view of the closest prior art. 

EPAI’s sensible proposition was that, once one establishes the presence in a claim of ‘any hardware,’ it is simply a requirement to ask the statutory question, ‘is there an invention and, if so, what exactly is that invention? In this sense, invention means the invention as claimed less any features which are properly excluded. One then just asks whether the remaining parts are obvious. This is not rocket science, although it seems that the current approaches are operated with a degree of subjectivity. I am reminded, in this respect, of a case I argued at the EPO where a Board of Appeal member said, “We are aware that Examining Divisions can operate with a degree of mental laziness in that it is too easy to throw too many features out of the basket when considering the issue of inventive step.”  Subjectivity is therefore a recognised and prevailing problem, and that is true for EPAI’s pending European application (discussed in Q.7). The suggested provision of an additional “rider” to Pozzoli essentially aligns the inventive approach in the UK more closely to that suggested by Comvik, and forces examiners to consider all interactions and everything in context of the invention as a whole. However, the issue of inventive step analysis is not bound by one specific approach since it makes no difference, if the invention is properly understood, whether you apply the UK’s Pozzoli approach, the EPO’s problem-solution approach and Comvik for mixed inventions, or the recent “concrete elaboration” realisation where the invention is simply a better solution over the prior art. In summary, the upside of the proposed rider is that one can do whatever one wants to do with inventive step because you will never be assessing inventive step by reference to exclude subject-matter as long as you properly identify the unit invention itself. The decision from the UKSC could – and arguably should – lead to some significant changes for the UK, but a decision of the UKSC to include some form of rider to Pozzoli would also act as guidance to the EPO on how to avoid subjectivity when applying Comvik. The resulting acquisition of patent rights should therefore be cheaper and quicker because an understanding of the limited effect of the exclusion is settled, and the subsequent assessment of inventive step and real-world contribution checked for contextual objectivity.

It is a question of understanding. 

I would suggest that any electrical or mechanical system can be modelled mathematically, but does this make the invention, implemented in some form of code, an abstract idea or a mathematical method that is excluded “as such?” So, probably not! It is about assessing the practical application, i.e., the real-world impact of the technology of the claimed invention. Software emulated ANNs are not considered to be any different, technically, to a hardware realisation. It is merely a choice of how one implements the ANN, as the UK CoA has already said. 

The CoA suggested, obiter, that the claimed invention might fall under the mathematical exclusion, but this might be irrelevant, a read-herring and/or superficial if the UKSC advocates ‘any hardware.’ 

The EPO’s Guidelines for Examination already indicate the applied use of ANNs supports a patentable invention. We also see heavily business-orientated methods granted when some technical interaction arises between features, such as T1051/07-SK Telecom. Given the clear technical nature of the learning objectives and the pairwise convergence of separation distances in different embedding spaces, it is difficult to understand how this is a mathematical method.  Admittedly, backpropagation used in the training of an ANN usually requires a mathematical comparison of vectors, but backpropagation is a minor backend part of the overall training regime. Backpropagation has nothing to do with initial setting of conceptual learning objectives. If that is indeed the case, then learning objectives for training ANNs should be patentable themselves because they are at a higher level of generality and conceptual ingenuity. I am not aware of a case where the EPO has had to consider this fact.

There are multiple aspects to this response. 

The EPO has taken a different approach to the UK Court of Appeal because the EPO agrees that the claimed invention is not excluded from patentability. The EPO allege that the invention is obvious over a single piece of prior art that, already, the UKIPO has said is significantly technically different. In fact, the EPO also has the same expert report that critiques and then exposes the major differences and limitations of the prior art over the claimed invention. The claims are also slightly different.  The EPO Examiner cites the right law – Comvik – and indeed correctly states that one has to identify all the features that contribute and only exclude that hide that do not contribute, but the Examiner then butchered the application of the law by excluding features that evidently are both technical and which contribute to the invention. How does one square away the nature of the learning objectives in the production of a new product by a new process? I refer to my comments about subjectivity and misapplication of the law; this is a case in point. On a correct analysis, the EPO situation is not relevant to the issues before the UKSC, namely (1) how to deal with exclusions to patentability and how broad are these, and (2) is guidance required on the approach to assessing obviousness.

A decision from the UKSC is not subject to any appeal, and the decision is only binding on the U.K. itself. However, if the UKSC decides one way and this contradicts the modus operandi and contra-determinations of the EPO, we clearly have a problem. The law is meant to be the same at this national and regional level, so how can commonly agreed statutory language give different results or different degrees of protection, including none or only in certain areas? So, although not binding, the UKSC decision is likely to have influence on at least the EPO and contracting states of the EPC. The EPO views the exclusions to patentability, in the context of the TRIPS required “all fields of technology,” to be extremely limited. The UK CoA has taken a much broader stance with its definition of at least a computer in the context of a program for a computer, and also what amounts to the presentation of information (discussed at Q.5). The scope of the limitations to patentability of Art.27(1) TRIPS will therefore likely be settled given that it is not likely to be seen again at a supreme court anywhere in the world in the near future, although a decision from the UKSC relative to the statement, in G1/19, that the exclusions are limited would also be difficult to square. I understand that the UKSC proceedings are being followed by many national patent offices because of the rarified review of some fundamental issues on patentability. Will this encourage other jurisdictions to follow the UKSC decision? That depends. Rulings from different jurisdictions will be accorded different weight as persuasive authority.

