Artificial Intelligence (or AI) raises numerous questions in relation to industrial property rights, both concerning the creation of a work or an invention and the protection of AI through intellectual property rights.
What is AI?
Definition
To properly discuss AI, it is necessary to explain what it is (and especially what it is not).
Artificial Intelligence is, in reality, nothing more than a computer program—no more, no less. However, it is a somewhat special computer program: a human has not specifically programmed its operation, but rather the program has « learned » its operation by analyzing a large amount of training data (we will explain this below).
There is no need to see any particular « intelligence » behind these AIs: AIs only do what they have learned, without any reflection.
Many people have attempted to define AI:
- AIs are « computer programs that engage in tasks which, for the time being, are performed more satisfactorily by human beings, as they require high-level mental processes such as perceptual learning, memory organization, and critical reasoning » (Marvin Lee Minsky);
- AIs are « computer programs entrusted with tasks for which even humans cannot truly define the underlying rules » (Me).
Some examples of AI
AI encompasses a very large number of concepts, which I have attempted to group in the following diagram (which is clearly not exhaustive):
All this to say that we must not confuse everything: AI is not deep learning and vice versa…
How some AIs work
How regression methods work
Regression methods are primarily used to try to determine relationships between variables.
For example, let us assume that my variables are:
- a person’s age;
- the percentage of their gray hair.
In this very simple example, it is sufficient to take a large number of people, analyze their hair, and record their age, and we will obtain a curve like this.
Of course, this is not sufficient: all you have is a scatter plot. Even if a human can generally see the emerging relationship here, we must not forget that the problem can be generalized to 800 variables, and in this situation, it becomes a bit more complex to visualize…
Therefore, we use so-called regression methods (linear, polynomial, logarithmic, etc.) to determine the curve or surface that best fits the scatter plot.
Ultimately, the red curve will give us the relationship between the different variables (and even the uncertainty regarding this relationship—it’s magical).
How a neural network works
The underlying idea of a neural network is to « mimic » the way our neurons function, but at their most basic level.
In a brain, each neuron has an axon that acts like an electrical wire, conducting the nerve impulse (in the form of an action potential) to the neighboring neuron, thus ensuring the brain’s functional activity.
However, if the sum of the nerve impulses arriving at the axon does not exceed a certain value (the neuron’s excitability threshold), the axon does not relay the nerve message: this is a thresholding mechanism.
Computer scientists have done exactly the same thing, but by replacing the brain’s neurons with « logic blocks. »
Thus, each neuron transmits values to its neighbors. The neuron receiving the values will then sum them, weighting them with weights depending on the connections (w1, w2, etc.). This sum is then thresholded using a threshold function (such as the sigma function here).
Of course, we do not have just 3 neurons, but many more in practice. There are numerous neural network architectures, but here is what an implementation might look like (5 input parameters and 1 output parameter).
All the « intelligence » contained in neural networks lies in the proper determination of these weights. Thus, these weights are determined by attempting to solve the following optimization problem: « Given a large number of input and corresponding output data, what are the weight values that maximize the network’s response. »
To solve this optimization problem, the weights are often initialized randomly, and then progressively modified to optimize the output.
As long as the network’s result is not the expected one (or its error rate is not below a predetermined threshold), the search for weight values continues.
(Yes, I know… my example is formally simplistic and therefore incorrect for purists, but it is to explain simply).
How a « Deep Learning » neural network works
In reality, a « Deep Learning » network is conceptually very similar to a neural network, but the number of layers is much higher.
It is important to understand that « Deep Learning » networks have a very large number of weights or parameters, and it can be very complex to make them converge during training (i.e., solving the optimization problem can be challenging).
« Deep Learning » networks have nevertheless emerged in recent years because computing power has greatly increased, particularly due to the use of graphics cards.
Technical implementations related to AI
Patentability
Patentability of AI-generated inventions
Principle
It may happen that AI « finds » technical solutions to problems that humans have been trying to solve for a long time.
These solutions may be in the medical field (e.g., identification of new molecules that may have a therapeutic effect), in the mechanical field (e.g., identification of a particular aircraft wing profile with high lift), or in any other technological field.
