Written by Patent Consultant Maria Laini.
In my previous article “Is Patenting AI the Smart Move? Weighing Legal and Strategic Value” (1), I discussed some of the main problems that inventors face (or should beware of) when trying to patent inventions based on the use of artificial intelligence. One of the main issues discussed was the necessity of a non-obvious technical effect.
In this article, I am going to dive deeper into:
- the issue of what is considered “technical” at the European Patent Office
- how this definition of the word can be different from what is commonly perceived
- why this causes so many hurdles and frustration when trying to patent AI tools.
I then give you five examples of AI patent applications, which were either approved or rejected by the EPO, and highlight some important decisions that were made in those cases.
What is an invention? It’s easier to say what isn’t!
As odd as it may seem, when the European Patent Convention (EPC) was first written, there was no unanimously approved definition of the term invention. Instead, a list of what cannot be considered an invention was written down in what is now Article 52, paragraph 2 of the EPC.
In particular, Article 52, paragraph 2 of the EPC states that a patent cannot be granted for inventions that are exclusively about the following:
- Discoveries
- Scientific theories
- Mathematical methods
- Aesthetic creations
- Schemes, rules and methods for performing mental acts, playing games or doing business,
- Programs for computers
- Presentations of information.
This does not imply that a patent can never be granted for an invention which includes one or more of these elements. If anything from this list is combined with technical features, then a patent could be obtained. However, one thing is certain: if, when describing your invention, you start using some of the buzzwords above, you will need to consider whether your application can be rejected, due to a lack of technical character.
In addition to the list above, a positive requirement provided by the EPO is that “the invention must be of both a concrete and a technical character”. (G-II, 1; T854/90). Beyond this, what we have is a vast number of very practical board decisions. A patent is refused, the applicant protests, the board reads the EPC and then publishes that: “invention X is not considered technical because of A, B and C”. This is what patent attorneys work with. We try to predict what examiners would be inclined to accept, based on examination proceedings of past applications and decisions taken in the relevant EPO courts, but we don’t have an exact recipe or formula to get your draft out of the woods, nor can we be absolutely certain that we cannot.
What does “technical” mean for the EPO?
The problem is that, as for the word “invention”, the EPO does not offer a detailed and precise definition of the word “technical” for the purposes of patentability assessment. Again, this may sound odd.
One thing is certain. In patent law, the adjective “technical” is used in a frustratingly narrower, although not precisely defined scope than in our everyday language.
The etymology of the word is Greek, and comes from tekhnē, which means “art”, “skill” or “craft”, which is extremely broad. The narrower sense of “of pertaining to the mechanical and professional arts; appropriate to a science, profession, or trade” comes from 1727, as illustrated in etymonline (2).
What does “technical character” mean at the EPO
Some hints on what the EPO means by “technical character” can, of course, be found in the list of exclusions of the EPC, in the court cases, and in the history of European patent law.
Except for the understandable need to exclude things that, although important, are not the direct product of human ingenuity and craft (e.g., scientific discoveries), there has always been a general agreement to exclude the so called “abstract ideas”, i.e., mere products of mental gymnastics such as games and mathematical formulations.
Then, there is the general hostility towards patenting “business methods”, which aims to draw a line between what is done to solve technical and concrete problems faced by humanity and what is done, put in a very reductionist way, solely and purely to make money.
And then there’s aesthetic creations, such as paintings, statues and music, for which the very idea of patenting them sounds just…wrong.
The rules of the game may seem easy, to some extent. A new song? Can’t patent it. A mysterious bird never captured on camera before? Can’t patent it. A cunning, new puzzle game idea? Can’t patent it.
But then came computer programs and then came artificial intelligence (AI). And things got more complicated.
Because the whole point of AI tools is to do what humans do, but better and faster. So, things like decision making, planning, presenting graphs, suddenly look very technical. Except they may not be, in the eyes of the EPO.
How do I know if my invention is technical by EPO standards?
In my job, I have seen many AI-based solutions that fall within grey areas of patentability exclusions. This is not surprising, as AI tools are often used to simplify decision making, calculations, classifications, analysis of complex data and pattern recognition (see what the EPO states about mental acts, in the list of exceptions above). Often, AI is used to innovate business practice, or financial services (look again at the list of exceptions above) and thus can be highly important for companies, but at the same time very hard to patent, especially in Europe.