A decision from the UKSC can certainly be influential (or persuasive) and could potentially be applied, particularly in Commonwealth countries or countries operating with common law, but this also depends on the context – patent law is fairly uniform in its objectives and the needs for novelty, inventive step and what is fundamentally excluded from protection.  A decision from the UKSC could influence national legislative change, although change at a national UK level would need to take account of multi-national agreements. To obtain agreement to adapt, say, Art.52 EPC would require hard to reach international consensus, so the final EPAI decision is likely to be with us for some considerable time.

Followers of this thread will be pleased to see that this is a short answer!

Until the UKSC delivers its decision, the current tests (like Aerotel and the EPAI CoA decision) will continue to be applied by the UKIPO. This contrasted with a sea-change when the High Court decision issued, with cases granted for a variety of technologies, including methods of training an ANN and more “subjective and cognitive” inventions in generative AI. If my observations above suggest that change is afoot for the benefit of applicants, then it will be a question of trying to delay a final refusal on pending applications or making use of extensions to slow down prosecution. I would expect the UKSC decision to be handed down towards the middle or end of November, but that is simply my best guess.  The NVIDIA case GB2112104.1 is unfortunate because of its timing. The UKIPO rejected its applied ANN technology for being both a computer program and a business method “as such”. Technology moves forward, whereas the UKIPO arguably remains thoroughly entrenched. In this NVIDIA application, the invention relates to the training of a neural network using medical imaging data and clinical data, and the use of the trained neural network to determine a treatment for a patient. Computer tomography (CT) scans and clinical data were used to train the neural network to predict oxygen therapy treatment. With the refusal, a clock starts for filing the appeal, and any delay in meeting a 4-week deadline would require justification. It may simply be unfortunate timing for NVIDIA.

Moving to ‘any hardware’ would be significant for applicants and supports acquiring rights more cheaply and more quickly. Moving to a greater objective application of a test of inventive step would also be beneficial. If the UKSC supports the High Court decision and reverses the CoA findings, then opportunity to obtain wide-ranging IP rights in the UK – and possibly internationally for the reasons expressed in the response to Q.8 – becomes increasingly easy, including in AI, ANN and generative technologies as well as, potentially, more esoteric fields relating to suitably technically-implemented business processes [i.e. those not being simplistic automations but embodying some secondary and meaningful technical considerations].  I would suggest that the UK, between late November 2023 and July 2024, was possibly the most pro-patentee country in at least the western world. A positive decision would also simplify business decisions and provide increased certainty whilst reducing patent acquisition costs. Acquisition of patent rights still requires careful thought to ensure that the application says what it needs to say and is technically focused on both the implementation and outcome, including identifying interactions, real-world effects, and/or practical applications. Answering this negatively, if the UKSC leaves in place the current state of play from the CoA decision and, particularly, what amounts to a computer, a program for a computer and how exclusions to patentability are to be assessed, then this will likely cause damage to industries and companies operating in the U.K. (both homegrown and UK-based subsidiary companies). Such damage would likely take several forms and would include (i) lost earnings from a loss of competitive advantage, (ii) loss of incentive for R&D investment, (iii) distortion of the UK market through encouragement of foreign “knock-off” imports that are not subject to enforceable monopoly rights, (iv) encouragement for SMEs to re-locate activities offshore from the UK to competitor jurisdictions; and (v) loss of company value through an erosion of UK patent asset rights. 

Whilst one cannot predict how a court will react and there is always a chance of a bad decision, the incentives for change (see answer to Q.6 at point (iv)) appear reasonable and fairly compelling.  Patent Offices around the world are looking for guidance in AI , as are their courts. A decision from the UKSC will therefore likely be influential, although national procedures will likely require the baton of EPAI to be carried into some national courts. Of course, the EPAI decision could also be leveraged to bring about a change, if possible, in national law, but that is a government issue.

For a strategy, I reiterate the points made above: 

1. Try to slow prosecution for existing patent cases in AI and some applied more esoteric applications until November(ish).
2. Realise now that there could be meaningful change on the horizon, and that a more pro-patentee phase may soon be with us all. Opportunities to protect technologies, hitherto difficult in the UK (at least), could realistically open, so consider how these changes can be leveraged for your company’s commercial benefit.
3. Ensure that your patent application says what it needs to say and is technically focused on both the implementation and outcome, including identifying interactions, real-world effects, and/or practical applications. Spend the time to make sure you cover the bases because this decreases the overall acquisition costs for each patent.
4. Remember that IP is, ultimately, a commercial opportunity to make a better return on investment.