Issue concerning the concept of invention
The question that immediately arises is: is an invention made by a computer an invention?
Fortunately, A52(1) EPC provides a definition of an invention:
European patents shall be granted for any invention, in all fields of technology, provided that it is new, involves an inventive step and is susceptible of industrial application.
Thus, it is clear that there is no difficulty: an invention is defined by its application (i.e., a new and inventive thing in a technological field) and not by its origin or genesis.
Therefore, an invention can perfectly well be qualified as an « invention » if its application is technological.
Issue concerning the inventive step
Some have argued that there is no inventive step for an invention created using AI because the inventor’s effort was non-existent.
I strongly disagree.
Indeed, the inventor’s effort or the difficulty they encountered in inventing are not relevant criteria for assessing the inventive step: the requirement for an inventive step is based on the difficulty for a person skilled in the art to arrive at the invention from the prior art documents, not on the difficulty actually encountered by the inventor.
For example, no one would think of saying that an inventor cannot protect their invention because they discovered it by chance (which happens quite often).
Of course, it could be argued that the person skilled in the art is an AI (or a human using an AI), but this approach would inevitably lead to the conclusion that all inventions are obvious (see « Everything is obvious » by Ryan Abbott).
Issue concerning sufficiency of disclosure
Some have argued that an invention made by an AI would be insufficiently disclosed because the inventive process leading to the invention is unknown.
This is, admittedly, a problem with AIs: most often, they provide a result but struggle to explain why such a result is given.
I firmly disagree with this position because sufficiency of disclosure (A83 EPC) does not aim to describe how the invention was conceived but ensures that a third party, based on the description, can carry out the invention.
Therefore, there is no need to know why the invention works. It is sufficient that the person skilled in the art can verify that the invention works by carrying it out.
Issue concerning the concept of inventor
The EPC does not provide any definition of the concept of inventor.
Nevertheless, the generally accepted interpretation is that the inventor, within the meaning of A60(1) EPC, is a « natural person » or a human (i.e., natural person).
Indeed, how can it be accepted that « rights » over the invention belong to a machine or an algorithm, given that no national law (i.e., of the member states) provides for ownership attached to an entity other than a human.
But is this a real problem?
First, the EPC only requires that an inventor be designated. There is no sanction if the inventor is not the correct one or if they do not even exist (see the requirements for filing in Europe).
But in truth, this is not even the point for me: the real question is to determine what the inventive act is and who is the person performing that act.
Indeed, it is quite clear that an AI is nothing more than a computer program assisting a human. The human chooses how to use the AI, selects the dataset for training it, reviews the output data to assess its relevance and applicability to the problem at hand.
Therefore, there is no real difference from the use by that human of a simulator or a calculator to assist them. No one would think of saying « the inventor did not make an invention because they used a computer to create it. »
Thus, it is always possible to argue that the inventor is the person operating the AI to obtain the desired result.
Issue concerning proprietorship
Regarding the proprietorship of the invention, we may question whether the proprietor of the AI can file a patent application for an invention generated by an AI.
In my view, and under European law, this does not pose much difficulty, as A60(1) EPC provides that the invention belongs to the inventor or their successor in title.
Nevertheless (and even if we were to consider that the inventor is truly a machine), the lack of entitlement to file a patent (see patents filed by an unauthorized person) can only be invoked by the true proprietor of the rights.
Similarly, in France, the nullity of the patent for lack of entitlement (A138(1)(e) EPC) is a relative nullity that can only be invoked by the true proprietor (see nullity in France).
So, in short, no one can challenge a patent on the grounds that the proprietor of the patent allegedly stole it from a machine…
A few examples
To clearly demonstrate that this is not purely a theoretical issue, we have recently seen several patent applications filed with the EPO for inventions created by an AI.
Some of these inventions were created by DABUS (which itself is filed… EP2360629 (A3)).
The first invention relates to a container with fractal walls, thereby enabling simple interlocking of two containers.
The second invention relates to an alert system (via the flashing of a diode) with a fractal repetition sequence, thereby enabling better recognition by the human eye.
