Tomaž Lajovic



Foundation Models as Infringers: Should

Large-Scale AI Training Trigger Collective

Licensing Obligations under EU Law?



Ασπασία



SPLATO

Title:



Foundation Models as Infringers: Should Large-Scale AI Training

Trigger Collective Licensing Obligations under EU Law?

Author:

Tomaž Lajovic

Evaluation by Prof. Dr. Thomas Hoeren, University of Münster,

Law of the Digital Economy Certificate Programme.

Series:

Ασπασία (Aspasia)

Editor:

Tomaž Lajovic

Publication Date:

9 October 2025

Edition:

First Digital Edition (Preprint)

Website:

www.splato.eu

Publisher:

SPLATO

Tomaž Lajovic s.p., Ljubljana, Slovenia

© 2025 SPLATO

This work is licensed under CC BY-NC-ND 4.0.

To view a copy of this license, visit

https://creativecommons.org/licenses/by-nc-nd/4.0/

This publication is distributed for academic purposes only

and is not available for commercial sale.

ISBN 978-961-07-2932-7 (PDF)



Kataložni zapis o publikaciji (CIP)

pripravili v Narodni in univerzitetni knjižnici v Ljubljani

COBISS.SI-ID 252408067

ISBN 978-961-07-2932-7 (PDF)

Foundation Models as Infringers: Should



Large-Scale AI Training Trigger Collective

Licensing Obligations under EU Law?



Legal Essay

as part of the certificate course

Law of the Digital Economy



Author:

Tomaž Lajovic

tomaz.lajovic@gmail.com



Date of Submission: 24/08/2025





Table of Contents





A. Introduction ........................................................................................................................... 1

I. Foundation Models and AI Systems ................................................................................... 1

II. Provision of FM-based AI Systems ................................................................................... 3

B. Use of Copyrighted Works in FM-based AI Systems ........................................................... 3

I. Preparation of Training Datasets ......................................................................................... 3

II. Creation of Foundation Models ......................................................................................... 4

III. Output Generation ............................................................................................................ 5

IV. Use of AI Generated Output Outside the AI System ....................................................... 7

V. Issues Pertaining to Moral Rights ...................................................................................... 7

VI. The Question of Lawfulness of Access, Source and User ............................................... 8

C. Exemptions and Collective Licensing for the Purposes of Generative AI Systems ............ 10

I. Considerations and Approach ........................................................................................... 10

II. General Principles of the Exemptions .............................................................................. 11

D. Conclusion ........................................................................................................................... 12



A. Introduction

The term “foundation models” (FM) has been coined in 2021 by a research

group at Stanford Institute for Human-Centered AI, defining it “as models

trained on broad data (generally using self-supervision at scale) that can be

adapted to a wide range of downstream tasks” and seeing their potential at the

time “as the subject of a growing paradigm shift, where many AI systems

across domains will directly build upon or heavily integrate foundation mod-

els”1. It has since been commonly used in reference to the AI systems of gen-

erative nature (Gen-AI).



I. Foundation Models and AI Systems

In this essay the focus within the range of FMs is set on large language models

(LLMs) “because how LLMs use copyrighted material is obviously relevant

to the legal analysis of potential liability for copyright infringement”2. Tex-

tual content that LLMs are trained on is essentially a number of compilations

of word sequences and since computers cannot directly comprehend the

words of human languages, the text must be represented by numerical values3.

The central components of LLMs are artificial neural networks that define the

way the model is able to be trained and “contain an internal representation of

the input that has been processed during the model’s training”4. During the

training process the model effectively develops is own internal language that

itself uses to describe the ingested data, hidden to the outside world within

the “model’s black box [that] is mystical, both before and after the training.”5

Initial step in training is “tokenisation” that transforms natural and artificial

languages’ representations into a set of tokens, forming the LLM’s vocabu-

lary used to represent any text as a series of numbers in a lossless manner – it

allows for “map[ping] raw text onto a set of numbers, and a set of numbers

back into text”6. The next step is “word embeddings” which “provide the rep-

resentation of a word [token] as a high-dimensional vector” 7 defining indi-

vidual token’s “probability distribution over enormously large vector



1 (Bommasani and Liang 2021, p. 1)

2 (Gervais, et al. 2024, p. 1)

3 (Gervais, et al. 2024, p. 1)

4 (Dornis 2024, p. 16)

5 (Dornis 2024, p. 10)

6 (Gervais, et al. 2024, p. 3)

7 (Gervais, et al. 2024, p. 3)

spaces”8 of other tokens and embeddings. In this process the model learns of

“the meaning of a word in the context of the words that surround it, as found

in the text that is used during training […, resulting in …] a mathematical

construct that can efficiently capture the meaning of words based on the var-

ious contexts (i.e., word sequences) in which a word can be found“9.

