In this second article of our AI Blogpost series “AI and the Law: A Journey of Innovation”, we look at the relationship between Artificial Intelligence (AI) and Intellectual Property (IP). IP rights, such as copyrights and patent rights, aim to offer protection and stimulation to the development of new – often innovative – products and services. In this article, we will investigate if AI models may also enjoy protection under IP rights, and when the development and use of AI models might infringe third party IP rights.
How do Artificial Intelligence & Intellectual Property rights interact?
The goal for an AI model is to simulate human-like intelligence: to recognize patterns, make decisions, make creations, predictions and/or judgments based on the information or the question it receives. Over the last years, we’ve seen the rise of “generative AI”, such as ChatGPT and Dall-E. These models have a focus on creating new content, such as text, video, images.
For an AI model to improve and have the desired outcome, the algorithm must be “trained”. This can be done in different ways, but often requires large data sets for the model to train on. This data may also be protected by IP rights.
Another important aspect of AI models is the “black box” element: even though the framework is created by AI developers, and the input and output may be known, an AI algorithm may partly remain a mystery, a “black box”. For that part of the process, even the developers are unaware of all the internal processes that generate the outcome.
As companies will often look towards IP rights to protect their intangible assets (such as creative works, software, inventions, etc.), several important questions arise when looking at these characteristics of AI models:
- Can you use (copyright) protected creations to train an AI model?
- Can an AI model be protected? And if so, under which type of IP right?
- Can the output of an AI model be protected? And if so, under which type of IP right?
These are the questions we aim to solve with the article. For this, we look at the existing Belgian and European legislation and caselaw. The proposed AI Act we discussed in our previous article [LINK] comes less into play, as it focuses primarily on the practical usage of AI systems. Nevertheless, the AI Act proposal does make clear that AI should not affect existing IP rights.
As this is a very complex matter, we focus on the two types of IP rights we deem most relevant for answering these questions: patent rights and copyright. But note that other IP rights and related rights (such as database rights and trade secrets) may also be relevant.
AI & Patents
A patent is an exclusive right granted for an invention, which allows the holder of the patent to prohibit another party from replicating, using or selling the patented product or method.
To obtain a patent, Belgian legislation and the European Patent Convention sets out certain conditions: the invention must be (i) new, (ii) inventive, (iii) industrially applicable and (iv) must be fully disclosed, meaning it must be described in a clear and complete way in such a way it can be replicated by a person skilled in the relevant technical field.
At the same time, a number of specific exclusions are mentioned, i.e. matters that cannot be patented. These include aesthetic creations, scientific theories and… programs for computers.
Can an AI Model be protected by a patent?
As computer programs are mentioned as a non-patentable matter in article XI.4 of the Belgian Code of Economic Law and article 52, paragraph 2 of the European Patent Convention, you will not be able to patent an AI model as such. However, caselaw has made it clear that together with a computer, computer network or programmable apparatus, software can be part of a patented invention (a so-called “Computer-Implemented Invention” or “CII”). A CII could for example be AI-software in an MRI, AI-software in 3D-printers or CNC machines or AI data security systems on a computer network.
The applicant might however encounter another hurdle: given the “black box” aspect of AI systems, the subject matter might bedifficult to fully disclose / describe. This lack of transparency hampers potential replication, which poses a problem to meet the patent requirements. Some aspects of the inner working of the model may even be beyond the limits of human understanding.
As such, an AI model can be patented as part of a Computer-Implemented Invention, if it can be sufficiently disclosed, and provided the invention meets all other patent requirements. As this is a very complex matter, the patent application will require close collaboration between the AI experts and the patent attorney.
Can the output of an AI model be protected by a patent?
In other words: what if the AI model fully describes a new and innovative invention?
This was the scenario of the ‘DABUS-cases’, in which the AI system ‘DABUS’ was designated as the inventor on patent applications. Several applications were made but, except for the South-African patent office, all applications were refused. The European Patent Office and its Legal Board of Appeals, the UK Court of Appeal, and the German Federal Patent Court all ruled that the application should be refused as an AI model cannot be the inventor.
As the creation of an innovative invention is a human activity, we conclude for now – until additional caselaw or legislation arises – that in the event there is no human intervention when making the invention, a European or Belgian patent cannot be granted.
However, in the event the AI model is used by a person as a way of improving the invention or as a tool to further develop the invention, we are of the opinion such invention could be patentable. The inventor should then be the person making use of the AI.
