Copyright Infringement of AI in the EU

• AI is a system designed to mimic how humans learn and create. Like humans they are “taught” at first. Then, a subsection of AI called generative AI create content based on those teachings. This teaching process is called training and it includes feeding the AI with vast amounts of data (input).
• This training data often includes copyright protected material: books, images, music, videos, and more. 
• The generative AI processes the data it is fed with to identify patterns and structures which then enables it make predictions and generate content (output).
• This data training process and the outputs generated raise questions on copyright infringement as they use and generate copyright relevant material. 

1) INPUT

The issue during the training stage is that making unauthorised copies of protected inputs to provide to the model could infringe the right of reproduction (Art 2 Information Society (InfoSoc) Directive) of copyright holders. Although there are discussions on whether such act of AI models is sufficient to be considered as a reproduction. 

a) TDM exception
One of the exceptions to copyright is text and data mining (TDM), outlined within Art 3-4 of the Digital Single Market (DSM) Directive. According to Art 2 of the DSM Directive TDM is “any automated analytical technique aimed at analysing text and data in digital form in order to generate information which includes but is not limited to patterns, trends and correlations.” It is largely agreed that this exception applies to the training of AI models on a case-by-case analysis. Before the AI Act there were concerns whether the TDM exceptions applied to the training of AI models. The AI Act referenced the TDM exceptions in the DSM Directive. Therefore, it mostly clarified that the TDM exceptions could be applicable to the training stage of AI models. 

i) Input used by research organizations
Art 3 of the DSM Directive applies to text and data mining of lawfully accessed works by “research organizations and cultural heritage institutions” for the “purposes of scientific research”. The copyright holders cannot opt-out of this process. Opting out refers to reservation by the copyright holders for their works to be excluded in datasets used for training artificial intelligence systems. 
See Kneschke v LAION (27 September 2024) of the District Court of Hamburg as the first decision applying TDM exceptions. 

ii) Input used in all tdm activities
Art 4 of the DSM Directive applies to “all” text and data mining, whether for commercial or non-commercial purposes, of lawfully accessed works. The copyright holders can opt-out of this process. They have to do so “in an appropriate manner, such as machine-readable means in the case of content made publicly available online”. For example, through metadata and terms and conditions of a website or a service (Recital 18 DSM Directive).

b) Transient copying
Another exception to copyright is transient copying as established within Art 5(1) of the InfoSoc Directive. This exception might be considered for some cases where the copy made by AI is temporary. The copy made by AI must fulfil the below cumulative criteria for this exception to be applicable: 

a) be transient or incidental
b) constitute an integral and essential part of a technological process
c) enable a lawful use
d) have no independent economic significance

2) OUTPUT 

Regarding the outputs generated by AI, the issue relates to the fact that: 

• training data of AI models might include copyright protected works, 
• Generative AI may memorise that training data, 
• and generate similar outputs to the copyright protected material used as part of the training. 

Under the right of reproduction of the author, even copying parts of the copyrighted work is considered infringement as long as the copied part is an original expression of the author (CJEU C‑5/08 Infopaq). Both literal and non-literal reproductions may be considered infringing depending on the case. This means that non-literal copying in the output of a generative may also infringe the right of adaptation of the author (legislated within national laws), being considered derivative work of the copyright holder. 

If an infringing use such as defined above is determined for an output generated by AI, still copyright exceptions and limitations might apply to that output. 

a) Especially the exceptions that relate to freedom of expression and the arts might be relevant. These are quotation for criticism or review (Art 5(3)(d) InfoSoc Directive) and use for the purpose of caricature, parody or pastiche (Art 5(3)(k) InfoSoc Directive). 

b) The three-step test might help interpreting the use to balance interests within the work and the use (Art 5(5) of the InfoSoc Directive). 

