Article 50(2) of the Artificial Intelligence Act and the traceability of text
Doctor of European Union Law | President of the Euro-American Committee of Digital Law - CEA Digital Law
The European Union has built much of Regulation (EU) 2024/1689 of 13 June 2024 laying down harmonised rules on artificial intelligence (Artificial Intelligence Act, or AI Act) on a premise that, broadly speaking, is hard to dispute: not all uses of artificial intelligence give rise to the same risks and, therefore, not all of them should receive the same legal response. That risk-based approach is one of the principal political and regulatory justifications for the AI Act. That is precisely why it is striking that, on reaching Article 50(2), the legislature largely abandons that logic and subjects to the same basic traceability obligation content whose nature, function and capacity to deceive are radically different.
What Article 50(2) requires, and what it does not
The provision requires providers of artificial intelligence systems, including general-purpose models, that generate synthetic audio, image, video or text content to ensure that their outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated. The obligation gives way only where the artificial intelligence performs an assistive function for standard editing or does not substantially alter the input data provided by the user or the semantics thereof. Since 2 August 2026 the transparency obligations in Article 50 have, as a general rule, applied in the European Union, without prejudice to the transitional regime introduced by Regulation (EU) 2026/1744 of 8 July 2026, known as the Digital Omnibus on AI, for certain generative systems already placed on the market before that date, whose providers have until 2 December 2026 to comply with Article 50(2).
It is worth making clear from the outset what is under discussion, because criticism loses force if it is directed at something the provision does not say. Article 50(2) does not require a visible notice to be placed on every text announcing to the reader that artificial intelligence has been involved. That technical obligation falls on the provider and consists in making it possible to identify the artificial origin at a later stage. Nor should it be confused with Article 50(4), which governs certain disclosure duties incumbent on those who publish content, in particular deepfakes and text intended to inform the public on matters of public interest. The Commission has expressly confirmed that, in the latter case, text that has undergone genuine human review or editorial control, and for which a person holds editorial responsibility, need not be publicly labelled as generated by artificial intelligence.
This distinction improves the Regulation technically, but at the same time it reveals the problem the Regulation contains. The European legislature recognises, at the point of publication, that substantive human intervention and the assumption of responsibility may be legally sufficient to render the earlier involvement of an artificial intelligence irrelevant as far as the public is concerned. At the earlier generation stage, however, it maintains a general marking obligation that makes it possible to preserve precisely that piece of information about origin. There is no normative contradiction in the strict sense, because paragraphs 2 and 4 have different addressees and different immediate purposes, but there is a conceptual tension that deserves far more scrutiny than it has received: the Regulation accepts that human responsibility may legally displace the relevance of the AI’s involvement and, at the same time, establishes an infrastructure designed to keep that involvement detectable.
Audiovisual deepfakes and text: the same risk?
The problem becomes clear if an audiovisual deepfake is compared with a text. A hyper-realistic image can show a person committing a crime they never committed. A synthetic recording can reproduce the voice of a minister announcing a decision that does not exist. A video can place an election candidate at a meeting that never took place. In all those cases the artificiality of the content is part of the very mechanism of deception, because the aim is to replace a perceptible reality with another that never existed. Knowing that the image, the voice or the video was artificially generated provides information directly related to its authenticity.
The same is not necessarily true of text. The fact that a statement was formulated with the help of a language model does not determine whether it is true or false. Nor does it determine its quality, its legal originality, its scientific rigour or the existence of a person responsible for its content. A false fact remains false even if it was written from beginning to end by a human being. True information does not become doubtful because an artificial intelligence system helped to find it or draft it. A non-existent judgment cited by a lawyer is an extremely serious professional problem, but it is so because the decision does not exist and because the lawyer failed in the duty to check, not because a particular tool was used to prepare the submission.
An over-inclusive obligation for textual content
Here the first structural flaw in Article 50(2) appears: it takes technological origin as its principal regulatory category when, in the case of text, the legally relevant risk very often lies elsewhere. What is decisive may be falsity, manipulation, plagiarism, infringement of rights, lack of professional diligence, fraud or the absence of responsibility for what is published. Generation by means of artificial intelligence may increase some of those risks, particularly because of its capacity for mass production, but it does not in itself constitute any of them.
