Reflection on artificial intelligence

Artificial Intelligence (or AI) raises many questions in relation to industrial property rights, both on the creation of a work or an invention, and on the protection of AI by IP rights.

What is AI?


But to speak well of AI, it will be necessary to explain a little what it is (and especially what it is not ...).

Artificial Intelligence is actually just a computer program, no more, no less. But a computer program a bit special: a human has not specifically programmed its operation, but this program has "learned" how it works by analyzing a large number of data (we will explain below) .

We must not see a particular "intelligence" behind these AIs: AIs only do what they have learned, without thinking.

Many people have sought to give a definition to RNs:

  • AIs are " computer programs which carry out tasks which are, for the moment, more satisfactorily performed by human beings, because they require high-level mental processes such as: perceptual learning, organization of memory and critical thinking »(Marvin Lee Minsky) ;
  • AIs are " computer programs that are given tasks that even the human can not really define the underlying rules " (Me)

Some examples of AI

The AI encompasses a very large number of concepts that I tried to group in the following diagram (which is clearly not complete):

All this to say that we should not mix everything: AI is not deep-learning and vice versa…

How some AI works

How regression methods work

Regression methods are mainly used to try to determine relationships between variables.

For example, suppose my variables are:

  • the age of a person;
  • the percentage of his white hair.

In this very simple example, just take a large number of people, analyze the hair and note their age and we will have a curve of this type.

Of course, this is not enough: you only have a scatter plot. Even if a human here sees the relationship that emerges overall, we must not forget that we can generalize the problem to 800 variables and, in this situation, it is a little more complex to visualize…

Therefore, we use so-called regression methods (linear, polynomial, logarithmic, etc.) to determine the curve or surface that best suits the scatterplot.

In the end, the red curve will give us the relationship between the different variables (and even the uncertainty concerning this relationship is magical).

How a neural network works

The underlying idea of a neural network is to "mimic" the functioning of our neurons, but in their most basic operation.

In a brain, each neuron has an axon that acts like an electrical wire, leading the nerve impulse (in the form of an action potential) to the neighboring neuron, thus ensuring the functional activity of the brain.

Nevertheless, if the sum of the nerve impulses arriving on the axon does not exceed a certain value (threshold of excitability of the neuron), the axon does not relay the nervous message: we thus have to make to a mechanism of thresholding.

Computer scientists did exactly the same thing, but replacing neurons in the brain with "logical blocks".

Thus, each neuron transmits values to its neighbors. The neuron receiving values will then sum these by weighting them with weights depending on the links (w1, w2, etc.). This sum will then be thresholded using a threshold function (as here the sigma function).

Of course, we do not have 3 neurons, but much more in practice. There are many neural network architectures, but here is what an implementation might look like (5 input parameters and 1 output parameter).

All the "intelligence" contained in the neural networks lies in the good determination of said weights. Thus, these weights will be determined by seeking to solve the following optimization problem: "Knowing a large number of input and output data corresponding, what are the weight values to maximize the response of the network.

To solve this optimization problem, the weights are often initialized in some way and we will try to modify them gradually to try to optimize the output.

As long as the result of the network is not the expected one (or its error rate is not lower than a predetermined threshold), we continue to look for values for the weights

(Yes, I know… my example is formally simplistic and therefore false for purists, but it is to explain simply).

Functioning of a “Deep-learning” neural network

In reality, a "deep-learning" type network is very conceptually very close to a neural network, but the number of layers is much higher.

It should be understood that Deep-learning networks have a very large number of weights or parameters, and it can be very complex to converge them during learning (ie solving the optimization problem can to be arduous).

Deep learning networks have nevertheless emerged in recent years, as computing power has increased significantly in recent years, notably thanks to the use of graphics cards.

Technical achievements related to AI


The patentability of inventions made by IA


It can happen that the AI "finds" technical solutions to problems that humans had long since posed.

These solutions may be in the medical field (eg identification of new molecules that may have a therapeutic effect) in the field of mechanics (eg identification of a particular profile of aircraft wing with a strong lift) or in n any other technological field.

Problem with the notion of invention

The question that can immediately be asked is: is an invention made by a computer an invention?