References

• Infringing AI: Liability for AI-Generated Outputs under International, EU, and UK Copyright Law- Eleonora Rosati 
  https://www.4ipcouncil.com/research/infringing-ai-liability-ai-generated-outputs-under-euuk-copyright-law
  https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4946312
• Copyright Infringement & AI: A Case Study of Authors Guild v. OpenAI and Microsoft-Maria Alexandra Mǎrginean
  https://www.4ipcouncil.com/research/copyright-infringement-and-ai-case-study-authors-guild-v-openai-and-microsoft 
• Generative AI, Copyright and the AI Act - João Pedro Quintais
  https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4912701
• Can Artificial Intelligence Infringe Copyright? Some Reflections-Enrico Bonadio, Plamen Dinev & Luke McDonagh
  https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4315222
• Artificial intelligence and copyright: use of generative AI tools to develop new content- EC website
  https://intellectual-property-helpdesk.ec.europa.eu/news-events/news/artificial-intelligence-and-copyright-use-generative-ai-tools-develop-new-content-2024-07-16-0_en
• Copyright Flexibilities: Mapping and Comparative Assessment of EU and National Sources-Caterina Sganga, Magali Contardi and others 
  https://ssrn.com/abstract=4325376

The Protection of AI Generated Outputs by Copyright in the EU

• Generative AI refers to artificial intelligence systems that create new content (output), such as text, images, or music, by learning patterns from existing data. These systems usually require a text or voice prompt to generate images, videos or texts. Some common generative AI tools are ChatGPT, Gemini and MidJourney. 
• For example, in MidJourney one can ask for ‘a living room full of elephants watching television’ and it will produce several images that describes that. This disruptive technology is enhancing content creation, streamlining various sectors, i.e. marketing and transforming industries like media, entertainment and design. At the same time, it poses a challenge to human-driven creative industries. It is impacting traditional roles and business models, while also raising concerns about job displacement. Also, the generated outputs could be misleading or unreliable at times. 
• As generative AI generates human-like content, discussion arises whether these outputs could be protected by copyright. 

Copyright Protection 

Subject Matter: Work

• The subject matter of copyright protection is literary and artistic works. 
  The Requirement for Copyright Protection: Originality
• To be protected by copyright, a work must be original. The definition of an original work in the European Union (EU) is “author’s own intellectual creation” reflecting the “personality” of its author and their “free and creative choices”. (Court of Justice of the EU (CJEU) C-5/08 Infopaq, C-145/10 Painer and C-683/17 Cofemel cases)

Authorship: Human

• All these suggest that copyright protection requires a human intellectual effort and a human author (the creator of the work). Therefore, outputs generated by AI without any human intervention cannot be subject to copyright protection.
• When it comes to copyright ownership, non-human entities could be designed as copyright owners in national laws. For example, for works created during employment, the copyright might vest in the employer (that are usually legal persons).  

Scope of protection:

• It is important to define the scope of copyright protection in order to analyse if an output could entertain copyright protection or not. Copyright extends to the ‘expressions’ and excludes ideas, style, facts, procedures, methods of operation and mathematical concepts. It is suggested that there is an implied causal link between the author’s creativity and its expression. It is suggested that this link is made through an authorial intent on top of the human creative contribution. A general authorial intent that allows to see the conception of the author before it produces the work is considered sufficient. There exists space for unintended expressive features here. For example, an unexpected shadow when taking a photo or an improvised act that makes it into the film.
• Also, the subject matter should not be dictated by technical function. Dictation by function requires that the choice to express an idea are very limited that there is no room for creativity. (CJEU C‑393/09 BSA case) However, subject matter containing a technical function can qualify for copyright protection as long as there is space to express creative choices. (CJEU C-833/18 Brompton Bicycle case)

Creative choices: 

• The creative choices in work eligible to copyright protection could relate to:
  a) the “conception/preparation” of the subject matter regarding planning, designing the style, technique, medium, materials etc.
  b) the “execution” which is producing or drafting the subject matter
  c) the “finalization/redaction” referring to the editing, framing and formatting (CJEU Painer case)
• If there is minimal or no creative space for the author, the creation cannot qualify as original.
  