This is no minor difference. Genuinely risk-based regulation ought to ask what danger it seeks to prevent, how intensely that danger arises and what causal connection exists between the measure imposed and that danger. Article 50(2) proceeds the other way round in respect of text: it first identifies a technology, attaches legal relevance to its use and then requires a mark of origin practically regardless of the subsequent context in which the content is to be used. The result is a structurally over-inclusive obligation. It will catch content intended for disinformation, but it will also catch professional reports, draft contracts, academic work, economic analyses, literary texts, private correspondence and any other textual output that has been substantially generated by one of these systems and does not fall within the exceptions in the provision itself.
The fact that the obligation formally falls on the provider does not remove the problem. The provider introduces the mark, but its communicative consequences are also borne by the person who uses the tool. The legal question is not merely who must technically implement the mechanism, but what effect that mechanism has on the intellectual product and on the person who subsequently adopts it as their own.
The principle of proportionality and Article 50(2)
The principle of proportionality offers far more solid ground for debate here than a general challenge to the Union’s competence. Article 5(4) of the Treaty on European Union (TEU) provides that the content and form of Union action shall not exceed what is necessary to achieve the objectives of the Treaties. Where, in addition, the action affects rights recognised by the Charter of Fundamental Rights of the European Union (the Charter), Article 52(1) of the Charter requires any limitation to respect the essence of those rights and to be necessary and proportionate to the legitimate objective pursued.
No one seriously disputes that combating disinformation, fraud, impersonation and manipulation is a legitimate objective. Recital 133 of the AI Act states this clearly by linking these obligations to the risks that synthetic content poses to the integrity of the information ecosystem. That is precisely why the problem lies not in the legitimacy of the objective but in the scope of the measure chosen to achieve it. The recital itself groups under a single logic the need to trace content whose capacity to deceive is extraordinarily disparate.
The legislature’s discretion and the Philip Morris judgment
Judicial review of a legislative choice of this kind would not be straightforward. The Court of Justice grants the EU legislature broad discretion where it must take complex decisions of a political, economic or technical nature. The case law on proportionality and on the discretion of the EU legislature and, of particular relevance, the judgment of the Court of Justice of the European Union of 4 May 2016, Philip Morris Brands SARL and Others v Secretary of State for Health, Case C-547/14, ECLI:EU:C:2016:325, show that it is not enough to establish that another solution would have been possible or even preferable. To declare a legislative measure invalid for disproportionality, it will normally be necessary to show that the legislature has exceeded the limits of its discretion.
But that discretion is not immunity. Nor does it prevent one from asking whether the general inclusion of text in Article 50(2) genuinely responds to the risk that the Regulation seeks to neutralise. Far from closing the debate, the comparison with the Philip Morris judgment allows it to be framed more precisely. There, the Court accepted restrictions on the presentation of products whose very marketing was associated with an established health risk, and in a very specific subject area. Generative artificial intelligence is not in itself a harmful product, nor does text assisted by it carry an equivalent inherent dangerousness. We are dealing with a general-purpose technology capable of taking part in practically any intellectual activity. The more cross-cutting the tool and the more heterogeneous its uses, the more important it becomes to justify why all of them should be subject to the same provenance trace.
Human review and hybrid authorship of text
Article 50(4) itself, moreover, provides a conceptual alternative that ought to have been explored in greater depth in paragraph 2. For certain textual publications of public interest, the Regulation regards as sufficient the existence of human review or editorial control and of a natural or legal person who assumes responsibility. The Commission has interpreted that review in demanding terms: a spelling or formal correction is not enough; there must be a substantive examination carried out by people with professional knowledge and judgment, including the possibility of checking facts and verifying the reliability of sources.