According to the taxonomy of FM-based AI systems, proposed by Qinghua

Lu and co-authors, FM may be used as a “connector” between interaction and





Figure 1: Role of foundation models in implementations of AI systems



Source: (Lu, et al. 2024, p. 4)

task execution components, as well as one of the AI models as a component

for task execution10. As a connector the FM may serve as a communication

connector enabling transfer of data between the components, and/or coordi-

nation connector coordinating computation of different software (task execu-

tion) components, and/or conversion connector adapting data formats com-

municated to individual software components, and/or facilitation connector

optimising interactions between components11.



The distinction between text-and-data-mining (TDM) and Gen-AI systems

lies in the fundamentally different ways they use the large-scale data in-

gested12. “TDM […] involves using software to process large volumes of text,

images, or other data in order to find patterns [… while Gen-AI systems] are

engineered to process large datasets and algorithmically synthesise out-

puts.” 13 Therefore, the purpose of TDM is to “generate information and new



8 (Gervais, et al. 2024, p. 4)

9 (Gervais, et al. 2024, p. 4)

10 (Lu, et al. 2024, p. 4)

11 (Lu, et al. 2024, p. 4)

12 (Lucchi 2025, p. 38)

13 (Lucchi 2025, p. 38)

knowledge”14 and the purpose of Gen-AI is ”the production of data that are

similar to the training data.” 15



II. Provision of FM-based AI Systems

An FM-based AI system may be set-up by the user obtaining a copy of FM

and implement it in an AI system deployment installed on their own computer

infrastructure (either owned or rented) or they can rely on an AI system ser-

vice provider as an online service, offered to subscribers (paid or free) in the

form of Model-as-a-Service (in many flavours, e.g. generalised like Soft-

ware-as-a-Service or custom deployments like Platform-as-a-Service). The

way the AI system is set-up defines the person in whose sphere of influence

(i.e. responsibility) an individual downstream process is performed.



B. Use of Copyrighted Works in FM-based AI Systems

The so called “generative-AI supply chain”, as accounted for by Feder Cooper

and James Grimmelmann, can be broken-down into eight interconnected

stages16. The first three stages address the original content by (1) creation of

works and other information by their respective creators, (2) conversion into

digitised data, and (3) collection and curation into training datasets. The next

three stages relate to the FM, stating its (4) pre-training, (5) fine-tuning to the

specific domain, and (6) public release or embedment in an AI system. The

final two stages deal with the use of the AI system by (7) prompt-induced

generation of output and (8) alignment of the model with human preferences

or usage policies. The authors note that “even to call this a supply ‘chain’

understates its complexity; it is a densely interconnected ecosystem whose

stages can branch, recombine, loop, repeat, and feed back into each other.” 17



I. Preparation of Training Datasets

FMs are trained on large datasets of diverse data that may include copyrighted

material. Moreover, the AI industry admits that since “copyright today covers

virtually every sort of human expression, [… training of FMs] without using

copyrighted materials […] would not provide AI systems that meet the needs

of today’s citizens.”18



14 (Dornis 2024, p. 9)

15 (Dornis 2024, p. 9)

16 (Cooper and Grimmelmann 2025, p. 12)

17 (Cooper and Grimmelmann 2025, p. 13)

18 (OpenAI 2023, p. 4)

Regardless of whether a training dataset by itself is copyrighted as a database

(copyrighted database) and/or granted sui generis database right protection

(sui generis protected database), the copyrighted material collected within –

that may include other copyrighted databases – is subject to reproduction

(possibly twice, should it be digitised first), and the sui generis protected da-

tabases within the datasets are extracted.

