In May 2021, I had the pleasure of being invited to St. Petersburg again (this time only virtually, though) to talk about the “Patentability of Artificial Intelligence Inventions in Europe”.

For me, one advantage of online conferences is that it’s very easy to make a video recording. On the other hand, I found out that most of my followers actually like their content in text form, so I’ve finally turned the talk into an article:

Is AI patentable in Europe? The two-hurdle approach

The answer to the question of AI’s patentability in Europe requires a short reference to the basics. To understand the fundamentals of European patent law in connection with patents for digital inventions, you have to be familiar with the two-hurdle approach at the European Patent Office (EPO).

The EPO's two-hurdle approach

The two-hurdle approach involves two patent hurdles:

  1. Patent eligibility: This first hurdle involves the rather dogmatic question whether something should be susceptible to the patent system in the first place. At the EPO, this is an absolute hurdle, i.e. it is decided without looking at what already exists.
  2. Patentability: This second hurdle is a relative one, where the invention has to be novel and non-obvious (aka “involve an inventive step”) compared to the prior art, i.e. everything which is publicly known already.

The first hurdle? Easy! About patent eligibility

The first hurdle (patent eligibility) is set quite low at the EPO as it just requires that the claimed invention as a whole has technical character. One single technical feature in the patent claim is enough.

In practice, this means:

  • Apparatus claims (which refer to some physical entity, e.g. a machine or a smartphone) are automatically patent eligible.
  • Method claims (which are directed to a process) are patent eligible already if the process uses technical means. The simple statement “computer implemented method” in the patent claim is enough at the EPO, because these words indicate that the method necessarily uses technical means (e.g. a computer).

In short, the first hurdle is really a non-issue in practice. Only completely abstract ideas or theoretical constructs fail at this point while most inventions pass. Note that this is quite different from the U.S., where there are lengthy discussions about how hard or easy it should be to take the patent eligibility hurdle.

The second hurdle? Tricky! About patentability

Once the invention is patent eligible, the second hurdle is patentability. Here, software-based inventions are particularly difficult to prosecute, depending largely on the technological field and the individual case. At the second hurdle, the quality of the invention is compared to the prior art. The decisive part is regularly the inventive step, but there is a specific twist to this assessment which needs some further explanation.

This is because computer-based inventions often involve a mix of technical and non-technical aspects, which raises the question which of these aspects should count towards inventive step.

Enter the COMVIK approach, the EPO’s long-established framework for assessing the patentability of mixed-type inventions:

The COMVIK approach at the European Patent Office

Let’s assume the prior art and the invention have a certain overlap and – hopefully – a “delta” (part of the invention not yetknown in the prior art), so that the invention is novel. The key question at the second hurdle is: Is this delta over the prior art non-obvious?

But: According to the COMVIK approach, inventive step can only be based on those features of the invention which contribute to the solution of a technical problem.

By the way, did you know that I write a short daily email with practical tips on how to make sense of patents in the digital transformation. Yes it’s daily, and yes people actually read it! Get on the list to not miss future updates.

Accordingly, the EPO examiner looks at the delta, i.e. the distinguishing features, and decides for each one whether it is

  • an inherently technical feature (e.g. a physical thing, a piece of hardware, a technical measurement value) or
  • an “as such” non-technical feature (e.g. a mathematical formula).

Then, only the inherently technical features enter the non-obvious assessment, as well as those non-technical features which still contribute to the solution of a technical problem. In other words, the delta over the prior art not only needs to be non-obvious, but also in a technically meaningful way.

Which leads to the following key takeaway of how the EPO assesses software patents:

European patents are all about the non-obvious technical contribution.

Which in turn leads directly to the follow-up question:

Is AI something “technical”?

If the answer is “no”, AI aspects would be disregarded in the inventive-step assessment. If the answer is “yes”, they would enter into the equation.

In the current EPO practice, the answer is an “it depends”, and at this point it gets complicated.

First of all, there is no legal definition of the term “technical” in the European Patent Convention (EPC). And for a good reason. The creators of the European patent convention decided that the notion of “technicality” should be flexible to adapt to new developments as technology evolves.

So in the absence of an exhaustive definition, we must consider the case law of the Boards of Appeal (the judicial instance of the European patent system). However, if you search in the public database of decisions of the Boards of Appeal for “machine learning”, the query returns only ten decisions.

By the way, I can talk you through most of these decisions in episode #2 of my podcast:

A second source of information besides the case law are the Guidelines for Examination in the European Patent Office. That’s basically a textbook and the “user manual” for every European patent examiner. In 2018, the EPO added a specific section about AI and machine learning.

