Artificial intelligence (AI): The qualification of AI creations as “works” under EU copyright law

Artificial Intelligence (“AI”) systems are capable of creating a wide range of artistic, literary and musical works with limited human intervention. In the future, AI systems might be capable of creating output without any human intervention. However, EU copyright law today is centered around the original author as a human being and not an AI-driven machine. In order to have copyright, two conditions must be fulfilled: (i) the creation must be a “work”, as defined in the case-law of the Court of Justice of the European Union (“CJEU”), and (ii) one must be the original author of the mentioned work or have obtained the copyright by transfer.

Section 1 | Introduction and outline of this article

The authorship of AI creations has already been the subject of much debate and some interesting articles have been published about this topic (1). However, the first condition about the qualification of an AI creation has been underexposed so far; can AI creations qualify as a “work” to start with? And, can AI creations even be protected under EU copyright?

An interesting example of the issue of copyright protection of AI creations is illustrated in the following case before the US Copyright Office (“USCO”). On February 14, 2022, the Copyright Review Board of the USCO refused the application for copyright protection
of an AI creation in the form of a two-dimensional artwork entitled “A Recent Entrance To Paradise” (see image).

The work was created by an AI system named “Creativity Machines”. The Applicant, Dr. Steven Thaler, is the inventor of Creativity Machines (2).

The work in question stems from a larger project involving Dr. Thaler’s experiments to design artificial neural networks simulating the creative activities of the human brain.

“A Recent Entrance to Paradise” is one in a series of images generated and described in text by the “Creativity Machine” as part of a simulated near-death experience Dr. Thaler undertook in his overall research into an invention of artificial neural networks (3).

Dr. Thaler, the applicant for the registration of the concerned work, stated that the work “was autonomously created by a computer algorithm running on a machine”, and he was “seeking to register this computer-generated work as a work-for-hire to the owner of the Creativity Machines”. The Copyright Review Board noted that the work in question was created by artificial intelligence without any creative contribution from a human actor. In accordance with the case-law of the US Supreme Court, it denied the registration on the basis that human authorship is an essential element of copyright protection (4).

In this article, we will examine whether, and under what conditions, AI creations such as “A Recent Entrance To Paradise” can qualify as “works” under the current EU copyright framework.

First, we address the distinction between AI-generated output and AI-assisted output (Section 2). Then we discuss the concept of “work” under EU copyright law, as defined in the case-law of the Court of Justice of the European Union (Section 3). Next, we discuss the four-step test proposed by the European Commission in order to determine whether AI-assisted outputs and AI-generated outputs can qualify as “works” under EU copyright law (Section 4). In our conclusion, we argue that the current legal framework in the EU is generally suitable to assess whether AI-assisted output can qualify as a “work”. However, this is not the case for AI-generated output (Section 5).

Section 2 | Output: AI-assisted versus AI-generated

There is a general tendency of making a distinction between “AI-assisted output” and “AI-generated output”. This distinction is confirmed by the World Intellectual Property Organization (“WIPO”) (5), the European Commission (6) and the European Parliament (7):

1. AI-generated output refers to the generation of an output by AI without any human intervention. In this case, AI can change its behavior during operations to respond to unanticipated information or events. According to the European Commission (“EC”), there are no examples of AI-generated works at this time (8).
2. AI-assisted output is generated with material human intervention and/or direction. AI-assisted output can be defined as outputs, applications or productions generated by or with the assistance of AI systems, tools or techniques, such as the 3-D painting “The Next Rembrandt” (9), created by Microsoft on the basis of paintings of Rembrandt van Rijn (10), or text translations produced by “DeepL” and musical compositions for the “AI Song Contest” (11).

Section 3 | The Concept of “work” under EU law

The concept of “work” has been recognized by the CJEU as an autonomous and harmonized concept of EU law which must be interpreted and applied uniformly, and requiring two cumulative conditions to be satisfied.

1. The subject matter of the work must be “original”. “Original” means that it reflects the personality of its author, as an expression of the author’s free and creative choices. If the subject matter is dictated by technical considerations, rules or other constraints, with no room for creative freedom, such subject matter will not be considered to be “original”.
2. The subject matter must be identifiable with sufficient precision and objectivity (12).

It is clear that there may be certain complications to demonstrate free and creative choices of an author in the case of AI-assisted output. What types of choices are deemed relevant in the creation of AI-assisted output? Can the requirement of originality even be fulfilled given the role of the AI system in the creation of AI-assisted output? Is it relevant in which phase of the creative process, the alleged free and creative choices have been made? Can AI-generated output fall under the above mentioned concept of “work”?

In its report of September 2020 entitled “Trends and Developments in Artificial Intelligence – Challenges to the Intellectual Property Rights Framework”, the EC has addressed the above mentioned questions. More specifically, the EC has reviewed how to apply the conditions set out by the CJEU to AI-assisted output and AI-generated output, resulting in the proposal of a “four-step test”. We underline that the CJEU is in no way bound by the EC’s report.

Section 4 | The four-step test: Can AI-assisted output and AI-generated output qualify as protected “works”?