Further on, the dataset itself, if it is a copyrighted database, and/or the copy-

righted material within, may be distributed or communicated to the public or

rented for the purpose of FM training. Such an act constitutes also re-utilisa-

tion of the training dataset as a sui generis protected database and of all sui

generis protected databases within.



II. Creation of Foundation Models

The stages 4 and 5 mentioned above represent the two phases of FM training

with the pre-training phase focusing on quantity of materials and the fine-tun-

ing phase19 20 focusing on quality of material used. In this regard, both data

included in the training sets and training set by themselves are deemed mate-

rials.

Specifically, LLMs are trained on textual material that is reproduced or trans-

lated (or adapted or otherwise altered) and stored within the FM during the

training process, using its own language expressed by enormous vector spaces

of tokens and embeddings. “The syntax is not filtered or sifted out during

generative AI training” 21 and “the number of possible combinations is astro-

nomical, [… therefore] the only combinations that would have non-negligible

weights (i.e., non-zero probability of occurrence) would be the ones that were

observed during training.”22 Within the LLM’s black box a lossless reproduc-

tion of textual material or “memorization” may (and, in most successful train-

ing cases, does) occur, defined as when “(1) it is possible to reconstruct from

the model (2) a (near-)exact copy of (3) a substantial portion of (4) that spe-

cific piece of training data”23. “If a generative-AI model has memorized



19 The fine-tuning phase may be further divided in three sub-phases: supervised fine-tuning,

reward modelling, and reinforcement learning. (Gervais, et al. 2024, p. 5)

20 (Gervais, et al. 2024, p. 5)

21 (Dornis 2024, p. 21)

22 (Gervais, et al. 2024, p. 4)

23 (Cooper and Grimmelmann 2025, p. 1), where the authors further explain that they “dis-

tinguish ‘memorization’ from ‘extraction’ (in which a user intentionally causes a model to generate a (near-)exact copy), [and] from ‘regurgitation’ (in which a model generates

copyrighted works, the memorized aspects of those works are present in the

model itself.” 24 It is “clear that copyright-protected works in the training data

are ingested and digested during generative AI training with regard to all in-

formation contained [… and that] the vector spaces inside generative AI mod-

els represent elements which, according to legal doctrine, must be considered

to be part of the works’ creative and expressive quality.”25 Thus, memorized

copyrighted works and copyrighted databases are subject to either reproduc-

tion or translation (or adaptation or another kind of alteration), and content of

sui generis protected databases is subject to extraction.

The reproductions or translations (or adaptations or some kinds of alterations)

and content of sui generis protected databases within the LLM (in the form of

its parameters) may subsequently be publicly released and/or embedded in an

AI system for deployment in a software service.26 Depending on the actual

modalities this may be considered (a) either distribution or communication to

the public or rental of the copyrighted works and copyrighted databases mem-

orized within the model and (b) re-utilization of the content of sui generis

protected databases extracted to the model.



III. Output Generation

LLMs generate output based on user-invoked prompts. These may be created

by the user directly, or automatically generated by software agents following

certain programmed procedures and/or using a FM as a connector performing

communication, coordination, conversion, or facilitation tasks within the AI

system used (as explained in A.I above).

In the process of responding to the user’s prompt an AI system does not just

refer to the one LLM embedded within and provide output. Instead, it uses

embedded FMs as connectors to address respective FMs and other non-AI

components (as components for task execution; see Figure 1 above). These

may either be embedded within the AI system or are accessible to it by any



a (near-)exact copy, regardless of the user’s intentions) [… as ‘memorization’ specifically referring to] ‘reconstruction’ (in which the (near-)exact copy can be obtained from the model by any means, not necessarily the ordinary generation process)”; “memorization is a back-end phenomenon; it describes the characteristics and capabilities of the model itself that directly result from its training, [while] regurgitation and extraction are front-end phenomena; they describe how the model behaves in generating outputs in response to a specific prompt.” (Cooper and Grimmelmann 2025, p. 18),

24 (Cooper and Grimmelmann 2025, p. 5)

25 (Dornis 2024, p. 21)

26 (Cooper and Grimmelmann 2025, p. 13)

means – both hosted on the web 27, including those behind paywalls or other-

wise technically protected, and any other data sources it may be able to ac-

cess.

It should be noted that the user’s prompt feed may also include reproductions

of copyrighted material and content of sui generis protected databases. More-

over, the prompt or a specific part of it might in certain cases be deemed a

copyrighted work/database and/or sui generis protected database by itself.