If that human control makes it possible to dispense with public disclosure of an AI’s involvement, it is hard to understand why the legal order does not provide an equivalent mechanism with regard to the possibility that the technical mark continues to allow the artificial origin to be detected when the text has undergone genuine human intellectual appropriation. This would not be about allowing anyone freely to remove the traceability of content in order to deceive third parties. It would be about recognising in law that an intellectual product may change in nature as it passes through different stages of preparation and that the origin of an initial version need not define for ever the status of the final version.
That is probably the most problematic aspect of the rules. Artificial intelligence forces us to abandon a binary conception of authorship that no longer adequately describes reality. Between a text written entirely without technological assistance and one generated automatically from a minimal instruction lies a vast field of hybrid production. A professional may contribute the thesis, the documentation, the methodology, the selection of sources and the conclusion; then use an AI to test hypotheses, reorganise arguments or produce successive versions; reject much of the output; modify another part of it; verify every reference; introduce new reasoning; and finally adopt the resulting text as their own. Reducing that process to the conceptual label of ‘AI-generated content’ provides very little information about where the intellectual contribution actually lies.
Origin, authorship and responsibility are distinct categories. Article 50(2) regulates the first. The risk arises when the infrastructure created to identify origin ends up being used socially or institutionally as a presumption about the other two.
The risk of stigmatising those who use AI
That risk is no longer theoretical. A particularly impoverished way of assessing intellectual work is spreading, according to which detecting the presence of artificial intelligence seems enough to diminish the value of a text. Checking the reasoning matters less than discovering how it was drafted; knowing whether the sources are correct is of less interest than determining whether there is a statistical signal of automatic generation. The debate then ceases to revolve around the quality of knowledge and begins to revolve around a sort of technological purity of the process by which it was produced.
The law should be especially careful before contributing to that cultural shift. An obligation created for the legitimate purpose of ensuring transparency may end up providing the infrastructure for the regulatory stigmatisation of those who make intensive use of artificial intelligence tools. There is no need to prohibit their use to produce that effect. It is enough to turn their use into a circumstance that must be capable of being discovered later and to allow universities, companies, public administrations, courts or the media to attribute to it a meaning that the provenance data itself cannot justify.
The experience of the law offers a useful contrast. Lawyers have been building arguments on other people’s material for centuries. They use forms, treatises, law reports, contract templates, reports from colleagues, databases and passages taken from earlier decisions. Courts reproduce earlier legal scholarship and case law, sometimes at length. None of those practices in itself determines the quality of the result. The deficiency arises when the material used is incorrect, bears no relation to the case at hand, is incorporated without critical examination or leads to merely apparent reasoning. Legal judgment turns on relevance and on the intellectual responsibility of whoever adopts the result, not on whether each sentence originally came from their own typing.
The arrival of artificial intelligence increases the capacity for delegation and, of course, demands higher standards of verification. Precisely because an AI can produce, with enormous fluency, a non-existent citation or an argument that is only apparently convincing, the professional who uses it has a heightened duty to verify what they make their own. That duty is entirely compatible with rejecting the idea that the use of artificial intelligence in itself diminishes the authorship or intellectual value of the result. These are different questions, and conflating them leads to conceptually flawed regulation.
Does the watermark affect freedom of expression?
There is also a dimension connected with freedom of expression, recognised in Article 11 of the Charter of Fundamental Rights of the European Union. Legal caution is called for. A technical mark of origin is not exactly the classic case of compelled speech, and there is, to date, no case law of the Court of Justice holding that Article 50(2), on that ground alone, infringes freedom of expression. It would be wrong to claim so. But it would be equally wrong to maintain that there is no constitutional issue at all because the mark is not visible. Article 11 protects the process of imparting and receiving information and ideas, and Article 52(1) requires public interference with that process to be examined in the light of proportionality. Where the law requires the generated output to incorporate a technical feature designed to enable third parties to detect its artificial origin, there is at least an effect on the architecture of that communication that warrants legal scrutiny.
The question can be put without grandiloquence. If a person substantively revises a text, checks all its sources, changes whatever they consider necessary and takes full ownership of its content, what public interest requires that a third party should still be able to know that an earlier stage of its preparation was assisted by artificial intelligence? If the answer is to prevent the recipient from believing that the text was produced without AI, it remains to be shown why that circumstance is legally relevant in every context and why the responsibility assumed by the author is not a sufficiently effective alternative. Transparency is not an absolute value detached from its purpose. Making transparent a piece of information that lacks substantive relevance may end up producing more distortion than information.