Fortunately, theA52 (1) EPC proposes a definition of the invention:

European patents are granted for any invention in all technological fields, provided that it is new, that it involves an inventive step and that it is capable of industrial application. 

Thus, we saw well that there is no difficulty: invention is defined by its application (ie new and inventive thing in a technological field) and not by its origin or its genesis.

Therefore, an invention can be described as an "invention" if its application is technological.

Problem with inventive step

Some have said that there is no inventive step for an invention made from AI because the inventor's effort was zero.

I can only disagree.

Indeed, the inventor's effort or the difficulty he had to invent are not relevant criteria for evaluating inventive step: the requirement of inventive step is based on the difficulty for a person skilled in the art of 'arrive at the invention from the documents of the prior art and not from the difficulty encountered, in practice, by the inventor.

For example, it wouldn't occur to anyone to say that an inventor cannot protect his invention because he found it by chance (and it happens quite often).

Of course, it would be possible to argue that the person skilled in the art is an AI (or a man using an AI) but this approach would inevitably lead to considering that all inventions are obvious (see "Everything is obvious" by Ryan Abbott).

Problem with sufficiency of description

Some have argued that an invention made by an AI would be insufficiently described, because we do not know the inventive process that brought the invention.

This is, it is true, a problem of AI: most often, they give a result, but struggle to explain why such a result is given.

I definitely disagree with this position because the sufficiency of description (A83 CBE) is not intended to describe how the invention has been conceptualized, but is intended to ensure that a third party, in light of the description, can realize the invention.

Therefore, there is no need to know why the invention works. It is sufficient for one skilled in the art to verify that the invention functions by realizing it.

Problem concerning the concept of inventor

The EPC has no definition of the concept of inventor.

However, the generally accepted interpretation is that the inventor, within the meaning ofA60 (1) EPC, is a "natural" person or a human (ie natural person).

Indeed, how to accept that "rights" to the invention belong to a machine or an algorithm, knowing that no national law (ie member states) provides a property attached to an entity other than a human.

But is it a real problem?

First of all, the EPC only requires the designation of an inventor. There is no penalty if the inventor is not the right one or if he does not even exist (see filing requirements in Europe).

But in truth, this is not even the subject for me: the real question is what is the inventive act and who is doing it.

Indeed, it is quite clear that an AI is not more than a computer program helping a human. The human will choose how to use the AI, will choose the data set to train it, will look at the output data to assess their relevance and applicability to the problem that it arises.

Therefore, there is no real difference with the use by this human of a simulator / a calculator that will help this person. It would not occur to anyone to say "the inventor did not make an invention because he used a computer to make his invention".

Thus, it is always possible to argue that the inventor is the person handling the AI in order to obtain the desired result.

Ownership issue

Regarding the ownership of the invention, we may ask whether the holder of the AI may file a patent application for an invention out of the hat of an AI.

In my opinion, and in European law, this does not pose much difficulty, because theA60 (1) EPC provides that the invention belongs to the inventor or his successor.

For all that (and even if we had to consider that the inventor is really a machine), the lack of right to file a patent (see inventions filed by an unauthorized person) can only be invoked by the real owner of the rights.

Similarly, in France, the invalidity of the patent for lack of ownership of rights (A138 (1) (e) EPC) is a relative nullity which can only be invoked by the real owner (see nullity in France).

So, to put it simply, nobody can question a patent on the pretext that the holder of the title would have stolen it from a machine…

Some examples

To show you that this is not a purely theoretical matter, we have recently seen several patent applications filed with the EPO for inventions made by an IA.

Some of these inventions were made by DABUS (which itself is registered… EP2360629 (A3)).

The first invention aims a container with fractal walls and to secure two containers in a simple manner.

The second invention aims a warning system (via the flashing of a diode) having a fractal repeat sequence and thus allowing better recognition by a human eye.

In another technical area, we can evoke this Gillette invention (EP1284621B1) concerning toothbrushes whose "tufts" intertwine in a particular way. And this, thanks to the " Creativity Machine »By S. Thaler…

The patentability of inventions implementing AI

Here, we are in the case where the invention is really in the implementation of an AI (eg image recognition with an AI).