For this type of inventions, the advice of specialised patent professionals is extremely valuable, as they are trained to thoroughly read the case law and monitor the most recent decisions to align their drafting techniques to the EPO’s expectations.
No patent attorney can promise a grant, but their expertise can minimise the chances of refusal as much as possible. This becomes especially true in case of refusal due to non-technical subject-matter.
Some practical examples
Because most of the guidelines we have is based on case law, it’s worth discussing some important decisions taken at the EPO courts. These cases are a core part of the guidelines for patenting artificial intelligence that have been recently published by the EPO (3).
Two approved AI patents
A heartbeat monitoring apparatus
This first case (4) concerns an invention for monitoring a heartbeat. The method relies on a neural network, more specifically a Kohonen neural network, to determine whether an electrocardiograph signal of a heartbeat is regular or irregular.
This method was the first method to use Kohonen neural networks applied to heartbeat monitoring. However, this was not the only reason why the EPO found the invention to be patentable. The technical character of this solution lies in the comparison of an input vector (containing data of the heartbeat that needs monitoring) with two sets of reference vectors, to first get rid of vectors indicative of spurious irregular heartbeats before the second comparison is carried out (see paragraph 4.2.3 of the decision).
This particular configuration was found to improve the signal to noise ratio and consequently reduce the number of false identifications of irregular heartbeats. This was deemed to be technical, by EPO standards. This decision is aligned with EPO guidelines, which find that “technical input and technical output are typically achieved through direct links with physical reality” (5).
From this first case, we can learn that the use of a neural network can be technical, when it is tied to specific steps that are taken to achieve a real and demonstrable technical effect. In this case, the effect was to reduce spurious irregularities in monitored heartbeats.
Image recomposition methods to improve scene classification
The second case (6) I am going to discuss concerns a computer implemented method to improve image classification of a digital image. The method is based on a semantic classifier, which is trained on several systematically created “recomposed versions” of an exemplar image, to provide more accurate scene classification. Thus, the method is aimed at improving the performance of a classifier. Interestingly, in this specific appeal case, the board did not even discuss the possibility that the claims were not technical. The case was rather focused on the comparison between the claims and prior art arrangements.
What we can learn from this case is that inventions aimed at improving known software tools (such as classifiers) can be considered technical. In these cases, it is important to have evidence that there is an improvement in performance, and that said improvement is not trivial to achieve through the claimed features.
Three examples of rejection
Simulation of pedestrian movements in an environment
The third court case (7) I am going to discuss relates to a computer-implemented method that simulates the movement of a pedestrian in an environment. One of the goals of this invention was to find insights on how people move, either individually or in crowds, so that buildings can be designed to be safe, for instance, to evacuate crowds. This is a computer-program, but it has a very concrete and practical use, which has to do with buildings, that are made of bricks and are a very concrete thing indeed.
However, this invention was ultimately deemed unpatentable. Why? Because the claims were limited to a simulation, i.e., sequence of computer-implemented steps to simulate movement of a pedestrian but did not recite any feature which would be based on any result of said simulation for a specific technical task, such as for the construction of a building for safe crowds. The method was only describing a simulation as such, that could have been used, in principle, for anything.
This decision is now referred to in standard EPO practice, when considering the patentability of computer-implemented simulation techniques.
The first lesson that we can learn from this case is that simulating a physical object is not, in itself, sufficient to confer technical character to an invention.
The second lesson is that it is not enough to state that the simulation runs on a computer and that is “for (insert here a very technical and concrete use)”. Intended use is not, in this sense, limiting the scope of the claim and would be disregarded when assessing the invention.
A better claim may recite a step where an output of the simulation is extracted, such as a particular parameter related to the dispersion of a crowd and then used to derive specific building parameters of a building which needs to accommodate and safely evacuate a crowd.
A neural-network-based machine translation (NMT)
This fourth case (8) relates to a computer program that exploits neural networks to translate rare words in texts. The aim of the invention was to improve the quality of translations, and more specifically to overcome the limitations of previous NMT machines that relied on small-sized and limited vocabulary.
The application relies on specific algorithmic features, such as the emission, by the neural network model, of “pointer tokens” and “null unknown tokens”, which are used to track the origin of unknown words in sentences.
The patent application was refused, on the basis that it lacked a technical effect. This is because “the translation of a text …is a matter of linguistics and not a technical effect”. Notably, the board highlighted that this is so “even if the computer program includes algorithmic aspects which are not directly based on linguistic concepts”, such as the use of the tokens briefly introduced above.