Equal treatment of audio, image, video and text
The principle of equality in Article 20 of the Charter adds a complementary argument, although I do not regard it as the main one. European case law requires comparable situations to be treated alike and accepts that different situations may require different treatment. The joint inclusion of audio, image, video and text is not automatically contrary to that principle; the legislature may establish general categories. However, where the difference between the risks is so marked, that homogenisation reinforces the need for objective and proportionate justification. A sexual deepfake, an imitation of the voice of a head of state and a legal report prepared with the assistance of an AI are synthetic products in a technological sense, but only excessively abstract regulation could consider that this common feature exhausts the relevant legal analysis.
Can the Commission’s guidelines correct the provision?
The European Commission itself seems partly aware of the difficulty when it insists that technical solutions must take account of the specificities and limitations of each type of content. The Regulation contains that requirement, and the European Commission Guidelines of 20 July 2026 on the transparency obligations of providers and deployers of AI systems under Article 50 of the AI Act seek to guide coherent and proportionate application, also introducing adjustments for certain machine-to-machine exchanges and for closed-loop industrial or product development environments. The Code of Practice on Transparency of AI-Generated Content, moreover, is voluntary and is intended to facilitate compliance with Article 50.
But neither interpretative guidelines nor a voluntary code can correct any disproportionality located in the legislative provision itself. They can delimit concepts, guide supervisory authorities and avoid certain absurd applications, but they cannot introduce an exception that Article 50(2) does not contain, nor can they replace the political decision taken by Parliament and the Council. This point is particularly important because there is a risk that the regulatory debate will quietly shift from what the legislature adopted towards the technical criteria subsequently developed by the Commission, standardisation bodies and the companies themselves. Technical flexibility is necessary, but defining the limits of a restriction of this kind cannot be left indefinitely to soft-law instruments.
Nor should what Europe requires be confused with the compliance decisions that each company takes. A provider’s decision to apply a particular watermarking system worldwide, or to build its architecture more broadly than is strictly necessary for the European market, may reflect technical or commercial reasons or a wish to simplify operations. That business decision may be open to criticism on its own terms, but it does not alter the heart of the matter: it is Article 50(2) that has made the marking of AI-generated textual content a European legal obligation. A company’s possible overreaction does not remove the need to examine whether the mandate that prompts it is properly calibrated.
An alternative: graduated traceability for text
Nor does the regulatory alternative require transparency to be dismantled. It would be enough to draw better distinctions. The provider could be required initially to embed provenance mechanisms in those outputs that fall within the scope of Article 50(2), but the legal order could recognise a legally regulated route whereby, following substantive human intervention and an unequivocal assumption of editorial or professional responsibility, artificial origin would no longer necessarily operate as a relevant technical feature of the final version. Mandatory traceability could be maintained in full for content whose artificiality directly affects its authenticity, particularly in cases of audiovisual impersonation, and graduated in the textual sphere according to the nature of the use and the degree of human intervention.
Nor would that solution have to rest on a mere declaration by the user. Review standards, professional certification schemes or technical mechanisms could be established that are capable of replacing a mark of artificial generation with a declaration of human responsibility for the final version. What matters is not to design a specific technology now, but to recognise the principle: a society that is going to work in a structurally hybrid way cannot build its authorship regime around the fiction that all content belongs for ever to the machine or to the human being according to who produced its first words.
How could Article 50(2) be challenged?
From a procedural point of view, talk of overreach also calls for precision. The weakest argument would be simply to maintain that the Union lacked competence to regulate this matter. The AI Act is based, among other legal bases, on Article 114 of the Treaty on the Functioning of the European Union (TFEU), and the case law affords considerable latitude to harmonisation where national divergences capable of affecting the internal market exist or may arise. The real debate lies not so much in whether the Union could act as in how far it could go in doing so.
