AI-Generated and AI-Assisted Code Under EU Law: Intellectual Property and Compliance Challenges
Part 5: Literal Elements vs. Non-Literal Elements in Computer Programs
Av. Mirela Takacs · May 19, 2026 · 9 min read
As I mentioned at the beginning of the series in Part 1, the proposed analysis is based on the factual elements described and documented by Dan Blanchard for the relicensing version of the open source library Chardet, with the express mention that the Chardet case is used only as a working hypothesis for testing the European legal framework. The effective solution for Chardet shall be analyzed by reference to the US law, considering that the events took place in the United States.
What is very interesting about that case, and what attracted me as soon as I read about it, is this potential detailed discussion that can take place regarding the notion of copyright for computer programs, as it was conceived almost 30 years ago through the TRIPS Agreement. There has always been a distinction between expression and functionality, but in the current and future technological context, because at the moment everything is happening at an accelerated pace, the border between what is protectable and what is not seems to be becoming more and more complex and difficult to identify. Current technology just shows us how carefully this distinction between expression and idea must be approached.
As seen in Part 4, the CJEU cases were quite clear, the most relevant example being the program developed by WPL: same functionality, different expression. At that time there was no debate related to AI, fantastically fast development, intelligent agents increasingly autonomous. In those days, software development was done completely differently. Although the referring court mentioned in the SAS case that WLP did not have access to the source code, an aspect reiterated by the CJEU, I don’t consider this reference as having a specific authority for the final judgment.
Analyzing everything that means protectable according to Directive 2009/24/EC, for infringement to exist, the protectable expression of the author has to be identified in the new work. Although the CJEU lists the object of protection under the directive (source code, object code, MCP), the referring court mentions that “the structural design of the source code” would also be part of those elements.
In order to deepen the reference made by the referring court, I have also analyzed the decision for the preliminary questions [1]. According to the referring court, copyright protection is not limited to the text of the source code of the program, but extends to protecting the design of the program, that is, what has been referred to in some cases as its “structure, sequence and organisation” (p. 232). The court argues its reasoning on the basis that the preparatory material is also protected. A few years later, another referring court, in Case C-159/23, also in the context of the enumeration of the object of protection under Directive 2009/24/EC, included the internal structure and organisation of Sony’s software.
From my perspective, in light of the courts’ reasoning, the non-textual/literal elements of a computer program such as structure, sequence and organisation also benefit from copyright protection under Directive 2009/24/EC to the extent that they represent creative choices of the authors. I believe that the notion of protected expression should be interpreted in a broad sense, to protect the expressions of the concept separately, as well as their arrangement or structure. That is, whether we are talking about verbatim copying (literal copying) or internal similarity (non-literal copying), there must be a concrete reproduction of the protected expression, something original to the author, however little, but his own, especially since in the context of AI generation, non-literal copying will become more relevant. And in this context, I consider that the rule of not protecting the concept itself is still respected, and thus it is not possible to create a monopoly on it, and also it is respected the originality, the creativity of the author.
However, using an AI system to develop a computer program brings new legal challenges. Given the analysis made in Part 3, if an output appears substantially similar to a preexisting protected work, publicly available online, there is a possibility that the model was trained on that data. Usually, a claim for infringement of copyright is advanced when a work seems to be similar to a protected work and when the claimant suggests that the defendant had access to the copyrighted work. However, this circumstance of access to the code, I consider to be more relevant as an circumstantial matter, and not as a decisive element in making a judgment. Ultimately, what are the chances of copying someone’s protected expression if you don’t have access to the protected work? Even in the SAS Case, the referring court found it surprising that someone would invoke infringement without the person responsible having even seen the protected work.
So returning to AI models, if the model’s training was done on source code available in open source repositories, and considering the technical aspects of how such a model is trained, which ends up learning parameters, which in isolation, do not benefit from legal protection, but in the end it still regurgitates something protected, the output can constitute an infringement as long as a protectable expression is reproduced. The question is whether that reproduction can ultimately be established as an infringement under the copyright law? But as long as the output does not reproduce anything from the protected work, either syntax, the arrangement, the architecture, or any recognizable element of the author’s own intellectual creation, how would you establish copyright infringement?
