Emotional Perception AI Limited v Comptroller General of Patents, Designs and Trade Marks

This Supreme Court case is about whether an AI invention built around an artificial neural network can be blocked by the exclusion for programs for computers as such.

The court made two commercially important points. First, an ANN can count as a computer program. Secondly, that does not automatically defeat the patent application. In this case, the claims were not excluded at that threshold stage and the application was sent back for the next patent tests.

That makes the decision important for any business trying to patent AI, software or recommendation technology in the UK. It changes the legal framing, but it does not create a shortcut to patent protection.

The applicant, Emotional Perception AI Limited, wanted patent protection for a recommendation system. The business idea was easy to recognise. A user who liked one file, such as a song, video or text, could be recommended another file likely to trigger a similar emotional or semantic response.

The company said its system could do this more quickly and simply, and with better recommendations, than what was already available. The invention was not the ANN itself. The claimed invention was a particular way of training an ANN to match measurable physical properties of files with human semantic descriptions.

The training used pairs of files. One measure came from semantic descriptions processed into a semantic space. Another came from measurable physical properties processed into a property space. The ANN was trained so that distances in property space would align more closely with distances in semantic space.

The judgment describes that alignment step as the significant trick in the invention. The ANN used backpropagation to adjust weights and biases so that the error between the two kinds of separation distance was reduced over repeated training cycles.

After training, the weights and biases were frozen. The trained ANN could then analyse a target file by physical properties alone, generate a vector for it, compare that vector with vectors for reference files in a database, and recommend files with similar vectors. The system also involved a communications network and a network-connected user device that received and output the recommended files.

The UKIPO Hearing Officer refused the application because it was said to be excluded as a program for a computer as such. The High Court disagreed and allowed the appeal. The Court of Appeal reversed that result and restored the refusal. The Supreme Court then had to resolve both the legal framework and the result for this application.

The Supreme Court had to answer three linked questions. First, should the UK continue using the Aerotel approach when applying article 52 of the European Patent Convention and the Patents Act 1977? Secondly, is an ANN a computer program at all? Thirdly, if it is, were these claims still excluded as a program for a computer as such?

Those questions mattered because the exclusion rules are only one stage of patent analysis. A claim can fail early if it is treated as excluded subject matter. But if it survives that stage, it still has to satisfy novelty, inventive step and industrial application.

So this was both an AI case and a wider case about how UK law should handle software-related patent claims. The answer would affect not only this application, but also how businesses and advisers frame future filings for computer-implemented inventions.

The Supreme Court allowed the appeal. It held that the Aerotel approach should no longer be followed and that the UK should adopt the same interpretation of article 52 as the EPO Enlarged Board in G1/19.

The court was careful about what that meant. It did not say that UK courts or the UKIPO must copy all EPO practice, procedure or methodology. It distinguished between adopting the same interpretation of article 52 and being bound to follow every aspect of EPO practice.

On the technology question, the court rejected the applicant’s main argument and held that ANNs of the kind in issue are programs for computers. It said that, whatever the hardware form, an ANN is in essence a set of instructions that causes a machine to process data in a particular way to produce a result.

The court explained that this remained true whether the ANN was implemented on a conventional digital computer, dedicated hardware, reconfigurable hardware, or hardware with fixed instructions. It also said that freezing weights and biases after training did not change the analysis. A trained ANN embedded in hardware was still, in substance, a program for a computer.

But that was not the end of the matter. The court also held that the claimed invention was not a program for a computer as such. So the claims did not fall foul of article 52 on the basis used by the Hearing Officer and the Court of Appeal.

The court did not try to complete the whole later-stage analysis itself. It said it would be unwise to go much further without fuller argument and the views of the UKIPO and specialist lower courts on how the intermediate step under G1/19 should work in practice.

That point matters. The judgment does not decide that claims involving technical means will always succeed, and it does not decide that this invention was novel or inventive. It only decides that the refusal on the computer-program exclusion ground was wrong.

The result was that the Hearing Officer’s decision was set aside and the application was remitted for consideration of novelty, inventive step and industrial application.

If you run an AI or software business, this case removes one argument and strengthens another. The weaker argument is that an ANN somehow sits outside the idea of a computer program. The Supreme Court has now said that is wrong.

The stronger argument is about the claimed invention as a whole. This case shows that an ANN-based claim is not automatically excluded at the threshold stage. But that does not mean every AI claim will pass. The court only decided that this application should not have been refused on the exclusion ground used.

That means your patent filing should not rely on branding language about intelligence, personalisation or machine learning. It should explain the technical setup clearly. In this case, the claims referred to an ANN, a database, a communications network, a network-connected user device, training data, property vectors and recommendation outputs.

Just as importantly, surviving the exclusion stage is only the front gate. The court expressly left novelty, inventive step and industrial application to the UKIPO. Businesses should not read this case as saying that AI inventions are generally patentable, or that adding standard hardware language will guarantee success.

The better reading is narrower and more useful. If your invention uses AI, you still need to identify what the claimed system does, how it is implemented, and what features may matter when the application reaches the later patentability stages.

For founders and product teams, the practical lesson is documentation. Patent outcomes often turn on how the invention is characterised and what the claims actually say. If your internal records only describe a commercial outcome, your advisers may struggle to build a strong filing strategy.

In a case like this, useful records would include how the ANN is structured, what training data is used, how similarity is measured, what physical properties are extracted, how outputs are generated, and how the wider system operates in production. The court’s reasoning shows that these details matter more than broad labels.

You should also separate three different questions in your own planning. First, what is the invention? Secondly, how is it implemented? Thirdly, what is new and non-obvious about it? This case mainly answers the first threshold question about exclusion. It does not answer the later patentability questions for you.

That separation can save time and cost. A business may have a technically detailed system description but still need a different body of evidence to show novelty and inventive step. Treat those as separate workstreams rather than assuming one explanation covers everything.

This case is important beyond AI. The Supreme Court said the Aerotel approach should no longer be followed. That matters for software-related patent analysis more generally.

At the same time, businesses should avoid overreading the judgment. The court adopted the same interpretation of article 52 as the EPO Enlarged Board in G1/19, but it also said that UK courts are not bound to replicate all EPO practice and methodology. So this is not a simple statement that UK procedure now mirrors the EPO in every respect.

The court also declined to go much further in applying the new intermediate analysis to the claims in this case. It said that would be unwise without fuller argument and input from the UKIPO and specialist lower courts. That means future cases may still shape how the new approach works in practice.

For businesses, the immediate takeaway is practical rather than theoretical. Review any older assumptions about software patent exclusions, especially if they were built around Aerotel-style reasoning. If you are filing now, make sure your advisers are framing the application under the post-G1/19 approach recognised by the Supreme Court.

A likely reader question is whether the patent was actually granted. The answer is no. The Supreme Court did not grant the patent and did not decide that the invention was fully patentable.

What the court did was remove one barrier. It held that the application should not have been refused on the computer-program exclusion alone. The case was then sent back so the UKIPO could consider the remaining patent tests, including novelty, inventive step and industrial application.

That distinction matters in practice. A business can win an appeal on excluded subject matter and still fail later if the invention is not new enough or is obvious in light of earlier technology. So this case is best read as a threshold-stage win, not a final patent success.

The judgment was given on 11 February 2026. The appeal was from the Court of Appeal decision reported as [2024] EWCA Civ 825. The Supreme Court allowed the appeal, set aside the Hearing Officer’s decision and sent the matter back for further consideration.

That makes the case a leading authority on the exclusion stage for AI and software-related patent claims in the UK. It is not authority that this particular invention was finally patentable or that a patent was granted.