Nor can the possible invalidity of Article 50(2) be raised today as though any user could go directly to the Court of Justice to seek its annulment. We are dealing with a legislative act, and the rules on standing under Article 263 TFEU are restrictive. The judgment of the Court of Justice of 3 October 2013, Inuit Tapiriit Kanatami and Others v Parliament and Council, Case C-583/11 P, ECLI:EU:C:2013:625, confirmed that the concept of ‘regulatory act’ in the fourth paragraph of Article 263 TFEU does not cover legislative acts. In addition, the ordinary time limit for bringing a direct action against the Regulation has expired. The most realistic legal route would arise when a national or EU measure of implementation, supervision or sanction brought the provision into operation in a specific dispute. In that context, a question as to the validity of the EU provision could be raised before the competent court and, where appropriate, give rise to a reference for a preliminary ruling to the Court of Justice under Article 267 TFEU. The case law established by the judgment of the Court of Justice of 22 October 1987, Foto-Frost v Hauptzollamt Lübeck-Ost, Case 314/85, ECLI:EU:C:1987:452, reserves precisely to the Court of Justice the power to declare acts of the Union institutions invalid.
Any challenge would have to be constructed with considerably more rigour than a political accusation of regulatory excess. It would have to be shown that, as regards textual content, the general marking obligation is not sufficiently closely related to the risks relied upon, that less restrictive alternatives exist that are capable of achieving an equivalent level of protection, and that the rules produce significant effects on legally protected rights or interests that were not properly weighed. Article 5(4) TEU, Articles 11 and 20 of the Charter and Article 52(1) of the Charter itself would provide the analytical framework, but success would depend on proving specific disproportionality rather than on invoking those provisions as rhetorical formulas.
Before litigation: a proportionate interpretation
There is, moreover, a path that comes before any litigation on validity: insisting on a proportionate, teleological and fundamental-rights-compliant interpretation of the expressions that delimit Article 50(2). Concepts such as ‘assistive function for standard editing’ or substantial alteration of the semantics should not become token exceptions always interpreted in favour of maximum traceability. If the Regulation genuinely intends to maintain a risk-based approach, those concepts must serve to prevent mere technological assistance from automatically turning a human text into artificial content subject to a technical mechanism making its origin detectable.
Regulating harm, not the intellectual process
The underlying debate, however, goes beyond the scope of any single provision. Europe runs the risk of institutionalising a misplaced suspicion precisely at the moment when artificial intelligence is becoming an ordinary tool of intellectual work. If within a few years practically all professionals use generative systems to search, cross-check, organise, translate, code or draft, continuing to divide intellectual output into human content and AI content will be as uninformative as asking today whether a contract was written on a word processor or by hand. What will remain decisive is who adopts it and who answers for it.
If a lawyer signs a pleading, they answer for its arguments and its citations. If a researcher publishes a piece of work, they answer for its data and its methodology. If a journalist disseminates information, they answer in accordance with the rules of their profession and the legal order. Artificial intelligence may be involved in each of those processes, and that involvement may be relevant in certain contexts, but it should not operate as a presumption of lesser intellectual value or as a free-standing indication of wrongdoing.
The European Union is right to tackle the risks of deepfakes, impersonation and the automated production of disinformation. It even has respectable reasons for building technical provenance standards. What should not be accepted without debate is the logical leap from those legitimate objectives to a general traceability of text whose real basis ends up being the mere fact of having been produced with a particular technology.
The law should not protect a supposed human purity of writing. It should protect people against deception, hold to account those who cause harm and establish safeguards where there is a real risk. When it turns the tool used to think or to draft into a free-standing object of suspicion, it stops regulating harm and starts regulating the intellectual process.
That is the point at which Article 50(2) deserves to be questioned. Not because artificial intelligence should be left outside the law, but precisely because technologically mature regulation ought to be able to distinguish between falsifying reality and using a machine to work on it. The difference is too important to be hidden inside a watermark.
النشرة
فعاليات وأخبار وتحليلات في القانون الرقمي تصلك على بريدك، ولا نراسلك إلا حين يكون لدينا جديد.