AI & Copyright
Copyright grants the author of a work exclusive moral and economic rights. The economic rights include the right the reproduce and adapt the work, as well as the right to communicate it to the public.
A work (text, image, design, code, music, etc.) can be protected by copyright provided it (i) has a certain form; and (ii) is original. In its landmark cases Painer and Infopaq, the Court of Justice has clarified how the criterium of ‘originality’ should be interpreted: a work is original when the author can make free and creative choices, and as such is able to stamp the work with his ‘personal touch’.
Copyright does not require an application, but comes into existence as soon as the requirements are met.
Can an AI model be copyright protected?
Software, computer code and applications, including software containing AI models, can be protected by copyright if the above-mentioned requirements are met. As such, the parts of the AI model that have a specific form (for example specific source code or object code) and that are “original” will be protected by copyright.
In contrast, the sections that fall within the “black box” of the AI model will not be able to qualify for copyright protection, as no they do not have a specific form. Of course, these are also the parts of the software that will be least easy for a third party to reproduce.
The concept of, or idea for a specific AI model (for example “a generative AI model for creating images”, “an AI model for reviewing contracts”, or “an AI model for examining CT scans”) cannot be protected under copyright, as an idea lacks a specific form.
Can the output of an AI model be copyright protected?
Do any rights come into existence when ChatGPT writes a beautiful poem or when Dall-E creates a complex image? And if so, who owns these rights?
Belgium has a human-centric approach when it comes to copyright: copyright will in first instance be owned by the author: the natural person creating the work. It can only be assigned or licensed to another party in a second stage (which can happen automatically/simultaneously following a legal provision or a contract).
The Court of Justice’s interpretation of ‘originality’ also leads us to believe that only a human can make the “free and creative choices” required for copyright to come into play.
Therefore, it is quite clear that a creation that is solely made by an AI model, without human intervention, will not meet the requirements for copyright protection. In a recent US case (Thaler vs Perlmutter, Order of 18 August 2023 in Civil Action No. 22-1564 (BAH)), the District Court of Columbia judged on just this matter, and came to the conclusion that a piece of visual art created solely a computer system (plaintiff’s “Creativity Machine”) did not result in any copyright, given the absence of human involvement (“human creativity is the sine qua non at the core of copyrightability, even as that human creativity is channeled through new tools or into new media”). As the underlying copyright principles are relatively similar, we believe that a Belgian court would reach a similar verdict.
At the same time, if an AI model was only used in a very limited manner (for example, only in the ideation phase to create a first set of ideas for a drawing or a story; or only in the final phase to color grade a video), and a human has made a vast amount of free and creative choices to create the work, it will be quite clear that the result can be copyright protected.
In between those two extremes, many different situations exist in which AI models are used to different degrees, and in different phases. Those situations require a case-by-case analysis in which IP specialists and ultimately national courts, will make an assessment, balancing the input of the AI model and the input of the author in the final work.
Taking all this into account, it will be essential to clearly identify and document the free and creative choices made when copyright might be contested, as it is the claiming party who needs to effectively demonstrate the existence of copyright protection.
Can I use (copyright) protected works to train an AI model?
In 2022, a new copyright exception for text and datamining was included in article XI.190 of the Belgian Code of Economic Law, transposing the Digital Single Market Directive.
This exception allows the reproduction of various copyright protected works to train an AI model, provided that (i) the works are legitimately accessible, and (ii) the copyright holder has not forbidden such use. For the works to be “legitimately accessible”, you must have legitimate access (for example a subscription to a database) or the data should be freely available. However, given the op-out system, authors may still expressly reserve these rights, in which case text and data mining will not be allowed without explicit consent from the owner of the rights to this data.
In this context, it is also worth noting that even if the reproduction of a protected work for text or datamining was legitimate, the original author’s copyrights will be infringed if the output generated by the AI model incorporates original elements from the work used to train the model (for example if the AI model reuses essential/characterizing features of the original work in the output).
The evolution of AI brings many interesting, but complex legal questions. We continue to monitor AI & IP developments closely, to keep you updated on the latest evolutions.
If you have any specific questions about AI & IP, don’t hesitate to reach out to Cresco’s innovation team.
Olivier Van Raemdonck, Partner
Ward Verwaeren, Senior Associate
Aida Kaloci, Associate
Emilie Van Heck, Associate