The new section says pretty clearly that the computational models and algorithms underlying AI and machine learning “are per se of an abstract mathematical nature, irrespective of whether they can be “trained” based on training data”. This is an important statement, because this means that the computational foundations of AI – e.g. a neural network, a classification algorithm or a training method – in themselves are mathematical and therefore not technical.

AI "as such" is not patentable at the EPO

Is the conclusion then that AI is not patentable?

The mathematical foundations underlying AI may not be patentable as they are considered to be non-technical. Bthis is not the end of the story, because there are still two ways how this – as such non-technical subject matter – can still contribute to the technical character of an invention:

  1. technical application and
  2. technical implementation.

Machine control and voice assistants About technical applications of AI

As soon as a new machine learning approach is applied to the solution of a technical problem, it is found by the EPO to contribute to the technical character of the invention and thus enters into the inventive-step assessment.

Positive examples of technical AI applications

Some positive examples for technical applications:

  • Machine control, the AI is applied to control a technical device, e.g. a self-driving car, an autonomous drone or a robot.
  • Image processing applications like face recognition, identification of people in videos or medical analysis (where AI looks at pictures of human tissue and identifies cancer cells).
  • Speech and language processing like a smartphone’s voice assistant or machine translation.

Here is a concrete example for a speech recognition patent. A new kind of neural network in itself would simply be regarded as an abstract math construct by the EPO. Even if it is completely unheard of, i.e. “novel”, it would not lead to a patent.

EP 3 078 020 B1

But in this example case, the neural network was specifically applied to speech recognition, which is considered to be a technical use case at the EPO. As the neural network with its novel structure was applied to a technical domain (speech recognition), its features were subject to the inventive step assessment and in the end acknowledged to be non-obvious. The patent was granted.

Negative example of an AI application: Classification algorithms that classify something not based on “hard” technical characteristics, but on the textual content of the information (e.g. a spam filter). In this prominent appeal case from Microsoft, an AI based spam filter was proposed which scans the content (the text) of e-mails and decides whether they should be moved to the inbox or to the spam folder.

The Board of Appeal decided that the machine learning part in this case did not contribute to the solution of a technical problem is was based not on technical, but on linguistic factors. In addition to that, the Board took the view that it is rather subjective whether some text is considered spam or not. In the ends, the spam filter invention was regarded to be non-technical and therefore could not contribute to an inventive step.

#6 Patents for AI inventions in Europe (🔴Conference talk recording) The Best Practice

  1. #6 Patents for AI inventions in Europe (🔴Conference talk recording)
  2. #5 G1/19 first reactions w/ Rudolf Teschemacher and many more
  3. #4 Novelty w/ Patrick Heckeler

Did you know that my podcast is available on all major podcast platforms? If you’re enjoying this article, chances are you’ll also like the podcast!

CPU/GPU interplay About technical implementations of AI

The second dimension apart from technical application is technical implementation. When an AI algorithm is designed such that it takes into account the specifics of the underlying hardware, this also makes a technical contribution.

Patentable AI implementations

In this example, certain instructions are calculated on the CPU and others on the GPU. The CPU initiates a training algorithm for a neural network, and then sends certain inputs to the GPU, because the inventors of this patent observed that the CPU is good for tasks that require lots of input and output while the GPU is particularly good for computations.

EP 1 569 128 B1

The GPU then sends certain error observations back, based on which the CPU can adapt the learning parameters. This interplay between CPU and GPU provides a specific technical implementation of an AI improvement, and was found to be patentable.

These arguments, on the other hand, are regularly not successful at the EPO:

  • High mathematical complexity of the AI as such is not a persuasive argument for the Boards of Appeal.
  • Pure automation. The simple fact that AI automates something that had been done by a human doesn’t make it inventive.

So, is AI patentable then?

To sum up the patentability question, these are the key takeaways:

  • There are no European patents for the mathematical foundations of AI, as long as purely theoretical advancements of AI are concerned.
  • But patents are available for “applied AI” which is either used in a technical context or technically implemented in a specific way.

So, finally, the key message and answer to the opening question is:

AI is patentable at the EPO

Yes, AI is patentable at the EPO if it is used in a technical context or if it is implemented in a clever way.

If you enjoyed this article, you’ll love my mailing list — I write a short daily email with practical tips on how to make sense of patents in the digital transformation. Yes it’s daily, and yes people actually read it! Join the inner circle now 👇

Subscription received!

Please check your email to confirm your newsletter subscription.

I use Revue to send these emails. By subscribing, you acknowledge that your information will be transferred to Revue for processing. Check out Revue’s privacy policy here.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.