The inventive concept can be located in several places:

  • in the particular selection of the data set for learning,
  • in the architecture of the neural networks used for a specific task,
  • in memory management when learning,
  • etc.

For this very different subject, I refer you to my article on mixed type inventions in Europe.

Protection of AI models

We have seen that inventions made by an AI or implementing an AI could be protected by patent law.

However, AI involves other entities such as the model (eg the configuration of neural networks).

The AI model is often very complex to obtain, because it requires a very specific selection of training data and requires a lot of effort (e.g. in terms of computing power or in terms of human effort to intelligently format the Input data).

So, how to protect them?

In my opinion, there are two tracks:

  • copyright protection of computer programs;
  • the protection of databases by an ad-hoc right.

Indeed, the fact that AI remains a computer program motivates the possible analogy with computer programs (ie software). What most resembles AI training is the compilation of source code into compiled code: AI training is a kind of compilation of a system whose purpose is to make this system conform to user expectations.

Therefore it is possible to consider that the AI model is protected by the provisions of copyright relating to software (L112-2 CPI).

I remember the definition of software given by the French Academy in its dictionary (9th edition) is a “ Structured set of programs performing a specific function, allowing the accomplishment of a given task". AI seems to fit this definition.

But it is also possible to make an analogy with the databases.

Indeed, as I said earlier, the model of an AI is a set of configuration / weight / etc. In other words, it is a kind of configuration parameter database.

Therefore, why not apply the ad-hoc right provided for in the article L112-3 CPI and the Directive 96/9 / EC of 11 March 1996 ?

According to this directive (and its interpretation by the CJEU), a database must have the following characteristics:

I think we can consider that the parameters of a model have all of these characteristics (the weights have a meaning even taken individually, the weights are arranged so as to know at which nodes (or links) they s 'apply, it is possible to navigate in the model to know the parameters).

The difficulty remains the protection of this database. Indeed, Article 7 of the Directive 96/9 / EC provides that protection of the content of the database may be granted " when obtaining, verifying or presenting this content attests to a substantial investment from a qualitative or quantitative point of view".

By obtaining, we speak not of the creation of the data but its acquisition. It must be recognized that in the training of the model, what is complicated is the creation of data and not its "acquisition".

However, if I had to argue, I would say that the training of the model is a kind of “verification” of the validity of the values of the configuration weights… Therefore, the protection of article 7 of the Directive 96/9 / EC does it apply?

Of course, I have no answer regarding the protection applicable to the model. The preceding elements are only leads. We will have to wait for the case law.

Artworks generated by AI


I am not an expert in the protection of works by copyright, but it is true that it would be a shame not to raise this point.

So I apologize in advance to the specialists for the subject of my approximations.

Some examples

Today, there are several “works of art” produced from AI:

Portrait of Edmond De Belamy
Landscape made by

But we also have music generated by artificial intelligence (example of a piece generated by Spotify Research):

Copyright protection?

Once we have seen these examples, it is questionable whether these "works" can be protected by copyright.

In French law, to benefit from a copyright (L111-1 CPI), it is necessary to verify that the work is "original" (jurisprudential criterion).

Basically, to be original, a work must include:

  • the imprint of the author's personality,
  • the mark of the intellectual contribution of the author and
  • the expression of the author's free and creative choices.

In my opinion, for the examples that we saw previously, this does not really pose a question: indeed, these examples were created by AIs which were configured (ie which have undergone training) in order to get a result wanted by the author.

Therefore, the previous examples clearly indicate the author's desire to obtain such a result.

We can consider that AI plays a role similar to software like Photoshop or as a mixer. AI simply simplifies the creation process.

Of course, you have to look at it on a case-by-case basis and I cannot make a single answer of the type "any image or sound coming out of an AI benefits from a copyright", but the use of an AI does not does not seem to exclude, by this simple fact, the application of copyright.

To conclude, I would simply like to recall that, when inventing photography, some wanted exclude this of "art". Indeed, they considered that photography belonged to the technical field.

Do not you think that we are in the same situation concerning AI?

Comments are closed.