  Minimum level of originality:
• It is required that the work is stamped by the personal touch of its author. 
• There is no aesthetic quality requirement.
• There is no need for the work to be new (in contrast to the novelty requirement in patents). It simply sufficient that the work was not copied.
• AI-assisted outputs may be protected under copyright as long as they reflect human creative choices at some point in the creation process. 

Copyright protection of AI generated outputs

A report by the European Commission determined a four-step test for an AI generated output to be a work protected by copyright. It is important to note that this test does not bind the CJEU.

1- An output in literary, scientific or artistic domain 
Generative AI systems are commonly able to generate type of works in the literary, scientific or artistic domain. Some examples of such generative AI outputs include music, poems, articles, drawings, photos, videos, films, scripts, software codes. 

2- Human intellectual effort
With the technological capacities of today there exists no fully autonomous creative machines. In any AI system there is some level of human intervention. This could relate to developing the software of the AI, the selection of the training data/dataset, prompting the generative AI with specifications, editing and formatting the output. There will exist some degree of human contribution at the process of generating AI outputs. However, the extent of human intellectual contribution to qualify for copyright protection is not certain. This leads us the next stage.

3- Originality/Creative Choice 
For an AI-assisted output to be considered original, the production process must demonstrate creative choices made by human authors, that is reflected in the final outcome. 

a) The “conception/preparation” phase: This stage involves the plan and design of the work. It is important for this phase to be elaborate and go beyond general ideas. Choosing the subject matter, genre, format, material, style, technique, medium could be relevant examples. These creative choices in most cases will be conducted by humans. However, one could get AI assistance in making these choices. The level of human creative contributions will matter at this stage.

b) The “execution” phase: The execution is about implementing the plans and designs made in the former stage. With humans this corresponds to writing the article, producing the notes of the music, coding the software, taking the shot in photos and videos etc. Here, AI systems are dominantly involved in the creative process. Meanwhile, the user of the AI could have more operational role to guide the AI to generate their desired output. Depending on the case, these aspects could be regarded as the creative interventions of humans.   

c) The “finalization/redaction” phase: The generated output goes through edits at this stage. It is the last touches on the work to have a refined and finalized version. These might include correction, rewrite, cropping, framing, editing, changing colours and selecting etc. There is much room for human creative contributions at this stage. It might be that these edits impact the human creative choices in the final result. However, some AI systems might create just the desired outputs that there would be less or no room for edits. In that case, the selection as a post-production work could present the creative choice of humans.

This whole process is repetitive in many instances. The execution and finalization phases could be repeated many times to get the desired result. There could even be re-conceptualization at some part of the creative process.  

4- Expression
Lastly, the human creativity should be expressed in the final result. The general authorial intent requirement of expression should allow to see the conception of the author before the output is produced while leaving space for unintended expressive features. The “black-box” feature of AI systems might create a challenge here. At this stage, the author would probably be remote to how their contention resulted in the output. This stage will depend on whether the black-box feature could be considered as sufficient to be in the frame of general authorial intent. Generally, it is stated the expression stage would not necessarily present an obstacle for the copyright protection of generative AI assisted outputs.

Reference:

• Trends and Developments in Artificial Intelligence Communications- Challenges to the Intellectual Property Rights Framework Final Report, European Commission (2020)
  https://digital-strategy.ec.europa.eu/en/library/trends-and-developments-artificial-intelligence-challenges-intellectual-property-rights-framework 