Whether entering a prompt should be treated as an act of publication (in cases

of previously unpublished works) depends both (1) on the internal mechanics

of the AI system in terms of its treatment of prompts received – either the

works/databases within user’s prompts may only be used for alignment lim-

ited to the individual user or their impact may exceed these limitations and be

effectively ingested into the vector space available to the other users of the

AI system or otherwise influencing other users’ use of it – and (2) on the other

eventual unattended publication procedures of the AI system in question.

Nevertheless, the AI system uses the copyrighted works/databases provided

with a prompt by reproducing/altering/extracting them within its components,

including vector spaces addressed, in order to come up with an appropriate

output.

The LLM’s output as a response to a specific prompt may include copyrighted

works/databases and content of sui generis protected databases – either mem-

orized during the model’s training or provided to it with prompts (including

the alignment process) or fetched from other components for task execution 28

during output generation process – in their original or translated (or adapted

or otherwise altered) form. In copyright terms such outputs (or parts thereof)

are either direct reproductions or translations (or adaptations or other kinds of

alterations) of originals and this is true regardless of whether the LLM’s out-

put is only temporary and used internally within the AI system in its function

as a connector. Should the outputs be made accessible to the user (or any other

public), an act of either or both communication to the public of the copy-

righted works and re-utilization of sui generis protected databases.



27 Apart from generally visible and indexed websites, also publicly non-indexed dark-web

sites might be “known” to an AI system.

28 See Figure 1 above.

IV. Use of AI Generated Output Outside the AI System

The output by the LLM containing copyrighted material that has been made

accessible to the user may be subject of reproduction, of communication to

the public, of translation/adaptation/alteration, and of any other use the user

put it to outside of the AI system. Regardless of the output containing copy-

righted material in their original or altered form being obtained from an AI

system, (1) the copyrighted works within the output retain direct links to the

original works and the respective rightsholders their rights and (2) the makers

of the databases retain their protection of content of sui generis protected da-

tabases29.



V. Issues Pertaining to Moral Rights

The moral rights of authors that must be respected with any use of the copy-

righted works (within the Union for the protection of the rights of authors in

their literary and artistic works under the Berne Convention for the Protection

of Literary and Artistic Works) comprise some form of “the right to claim

authorship of the work and to object to any distortion, mutilation or other

modification of, or other derogatory action in relation to, the said work, which

would be prejudicial to his honor or reputation.”30 A relatively straight-for-

ward authorship attribution (or at least source attribution, should authorship

itself be undisclosed or otherwise missing) of works is key to respecting the

former right to claim authorship, while use of works in AI systems raise com-

plex issues with the author’s right of integrity as subject of the latter right to

objection.

Regarding the right to claim authorship the creators of training datasest and

developers of LLMs should (jointly) be able to ensure that at least attribution

is not intentionally omitted in preparation of the training data, so it can be

found within the material ingested and properly accounted for during training,

and subsequently stated with the (substantial parts of) copyrighted works in

the LLM’s outputs. Whether the developers might be willing to comply with

the eventual requirement to make a comprehensive list of all (copyrighted)



29 Protection applies to both extraction and re-utilization of substantial parts of the content

and in cases of repeated and systematic extraction and/or re-utilization extends also to insubstantial parts thereof; see Article 7 of the Database Directive (Directive 96/9/EC of the European Parliament and of the Council of 11 March 1996 on the legal protection of databases)

30 See Berne Convention for the Protection of Literary and Artistic Works, Article 6bis(1).

material ingested available with the release of their LLMs is questionable,

since they “tightly guard their training pipelines as trade secrets [… avoiding]

full transparency of dataset composition [as it] would undermine commercial

confidentiality and complicate web-scale crawling.”31

Ensuring the right of integrity in AI systems may be the easiest in cases when

substantial verbatim sections of source material are reproduced in the output.