As a reminder, we wish to understand if and to what extent AI-assisted outputs (i), and AI-generated outputs (ii), can be qualified as a “work” under EU copyright law. In an attempt to provide an answer to these questions, the European Commission has proposed a four-step test (13). Below, we will discuss every step of this test and apply those to “The Next Rembrandt” (13), a 3-D-printed painting created by an AI system (see image below).

AI-assisted output

Step 1: The AI-assisted output must be a production in the literary, scientific or artistic domain. This includes for example news articles, poems, compositions, geographical maps, photos, films, etc. (Article 2(1) of the Berne Convention). For many AI systems, the output will fall under this category. In our example, the Next Rembrandt fulfills this condition as the final output generated consists of a
3-D-painting.

Step 2: The AI-assisted output must be the result of human intellectual effort. In the Painer case, the CJEU clarified that it is possible to create works of authorship with the aid of a machine or device (15). According to the EC, AI-assisted outputs will always go hand in hand with some form of human intervention, be it the development of software, the gathering or choice of training data, editing, etc.
Therefore, this condition will generally be fulfilled for AI-assisted output.

In our example, the Next Rembrandt meets this requirement. The creators of The Next Rembrandt gathered data from the entire collection of the works of Rembrandt. In order to obtain this data, the creators analyzed a broad range of materials like high resolution 3-D scans and digital files, which were upscaled by deep learning algorithms to maximize resolution and quality. This extensive database was then used to create The Next Rembrandt. In addition to this, they designed a software system that uses different algorithms to identify Rembrandt’s style (composition, geometry, painting materials), and to construct a face out of the features. Yet another software was developed and used
in order to convert the digital file into a 3-D printed work that could mimic the brush strokes used b Rembrandt. Therefore, the result was at least partially the result of human intellectual effort.

Step 3: Originality. In the Painer case, the CJEU stated that a creative combination of ideas at distinct
stages of the production process might be sufficient for the result to qualify as a “work” protected under EU copyright law (16). The originality of an AI-assisted output will depend on whether a human author has made creative choices during the production process, and that these creative choices are reflected in the final result.

In the assessment of originality, creative choices in three phases of the production process (17) are taken into account:

Conception phase: This phase involves the creating and elaborating of a work and requires a series of detailed design choices, like genre, style, materials, technique, etc. These decisions can also entail the choice of the AI system as well as the selection of input data. With AI assisted outputs these choices will mostly be exercised by a natural person. The AI system will, in general, play no role in this phase.

In our example, the creators of the Next Rembrandt have made choices concerning the input and training data that was used to generate the final output (see above). One may argue that the choices concerning style, composition, color etc. are not “free and creative”, because they are dictated by the earlier works of Rembrandt.

Execution phase: This phase involves converting the design or plan into draft versions of the final work. Examples include producing text, recording music, taking photographs, coding of software, etc. In this phase, the AI system will often play a dominant role in the creative process whereas the user will have a rather operational role by guiding the AI system towards the desired output. However, this can still entail creative choices on the part of the user. In the case of supervised Deep Learning systems especially (18), the user’s role is vital in constantly monitoring the output and giving feedback and adjusting the AI system.

In our example, the AI system played a dominant role in the creation of the draft versions of The Next Rembrandt. It used the principles it had learned to replicate Rembrandt’s style and to generate new facial features for the new digital painting. The same is true for the 3-D printed painting that mimics the brushstrokes of Rembrandt. However, the creators must have actively monitored, adjusted and given feedback to the AI system throughout the creative process, in order to achieve the final output. Redaction phase: This phase involves processing and reworking the draft versions into a finalized cultural product before it is published and marketed, for instance rewriting,
formatting, editing, mastering, and other “post-production” activities. For AI-assisted outputs, the human actor will often make creative choices in this phase, e.g. a musician using an AI music composer will often edit the output before finalizing his track. On the
other hand, the output generated by “DeepL translator” will usually not require extensive redaction by the user.

In our example, we may argue that the creators have made many choices throughout the creative process in the redaction phase, such as the selection of the AI-generated facial features or the position of shadows facial features. The decision to use a 3-D print in 13 layers for the digital output of the AI system, can arguably also be seen as a free and creative choice.

The EC concludes that the amount of free and creative choices made by a human actor in the creative process of AI-assisted output should not be underestimated. Therefore, these choices may contribute  to the result of an original work. An AI-assisted output could qualify as a work protected by copyright if a human actor initiated and conceived the work and subsequently redacted the output in a creative
manner.

Step 4: Expression. The work needs to be identifiable with sufficient precision and objectivity. Given the “black box” characteristic of Machine Learning systems, the human actor may not be able to predict the outcome of the execution phase. However, as long as the output stays within the ambit of the author’s general authorial intent, this condition should not form an obstacle.

