AI and the Redefinition of Law in Code: The Future of Legal Intention under Article 4 DSM

Across Europe, rules once written for human readers are being interpreted by machines. This quiet transformation comes into view through Article 4(3) of Directive (EU) 2019/790 on Copyright in the Digital Single Market (DSM Copyright Directive). The Article allows text and data mining where a user has lawful access, unless a rightholder has ‘expressly reserved’ that right ‘in an appropriate manner’. What counts as ‘appropriate’ now reaches the Court of Justice of the European Union (CJEU) in Like Company Kft v Google Ireland Ltd (C-250/25, reference lodged 3 April 2025) following a request from the Budapest Környéki Törvényszék (Budapest District Court), which asks how a rightholder can reserve rights effectively when artificial-intelligence models absorb vast quantities of online material, and how such reservations interact with reproduction and communication to the public under the Directive 2001/29/EC (the InfoSoc Directive).

In Germany, the Hamburg Regional Court confronted the question first in Robert Kneschke v LAION e.V. (27 Sept 2024). The claimant, a professional photographer, found his image inside the LAION 5B dataset, amongst billions of image-text pairs drawn from public URLs. The court held that the dataset’s creation served research within § 60d UrhG, implementing Article 3 of the DSM Copyright Directive, and dismissed the claim. Yet the judgment added that Article 4(3) of the DSM Copyright Directive does not require any particular technical form for reservations; a clear human-readable statement, visible and unambiguous, could suffice. The court called such a notice an Erklärung, which means a declaration of will, and thus a communicative act.

In the Netherlands, the Amsterdam District Court reached a different conclusion in DPG Media v Knowledge Exchange B.V. (30 October 2024). This analysis is based on the official Dutch-language judgment published on rechtspraak.nl, reviewed through an unofficial Google translation, and verified against the English summaries provided by IP Kat. The case concerned the automated reuse of publishers’ RSS feeds by a news-aggregation service. The court read Recital 18 of the DSM Copyright Directive as a binding instruction and held that, for content made publicly available online, a rights reservation must be expressed in machine-readable form such as metadata, tags, or other technical means; otherwise, the Article 4(3) exception remains available. Through this reasoning, the Amsterdam judges treated legal communication as a technical protocol. They accepted that in a network of automated systems, meaning must be legible to machines as well as to people. The judgment marks a quiet but decisive move away from the understanding of law as a language of human expression and towards recognising it as a structure that must also be executed by code.

The Hamburg and Amsterdam conclusions reveal two instincts about law’s nature. One conceives law as a human-facing declaration; the other treats it as a machine-facing protocol. Both readings find textual support in Article 4(3) of the DSM Copyright Directive, yet they cannot coexist indefinitely. Artificial intelligence, which both reads and acts upon these signals, forces a choice between communication and code.

The Hungarian decisions take the discussion further. The Budapest Court of Appeal has already treated large-scale web-scraping and indexing as text and data mining within Article 4 DSM Copyright Directive. Building on that premise, the Budapest Környéki Törvényszék (Budapest District Court) referred the matter in Like Company Kft v Google Ireland Ltd (C-250/25, reference lodged 3 April 2025) to the CJEU. The questions put to the Court are whether model training involves a reproduction under the InfoSoc Directive, whether generated output amounts to a communication to the public, how lawful access should be understood for the purposes of Article 4 DSM, and what qualifies as an ‘appropriate’ reservation of rights in that framework.

If the CJEU endorses this analysis, Article 4 of the DSM Copyright Directive will govern not only AI training but virtually all automated reading of the open web. The line between lawful learning and infringement will depend on how clearly rightholders express their reservations.

A contextual standard can reconcile them. For open-web content, a reservation should be machine-readable; for access-controlled material, a human-readable statement should suffice. This solution reflects proportionality and links copyright with AI governance under the Regulation (EU) 2024/1689 on artificial intelligence. Article 53 of that Regulation requires general-purpose AI providers to maintain a copyright-compliance policy identifying and respecting reservations made under Article 4(3) of the DSM Copyright Directive. The European Commission’s Code of Practice for General-Purpose AI, adopted on 10 July 2025, confirms that reservations may appear in either form and expects providers to detect both. The document explains that general-purpose AI developers must use state-of-the-art technical and organisational measures to identify and respect copyright reservations expressed under Article 4(3) of the DSM Copyright Directive. A contextual interpretation from the CJEU that mirrors this understanding would bind the copyright and AI governance frameworks into a single, coherent structure.

Rigid formality invites inequality. If only machine readable reservations count, power passes to large publishers and platforms able to encode them, while smaller actors lose any practical means to assert their rights; if plain language alone suffices, developers face an unworkable compliance burden. A graduated rule aligns responsibility with capacity. The Union’s law has long assumed that authority rests on rules a person can understand, an assumption given constitutional force by the Court’s reading of Article 19(1) TEU as a guarantee of effective judicial protection in Associação Sindical dos Juízes Portugueses (C-64/16), a line of authority developed in Commission v Poland (C-619/18). That guarantee presupposes norms that can be known and contested, which is why the Court’s general principle of legal certainty insists that rules be clear, precise and foreseeable, as stated in Duff and Others (C-63/93) para 20 and VEMW (C-17/03) para 80. Read together, those authorities frame the present dilemma. If reservations are expressed only in code, we risk a model of compliance that is executable by machines yet opaque to citizens; if they are expressed only in words, we deny effect to the automated systems that now mediate access to law. The task for the CJEU is to reconcile the Article 19(1) TEU guarantee of an intelligible path to judicial protection with the settled demand of legal certainty for clarity and foreseeability, so that the boundary between lawful learning and infringement depends on signals that a person can read and a system can detect.

When the CJEU delivers its judgment in Like Company Kft v Google Ireland Ltd, it will not merely clarify a copyright exception; it will decide how EU law communicates intention in a world shared with machines. If it demands machine-readability everywhere, legality will depend on protocol. If it accepts contextual formality, law will remain bilingual, able to speak to both humans and systems without losing its voice. Either way, Article 4(3) of the DSM Copyright Directive will be remembered as the moment European law faced its own translation into code.

Verification note: All cited materials correspond to official EU or national judiciary publications available through EUR-Lex, Curia, WIPO Lex, rechtspraak.nl, the European Parliament, and the European Commission, verified as of 25 October 2025.

Ian Gauci is the Managing Partner at GTG, Malta.