Any depart from it raises concerns – as even relatively modest alterations

might lead to the author’s objection, the eventual translations and adaptations

are significantly more exposed to such actions. Moreover, “a recent EU

study[32] found that even when AI outputs do not reproduce original content

in a recognisable way, they may still infringe moral rights if they mimic the

author's style or cause reputational harm. This perspective was endorsed by

67% of surveyed experts, who supported allowing rightsholders to invoke

moral rights to oppose AI training, even where economic rights exceptions

like TDM might apply.”33



VI. The Question of Lawfulness of Access, Source and User

“The concept of lawfulness in relation to user status or user acts has been

gradually established in EU digital copyright law as a condition for the en-

joyment of certain copyright exceptions.” 34 If the principle is to be applicable

to all cases of exceptions and limitations, regardless of its application being

founded on explicit requirements by the law35 or on the jurisprudence by the

CJEU36, leads to the conclusion that, in the absence of a license granted by

the rightsholder directly, any such use may be deemed illegal37. In this regard,

licenses granted by collective management organisations may have effects



31 (Lucchi 2025, p. 82)

32 European Commission: Directorate-General for Communications Networks, Content and

Technology, Crowell&Moring, IMC University of Applied Sciences Krems, Philippe Rixhon Associates, Technopolis Group and UCLouvain. “Study on copyright and new technologies – Copyright data management and artificial intelligence.” Publications Of-fice of the European Union. 16 March 2022. https://data.europa.eu/doi/10.2759/570559 (accessed August 10, 2025).

33 (Lucchi 2025, p. 82)

34 (Synodinou 2025)

35 Such requirement is stated in relation to TDM exception in Articles 3 of the DSM Directive. 36 See CJEU cases ACI Adam BV and Others v Stichting de Thuiskopie (C-435/12), Copydan

Båndkopi v Nokia Danmark A/S (C-463/12), Vereniging Openbare Bibliotheken v Sticht-ing Leenrecht (C-174/15).

37 In some jurisdictions (i.e. France, Belgium, Spain, Slovenia) the exclusive right to first

publication is also listed among the moral rights of the author, effectively precluding any kind of lawful access to the work beyond extremely narrow private/confidential use of non-published works explicitly allowed by the author himself.

similar to exceptions and limitations due to their obligation to either issue

license or provide reasons, based on objective and non-discriminatory crite-

ria, not to.38 Such an obligation does not exist for the rightsholder exercising

his rights individually, therefore (at least) the cases of mandatory collective

management of exclusive rights (otherwise not subject to limitations and ex-

ceptions) present an important limitation in the exercise of the right39 and

should be treated as such.

Despite being referred to, neither comprehensive definition of the terms

“(un-)lawful access” 40 41 42 , “(un-)lawfully accessible” , “(il-)legal access”,

“(un-)lawful acquirer” 43 44 45 , “(un-)lawful user” , “(un-)lawful source” nor the

eventual distinctions between them is set by the copyright law and relevant

jurisprudence. “All these different terms cause quite the headache – they all

seem related, yet they are differently formulated and some can be understood

to refer to completely different things, such as the bona fide character of the

user, the status of the copy, or the status of the source from which the copy is

obtained.”46

Recital 14 of the DSM Directive appears to counter the approach to “lawful-

ness” as introduced by the CJEU by allowing for a kind of “source-washing”

in the case of TDM exception effectively included in the Article 3 of the same

directive. It states that “lawful access should also cover access to content that

is freely available online”. “A literal interpretation of the free access basis of

lawful use […] could a priori mean that if a copyright protected work is found



38 See Article 16 of the CRM Directive (Directive 2014/26/EU of the European Parliament

and of the Council of 26 February 2014 on collective management of copyright and re-lated rights and multi-territorial licensing of rights in musical works for online use in the internal market).

39 As noted also in the Recital 19 of the CRM Directive, referring to rightsholders’ freedom

of choice in relation to exercising their respective rights: “Where a Member State, in compliance with Union law and the international obligations of the Union and its Member States, provides for mandatory collective management of rights, rightholders’ choice would be limited to other collective management organisations.”

40 See Article 3 of the DSM Directive (Directive (EU) 2019/790 of the European Parliament

and of the Council of 17 April 2019 on copyright and related rights in the Digital Single Market and amending Directives 96/9/EC and 2001/29/EC).

41 See Article 4 of DSM Directive.

42 See Article 6(4) of the Infosoc Directive (Directive 2001/29/EC of the European Parlia-

ment and of the Council of 22 May 2001 on the harmonisation of certain aspects of cop-yright and related rights in the information society).