Still, considering the CJEU judgements studied in Part 4, if following a technical comparative analysis between the codes of the programs there is an overlap of source code, object code, structure, organization, then we should further verify the originality and complexity of each, if those overlaps are protected or are ideas, principles, standard algorithms and would not benefit from protection, or if are elements dictated by their technical function. And subsequently, for the remaining protectable elements it should be analyzed whether the overlap is the result of direct copying or the result of independent development, through the mechanism of a clean room implementation or by legally observing, studying and analyzing the functioning of a computer program as in the SAS case.
However, in case of implementation with AI, as long as the model’s training data contained the protected work from the beginning, it is even harder to evaluate if an independent implementation with that AI model is possible. As we saw in Part 3, from a technical perspective, similarity can result from specific training causes, not necessarily intentional copying of the protected work. Consequently, just as a detailed technical analysis is done to find similarities between software programs, I consider that such a technical analysis should also be done on how the AI model generated the alleged infringing program. Ultimately, the technical aspects related to the notion of memoristion, as well as the probability of an independent creation with an AI model should be part of the final analysis. But before you can even get to that layer of analysis, there still has to be an overlap: verbatim, approximate or structural.
For example, in a relevant case from Romania[2], about former employees (software engineer and programmer), who developed a new program, functionally similar to the employer’s program, which according to the employer’s claims presented very strong similarities with its own program, and the developers of the new program are its former employees who had access to the code of its protected program, the court decided that as long as it was not proven through technical expertise that the developed program does not contain identical source codes, does not use parts, lines of code, have different architectures, use different technologies and are developed by different methods, there is no infringement of the copyright of the employer’s program. There were no analysis of the relevance of the access to the code.
Returning to AI models, according to Article 53(1)(d) of the EU AI Act [3], providers of general-purpose AI models have the obligation to publish a “sufficiently detailed summary of the content used for training“. This will be a feasible possibility for claimants in copyright infringement proceedings to establish that a particular protected work was part of the training data. For now, this is the current legislation, the perspective of some courts, and ultimately AI reimplementation is a completely new subject, which, to my knowledge, has not yet been tested in EU courts. Until further changes, current legislation and case law have to be applied to completely new factual elements.
To conclude, the referring court in SAS case was fair enough to admit that computer programs are different from other kinds of literary works and present peculiar problems for copyright law …. although the general principles of copyright law must be applied to computer programs, it does not necessarily follow that particular doctrines which were established in relation to other types of literary works can be applied to computer programs, or applied in the same way (p. 198). Today, 16 years later, that conclusion is all the more relevant. Those peculiar problems of the past are becoming increasingly nuanced and tested to the maximum by the new technologies.
The purpose of this series was to analyze the Chardet case as a working hypothesis for testing the EU copyright framework. The maintainer’s decision to replace the old version directly with the new one, within the same project, raises separate questions, involving the project name, open source practices, GitHub rules, or the protection of the original author’s interests, topics that go beyond the discussion of protected expression addressed here.
Part 1: A real-world example: the relicensing of Chardet
Part 2: Terms of Service of AI platforms & EU main regulations for computer programs
Part 4: CJEU case law on computer programs
[1] https://www.bailii.org/ew/cases/EWHC/Ch/2010/1829.html
[2] Hotarâre nr. 1121/2015 din 07.05.2015 pronunțată de Tribunalul Iași, cod RJ 525e969e8 (https://www.rejust.ro/juris/525e969e8)
[3] Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 OJ L, 2024/1689, 12.07.2024
The content of this website is provided for informational purposes only and does not constitute legal advice. The law in this area is subject to ongoing developments. Please consult a qualified lawyer for advice specific to your situation.