In our example, the Next Rembrandt will most likely meet this requirement. Indeed, there was a clear initial intent of the creators to make a 3-D painting in the style of Rembrandt. Therefore, the generated output will fall within the creator’s general authorial intent.
It is important to note that copyright protection will occur automatically at the conception of the work and there is no registration requirement for copyright in the EU. The validity of copyright is only assessed ex post before national courts. Time will tell if national courts will indeed decide that the choices made by the creator in the production of an AI-assisted output, are in fact free and creative choices. More specifically, it remains to be seen whether national courts will concur that choices made only in the conception phase and in the redaction phase – as suggested in the four-step test of  the EC – will suffice to fulfill the condition of originality. From a practical perspective, we advise the author to meticulously document all his creative choices made in every phase of the creative process of an AI-assisted output. This will facilitate the author to demonstrate the original character of the work, and to claim copyright protection.

AI-generated output

As we have discussed above, AI-generated output is created without any human intervention. Therefore, it does not meet the requirements to qualify as a “work”, as laid down in the four-step test. For AI-generated output to enjoy copyright protection in the EU, the current legal framework does not seem suitable. In order to tackle this issue, the legislator will have to assess whether it is desirable in our social and
economic context to grant copyright protection to machine creations. WIPO takes the view that, by excluding AI-generated works from eligibility for copyright protection, the copyright system would be seen as an instrument for encouraging and favoring the dignity of human creativity over machine creativity. Still according to WIPO, if copyright protection were accorded to AI-generated works, the copyright system would tend to be seen as an instrument favoring the availability for the consumer of the largest number of creative works and of placing an equal value on human and machine creativity.

If the existing legal framework for copyright protection would remain as it is, notably tailored around the author as a human being, AI-generated outputs will not be eligible for copyright protection. However, the EU legislator may then consider creating a separate type of protection for AI generated works, such as the sui generis type of protection for databases in the EU.

Section 5 | Conclusion and remarks

In this article, we have examined the question whether, and under what conditions, AI creations can qualify as “works” protected under EU copyright. Generally, a distinction is made between AI creations made with human intervention and/or direction (AI-assisted output), and those made without any human intervention (AI-generated output).
The current EU copyright framework is generally suitable to assess whether AI-assisted output – for
example 3-D painting “The Next Rembrandt” and Rembrandt and 2-D work “A Recent Entrance To
Paradise” – can qualify as a “work”. The four-step test of the European Commission suggests that AI assisted outputs can be protected under EU copyright law, but the condition of originality will have to be assessed by national courts in the EU. More specifically, there will need to be free and creative choices by a human interlocutor during the conception phase and the redaction phase of the AI work.

As AI-generated output is created without any human intervention, it does not meet the requirements to qualify as a “work”, as laid down in the four-step test. For AI-generated output to enjoy copyright protection the current legal framework is not suitable. There will probably be a need for the EU legislator to address this point in the future. The GEVERS AI Team will be happy to assist with any questions you may have, including questions concerning the copyright protection of your AI creations.

(1) See for example: A. RAMALHO, “Will robots rule the artistic world? A proposed model for the legal status of creations by artificial intelligence systems”, Journal of Internet Law 2017, https://ssrn.com/abstract=2987757
(2) Creativity Machines is the predecessor to a subsequent AI system invented by Dr. Thaler “DABUS”. For more information on Creativity Machines and DABUS please see: https://cdn.ymaws.com/www.apaonline.org/resource/collection/EADE8D52-8D02-4136-9A2A-
729368501E43/ComputersV19n2.pdf

(3) See Paradise Rejected: A Conversation about AI and Authorship with Dr. Ryan Abbott – Center for Intellectual Property x Innovation Policy (gmu.edu)
(4) Copyright Review Board 14 February 2022, SR # 1-7100387071.
(5) See WIPO Conversation on Intellectual Property (IP) and Artificial Intelligence (AI) – Revised issues paper on Intellectual Property and Artificial Intelligence, 21 May 2020.
(6) See European Commission Trends and Developments in Artificial Intelligence – Challenges to the Intellectual Property Rights Framework; September 2020.
(7) See European Parliament Resolution of 20 October 2020 on Intellectual property rights for the development of artificial intelligence
technologies, 2020/2015(INI)).

(8) See European Commission Trends and Developments in Artificial Intelligence – Challenges to the Intellectual Property Rights Framework; September 2020, p. 80.
(9) For more information please see: https://www.nextrembrandt.com/
(10) For more information please see: https://www.deepl.com/translator
(11) For more information please see: https://www.aisongcontest.com/
(12) CJEU 12 September 2019, C-683/17, §29-32.

(13) See European Commission Trends and Developments in Artificial Intelligence – Challenges to the Intellectual Property Rights Framework; September 2020, p. 79-85.
(14) “The Next Rembrandt is a computer-generated 3-D–printed painting developed by a facial-recognition algorithm that scanned data from 346 known paintings by the Dutch painter Rembrandt van Rijn in a process lasting 18 months. The portrait consists of 148 million pixels and is based on 168,263 fragments from Rembrandt’s works stored in a purpose-built database” (see: https://www.wipo.int/wipo_magazine/en/2017/05/article_0003.html and https://www.nextrembrandt.com/).

(15) CJEU 1 december 2011, C-145/10, Painer.
(16) CJEU 1 december 2011, C-145/10, Painer.
(17) Should you desire more information on the different stages of an AI system and the stakeholder involved, we refer to our article “Dissecting the AI value chain” https://www.gevers.eu/blog/artificial-intelligence/video-post/