In another technical field, we can mention this Gillette invention (EP1284621B1) concerning toothbrushes with bristles intersecting in a particular manner. This was achieved thanks to S. Thaler’s « Creativity Machine« …
Patentability of inventions implementing AI
Here, we are considering cases where the invention truly lies in the implementation of an AI (e.g., image recognition using AI).
The inventive concept may reside in several aspects:
- in the specific selection of the training dataset,
- in the architecture of the neural networks used for a specific task,
- in memory management during training,
- etc.
For this very different topic, I refer you to my article on mixed-type inventions in Europe.
Protection of AI Models
We have seen that inventions created by an AI or implementing an AI could be protected under patent law.
However, AI involves other entities such as the model (e.g., the configuration of neural networks).
The AI model is often very complex to obtain, as it requires a very specific selection of training data and demands significant effort (e.g., in terms of computing power or human effort to intelligently format the input data).
So, how can they be protected?
In my view, there are two possible approaches:
- the protection of computer programs under copyright law;
- the protection of databases by a sui generis right.
Indeed, the fact that AI remains a computer program justifies the possible analogy with computer programs (i.e., software). What most closely resembles the training of an AI is the compilation of source code into compiled code: the training of an AI is a kind of compilation of a system whose purpose is to make this system conform to the user’s expectations.
It is therefore possible to consider that the AI model is protected by the provisions of copyright law relating to software (L112-2 CPI).
I recall that the definition of software given by the Académie Française in its dictionary (9th edition) is a « Structured set of programs fulfilling a specific function, enabling the accomplishment of a given task« . AI seems to fit well within this definition.
However, it is also possible to draw an analogy with databases.
Indeed, as I mentioned earlier, an AI model is a set of configurations/weights/etc. In other words, it is a kind of database of configuration parameters.
Therefore, why not apply the sui generis right provided by Article L112-3 CPI and 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 believe that the parameters of a model can indeed be considered to have all these characteristics (the weights have meaning even when taken individually, the weights are arranged so as to know to which nodes (or links) they apply, and it is possible to navigate the model to know the parameters).
The difficulty remains the protection of this database. Indeed, Article 7 of Directive 96/9/EC provides that protection of the contents of the database may be granted « where the obtaining, verification or presentation of that content demonstrates a substantial investment in either qualitative or quantitative terms« .
By obtaining, we refer not to the creation of the data but rather its acquisition. It must be acknowledged that in the training of the model, what is complex is the creation of the data, not its « acquisition ».
Nevertheless, if I were to argue, I would say that the training of the model is a kind of « verification » of the validity of the configuration weight values… Therefore, does the protection under Article 7 of Directive 96/9/EC apply?
Of course, I do not have an answer regarding the applicable protection for the model. The above elements are merely avenues to explore. We will have to wait for case law.
Works Generated by AI
Introduction
I am not a specialist in the protection of works under copyright law, but it is true that it would be a shame not to address this point.
So, I apologize in advance to specialists on the subject for my approximations.
A Few Examples
Today, there are several « works of art » produced using AI:
We also have music generated by artificial intelligence (example of a track generated by Spotify Research):
Protection by Copyright?
Having seen these examples, one might wonder whether such « works » can benefit from copyright protection.
Under French law, to benefit from copyright protection (L111-1 CPI), it is necessary to verify that the work is « original » (a jurisprudential criterion).
In essence, for a work to be original, it must exhibit:
- the imprint of the author’s personality,
- the mark of the author’s intellectual contribution, and
- the expression of the author’s free and creative choices.
In my view, for the examples we have seen, this does not pose a real issue: indeed, these examples were created by AI systems that were configured (i.e., trained) specifically to achieve a result intended by the author.
Thus, the preceding examples clearly reflect the author’s intent to achieve such a result.
We can consider that AI plays a role similar to software like Photoshop or a mixing console. AI simply streamlines the creative process.
Of course, each case must be examined individually, and I cannot provide a blanket answer such as « all images or sounds produced by AI benefit from copyright, » but the use of AI does not, in and of itself, preclude the application of copyright.
In conclusion, I would simply like to recall that, when photography was invented, some sought to exclude it from « art. » Indeed, they considered photography to belong to the technical domain.
Do you not think we are in a similar situation regarding AI?