43 See Article 5(1) of the Software Directive (Directive 2009/24/EC of the European Parlia-

ment and of the Council of 23 April 2009 on the legal protection of computer programs)

44 See Articles 6(1) and 9(1) of the Database Directive.

45 See CJEU cases referred to in the note 36 above.

46 (Szkalej 2025, p. 312)

on free access (non-technically restricted) online, then it should be lawful to

conduct text and data mining, without taking into consideration any contrac-

tual restrictions or even the unlawful source of the work. […] This approach

disregards the previous case law of the CJEU in the “lawful source” cases.

[… Therefore], accepting that the “lawful source” requirement is also appli-

cable when mining content that is freely available online would mean that the

concept of free access must be more restrictively interpreted. This implies that

the free access must have been authorised from the right holder.”47



C. Exemptions and Collective Licensing for the Purposes



of Generative AI Systems

Use of copyrighted works may be performed either with appropriate licensing

secured directly with the authors, other rightsholders, independent manage-

ment entities, and collective management organisations, or relying on the use

within the boundaries of effective exemptions and limitations in accordance

with the provisions of the law. Generally, the same applies to exploitation of

sui generis protected databases. The approach presented below builds on the

considerations of Nicola Lucchi in his study for the EU Parliament 48 and on

the established principles of exemptions and collective management within

EU.



I. Considerations and Approach

The continuously growing number of AI system developers and users affects

practically every domain of human activity and generates substantial new

value (and/or potential harm49). With the high volume of works required for

FM training, many potential uses of works (both used for training and used

on-the-fly) in the process of output generation by AI systems, their authors

are right to demand a fair share as compensation for their contribution and

loss of income. As even moderately effective system based on individual li-

censing of billions of copyrighted works used in AI systems seems inconceiv-

able and taking into account that interests of AI developers and users on the

one hand and of creators on the other are inherently conflicting, introduction



47 (Synodinou 2025)

48 (Lucchi 2025)

49 So called “market dilution” by AI systems’ generated output may have important conse-

quences in the markets for the types of works included in the training datasets. (Lucchi 2025, page 127)

of a combination of specific exemptions for the AI systems’ purposes includ-

ing creators’ remuneration right and either mandatory or extended collective

licensing has the potential to provide for a proper balance of rights and obli-

gations for all the stakeholders concerned. The exemption should not apply

to at least certain categories of databases should their use in AI systems be

deemed either their normal exploitation or the application of exception would

be prejudicial to the legitimate interests of the rightsholders (training datasets

and LLMs, were they deemed as databases, are the most obvious examples).

Given that legislative introduction of copyright exemptions for development

and use of AI systems can generally be justified by (when legislated) securing

proper balance between the conflicting interests, they may still be applied

only within the boundaries set by the three-step test50 and effectively denied

should the balance be significantly shifted. The remuneration right should be

unwaivable and either managed under extended collective management re-

gime, should an opt-out right be provided for and enforceable, or simply man-

datory collectively managed.



II. General Principles of the Exemptions

The analysis in chapter B above has shown that there are three distinct sets of

operations in the generative-AI supply chain where, moving downstream,

change of operator and eventual commercialisation is most likely. Therefore,

the exemptions should apply accordingly, taking their actual proceeds and/or

value generated into account.

Purpose of Rights Remuneration principle Reporting and other

exemption exempted obligations

Preparation reproduction tariff per data volume, weighted by itemised reporting on of training category of material material included extraction



datasets levy as share of proceeds, for fingerprinting and/or distribution communication commercial exploitation watermarking of (optionally tariff deduced) dataset to public



rental levy as share of proceeds, for use inclusion of metadata of datasets in private reporting on proceeds re-utilisation environments (private use levy) acquisition and transfer

potential special conditions, for of certificates of

non-profits, start-ups … until private use levy paid

commercial exploitation or

certain threshold of proceeds



50 See Article 5(5) of the Infosoc Directive.

Purpose of Rights Remuneration principle Reporting and other

exemption exempted obligations

Creation of reproduction tariff per major versions, based on reporting on

foundation training data volume, weighted commercially alteration



models by category of material obtained training sets extraction used levy as share of proceeds, for distribution commercial exploitation itemised reporting on communication (optionally tariff deduced) used material included to public in the training sets levy as share of proceeds, for use rental of FM in private environments fingerprinting and/or re-utilisation (private use levy) watermarking of FMs