A couple of lessons can be learnt from this decision. The first, perhaps unsurprising, is that translation tools, even if they rely on neural networks, are per se not technical in the eyes of the EPO. This derives from the fact that translation from one language to another is a mental act, which falls within the list of exclusions from patentability discussed at the beginning of this article. The second lesson is that the mere addition in the claims of algorithmic features run on a computer may not be enough to make the claims technical.
A neural network apparatus for hierarchical learning (i.e., a mimicked “human brain”)
The fifth and last case (9) I am going to discuss is a particularly interesting one, because it concerns a neural network that aims to solve the problem of overfitting. This is done through the establishment of “loose” connections in the neural network, prior to its training, independently of the learning data, and according to a check matrix of an error correcting code (see Figure 4 and para. 10 of the application (10)).
Whilst the concept of “loose connections” in neural networks to prevent overfitting was known at the time this application was filed, determining said “loose connections” based on the check matrix of an error correction code was found to be new with respect to previous systems. Because the technical effect must be based on the new features in the claims, this is the feature that was evaluated. Unfortunately for the Applicant, the EPO found this distinguishing feature to be non-technical.
The Applicant argued that creating loose connections in the neural network through this check matrix was technical, because this feature solved the problem of improving the learning capability and efficiency of a machine. This effect itself was known and supported by evidence from scientific papers. The Applicant argued that this feature provided a machine that could “mimic the human brain” by “replicating biological optimisation” and that, for this reason, could replace a human in handling complex tasks (see para. 6 of T 702/20 (11)). A point the Appellant relied on was the acknowledgement, by previous case law, that automation of tasks is generally recognised as a technical problem.
This case is complex, but there are two very interesting points to take from this decision:
- The first one is that the neural network structure, which is new because of the particular way in which the loose connections are determined, only defines a class of mathematical functions, which cannot be patented on their own.
- The second interesting point is the rejection of the Applicant’s argument that the invention was mimicking the human brain. The board remained unimpressed by this statement and concluded that there was no sufficient evidence that this was the case.
Notably, this case contains many precious teachings. The first one is that a complex bundle of mathematical functions is still, at the end of the day, a mathematical tool, and thus it is not per se technical. Furthermore, calculating something that could in principle be written on paper in a faster fashion through computer automation, is not sufficient for obtaining a patent at the EPO. In this case, the only technical element would be the presence of the computer, which is simply not clever enough.
Finally, another important point is that it is not enough to rely on the automation argument. In other words, claiming that the computer can replace the human in “mental acts” is not enough to obtain a patent.
Conclusion
With this small glimpse into the vast case law that we now have about software and AI patents, I hope I conveyed the message that this is a very intricate and complex field.
It’s worth concluding this article with the words of the Enlarged Board, in paragraph 6.3 of their G1/19 decision (12), which state that it is “never possible to give an exhaustive list of (positive or negative, alternative or cumulative) criteria for assessing whether a computer-implemented process solves a technical problem”.
This statement highlights the complexity and the lack of a formulaic, one-size fits all approach when evaluating the patentability of software (and thus AI) inventions.
The best that we can do is to look thoroughly at each individual case, and compare it with the most relevant case law, if any, that we can find, to anticipate the EPO’s moves during examination and beyond.
Citations
(1) https://www.patrade.com/knowledge/is-patenting-ai-the-smart-move-weighing-legal-and-strategic-value
(2) https://www.etymonline.com/word/technical
(3) https://www.epo.org/en/legal/case-law/2025/clr_i_d_9_2_12_e.html
(4) https://www.epo.org/en/boards-of-appeal/decisions/t070598eu1
(5) Para. 85 of https://www.epo.org/en/boards-of-appeal/decisions/g190001ex1
(6) https://www.epo.org/en/boards-of-appeal/decisions/t091286eu1
(7) https://www.epo.org/en/boards-of-appeal/decisions/t140489eu2
(8) https://www.epo.org/en/boards-of-appeal/decisions/t201903eu1
(9) https://www.epo.org/en/boards-of-appeal/decisions/t200702eu1
(10) https://worldwide.espacenet.com/patent/search/family/053777512/publication/EP3089081A1?q=14882049.1
(11) https://www.epo.org/en/boards-of-appeal/decisions/t200702eu1
(12) https://www.epo.org/en/boards-of-appeal/decisions/g190001ex1