Remuneration for AI use in the EU

• AI models train with vast amount of data (input) that include copyrighted works as well and then produce an output as a result of that training.  
• It is commonly accepted that the use of copyrighted works in the input phase (training of the AI model) is a copyright relevant action, regarding to the right of reproduction of the authors. (Art 2 of the Information Society Directive) The same affect is also being discussed for the AI generated output depending on the specific output generated.
• The EU, has an exception based framework to address the question of copyright infringement of the right of reproduction of the authors. Art 3 and 4 of the Digital Single Market Directive (DSM Directive) provide exceptions for text and data mining (TDM) actions. The reference of the AI Act in Recital 105, 106 and Art 53 (1)(c) is proof that the TDM exceptions might be relevant for the use of copyright protected works by AI. 
• While the exception in Art 3 of the DSM Directive regards to TDM for scientific research purposes, Art 4 of the same directive is pertinent to all TDM actions, including commercial purposes. Unlike the scientific research exception, the general TDM exception in Art 4(3) provides an option for the right holders to the reserve their rights from TDM uses. Such reservation (opt-out) must be provided in a machine readable manner. 
• Right holders claim remuneration for such TDM uses of their works because their works are being copied at the input phase and the outputs generated by AI run the risk of substituting their works. Different systems and basis for such remuneration are being raised: 
  
  
  
  1) Some argue that the opt-out option held by creators within Art 4(3) of the DSM Directive, could eventually lead to a remuneration mechanism. The provision provides that the opt-out is to be made through machine readable means. This is exemplified as metadata and terms and conditions of a website. Through these technical tools, it is suggested that right holders could have different options. They could either choose an all-out option where they entirely reserve their rights of use. Another way could be for them to reserve their rights from AI use if they are not remunerated for such uses. Collective management organisations could be part of that remuneration process (Collective management organisations are entities that manage rights on behalf of copyright holders by licensing works, collecting royalties, and distributing them to the rights owners). The Art 53 (1)(d) of the AI Act provides for a transparency obligation for the AI providers to provide “a detailed summary about the content used for training” of AI models. This provision could support the opt-out and remuneration mechanisms for the right holders to detect if their work has been used for an AI training. 
  
  2) Some suggest a “statutory license” for remuneration of the right holders instead of the opt-out provision of Art 4(3) of the DSM Directive. A statutory license would limit the exclusive right of the copyright holders by permitting TDM uses and in balance would offer a remuneration to the right holders. This argument is drawn from the private copying exception of copyright (Art 5 (2)(d) of the Information Society Directive) that is provided in exchange of fair compensation to the right holders. 
  
  
  
  3) There exist arguments that are against a remuneration based on copyright protected input used for AI training. Instead, they propose a compensation based on the output generated by AI that could be a substitute to copyright protected works. It is an advised to manage this in a lump sum payment made to collective management organizations. This view is based on Art 15(1) of the WIPO Performances and Phonograms Treaty which provides for a “single equitable remuneration” to the phonogram producers and performers for the commercial use to phonograms in their communication to the public. 
  
  4) Some advocate a “revenue sharing” between the AI providers and the copyright holders where there is an agreement for a fair share of the revenue of the AI to be provided to the right holders for using their works within their AI models. 
  
  5) There are arguments for a “royalty payment” compensation in which the AI providers would pay for a set fee determined in an agreement between the AI provider and the copyright holder for each use of the work in the AI. 

Resources:

• Redefining Artificial Intelligence boundaries: how to address remaining copyright related issues following the latest draft of the AI Act
  https://www.4ipcouncil.com/research/redefining-artificial-intelligence-boundaries-how-address-remaining-copyright-related-issues-following-latest-draft-ai-act 
• Generative AI and Author Remuneration- Martin Senftleben
  https://link.springer.com/article/10.1007/s40319-023-01399-4 
• Generative AI, Digital Constitutionalism and Copyright: Towards a Statutory Remuneration Right grounded in Fundamental Rights, Part 1 and 2-Christophe Geiger
  https://copyrightblog.kluweriplaw.com/2023/10/17/generative-ai-digital-constitutionalism-and-copyright-towards-a-statutory-remuneration-right-grounded-in-fundamental-rights-part-1/ 
  
  https://copyrightblog.kluweriplaw.com/2023/10/19/generative-ai-digital-constitutionalism-and-copyright-towards-a-statutory-remuneration-right-grounded-in-fundamental-rights-part-2/ 
• ChatGPT: A Case Study on Copyright Challenges for Generative Artificial Intelligence Systems- Nicola Lucchihttps://papers.ssrn.com/sol3/papers.cfm?abstract_id=4483390