potential special conditions, for inclusion of metadata

non-profits, start-ups … until reporting on proceeds

commercial exploitation or acquisition, redemption certain threshold of proceeds and transfer of

certificates of private

use levy paid

Output reproduction minimal commercial tariff, based reporting on

generation on number of queries received commercially alteration



extraction or time of normalised obtained FMs used compute×throughput itemised reporting on communication levy as share of proceeds, for material used for to public commercial exploitation (tariff output generation re-utilisation deduced and certain share of reporting on proceeds private use levy paid deduced) redemption of certificates

potential special conditions, for of private use levy

non-profits, start-ups … until paid

commercial exploitation or

certain threshold achieved

Table 1: A possible approach to general principles of AI-systems-related exemptions



D. Conclusion

The TDM exception and AI Act are not sufficient to ensure compliance of AI

systems’ developers and users with the EU copyright framework and cultural

and creative sectors are right to demand appropriate steps to be taken to

“safeguard European intellectual property rights in the age of generative

AI.”51 Introduction of specific exemptions with remuneration rights and

collective management mechanisms is the right way to secure long-term

balance of the stakeholders’ conflicting interests.



51 (Broad coalition of rightsholders active across the EU’s cultural and creative sectors 2025)

Bibliography





Bommasani, Rishi, and Percy Liang. “Reflections on Foundation Models.” Stanford HAI. 18

October 2021. https://hai.stanford.edu/news/reflections-foundation-models (accessed

August 9, 2025).



Broad coalition of rightsholders active across the EU’s cultural and creative sectors. “Joint

statement by a broad coalition of rightsholders active across the EU’s cultural and

creative sectors regarding the AI Act implementation measures adopted by the

European Commission.” CISAC. 30 July 2025.

https://www.cisac.org/Newsroom/articles/joint-statement-broad-coalition-

rightsholders-active-across-eus-cultural-and (accessed August 7, 2025).

Cooper, Feder, and James Grimmelmann. “The Files are in the Computer: On Copyright,

Memorization, and Generative AI.” arXiv. 24 March 2025.

https://arxiv.org/pdf/2404.12590 (accessed August 11, 2025).

Dornis, Tim W. “The Training of Generative AI Is Not Text and Data Mining.” SSRN. 19

December 2024. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=4993782

(accessed August 8, 2025).

Gervais, Daniel J., Haralambos Marmanis, Noam Shemtov, and Catherine Zaller Rowland.

“The Heart of the Matter: Copyright, AI Training, and LLMs.” SSRN. 21 September

2024. https://ssrn.com/abstract=4963711 (accessed August 3, 2025).

Lu, Qinghua, Liming Zhu, Xiwei Xu, Liu Yue, Zhenchang Xing, and Jon Whittle. “A

Taxonomy of Foundation Model based Systems through the Lens of Software

Architecture.” 2024 IEEE/ACM 3rd International Conference on AI Engineering –

Software Engineering for AI (CAIN). Lisbon, Portugal: IEEE, 2024. 1-6.

Lucchi, Nicola. “Generative AI and Copyright - Training, Creation, Regulation.” European

Parliament. 9 July 2025. http://www.europarl.europa.eu/supporting-analyses

(accessed August 8, 2025).

OpenAI. “OpenAI—written evidence (LLM0113).” UK Parliament. 5 December 2023.

https://committees.parliament.uk/writtenevidence/126981/pdf/ (accessed August 9,

2025).



Synodinou, Tatiana. “Navigating User Lawfulness in European Copyright Law: From Lawful

Use to Lawful Access.” Kluwer Copyright Blog. 19 March 2025.

https://legalblogs.wolterskluwer.com/copyright-blog/navigating-user-lawfulness-in-

european-copyright-law-from-lawful-use-to-lawful-access/ (accessed August 10,

2025).

Szkalej, Kacper. „The Paradox of Lawful Text and Data Mining? Some Experiences from the

Research Sector and Where We (Should) Go from Here.“ GRUR International, 25.

March 2025: 307–319.

Evaluation by Prof. Dr. Thomas Hoeren,

University of Münster, Law of the Digital Economy Certificate Programme.





Ασπασία