{"id":21243,"date":"2026-08-21T09:57:40","date_gmt":"2026-08-21T09:57:40","guid":{"rendered":"https:\/\/letslaw.es\/?p=21243"},"modified":"2026-08-21T09:58:24","modified_gmt":"2026-08-21T09:58:24","slug":"code-of-practice-on-transparency-of-ai-generated-content-what-it-requires-from-your-business-as-of-2-august-2026","status":"publish","type":"post","link":"https:\/\/letslaw.es\/en\/code-of-practice-on-transparency-of-ai-generated-content-what-it-requires-from-your-business-as-of-2-august-2026\/","title":{"rendered":"Code of Practice on Transparency of AI-Generated Content: what it requires from your business as of 2 August 2026"},"content":{"rendered":"

The transparency obligations under Article 50 of the AI Act<\/a> are enforceable as from 2 August 2026, and the framework developing them has been finalised over the past few weeks. The Code of Practice on Transparency of AI-Generated Content<\/a> was published on 10 June, the European Commission confirmed its adequacy on 8 July, and the AI Board did likewise the following day. On 20 July, the final Guidelines on Article 50 were adopted. The widespread perception is that all of this is aimed at the major model developers, and that is precisely where the mistake lies: the part of the Code that affects the greatest number of companies is not the one addressed to providers, but the one addressed to those who simply use these tools to generate or publish content.<\/p>\n

What is the Code of Practice on Transparency of AI-Generated Content<\/h2>\n

The Code is a document drawn up by independent experts in a multi-stakeholder process facilitated by the European AI Office and published on 10 June 2026. It is structured in two sections that are signed independently of one another: Section 1 sets out the commitments of providers of generative AI systems<\/a> as regards marking and detection (Article 50(2) AI Act), and Section 2 those of deployers as regards the labelling of deep fakes and of text published on matters of public interest (Article 50(4)). A single organisation may be both at the same time, in which case both sections apply to it.<\/p>\n

Adherence is voluntary, but the obligations under Article 50 are not. This distinction is the key to the entire scheme and should be borne in mind from the outset. The Code creates no new duties and does not extend the existing ones: it merely specifies how to comply with those already laid down in the Regulation. Indeed, its own text expressly differentiates between the measures using the verb will<\/em>, which are mandatory for a signatory to be considered compliant with Article 50 and whose fulfilment will be monitored by the market surveillance authorities, and those using encouraged<\/em> or may<\/em>, which are purely voluntary.<\/p>\n

On 8 July 2026 the Commission concluded that the Code adequately covers the obligations laid down in paragraphs 2, 4 and 5 of Article 50 and facilitates their effective implementation. The following day, the AI Board adopted its own adequacy assessment. It is this twofold pronouncement that turns the Code into the reference instrument at European level: signatories may rely on its measures to demonstrate compliance irrespective of where they are established and of which national authority supervises them. Those choosing to comply by other means will have to prove that such means are adequate, and that assessment will be carried out individually by each market surveillance authority. The public list of initial signatories is published in July, before the obligations become enforceable, although signing remains open thereafter.<\/p>\n

Transparency and labelling obligations for AI-generated content<\/h2>\n

On the providers’ side, the Code starts from a relevant technical finding: as matters stand, no marking technique on its own satisfies the four requirements of Article 50(2) (effectiveness, interoperability, robustness and reliability). Hence its requirement of a multi-layered approach, combining digitally signed metadata with imperceptible watermarks. There are narrowly defined exceptions: free-form text cannot carry metadata, so that a single layer suffices, and very short text, understood as text below 200 tokens, falls outside the watermarking obligation. Providers must also make available a detection solution that is accessible and, as a general rule, free of charge, a reasonable fee being permitted only for those with fewer than one million monthly users and only in the face of disproportionate query volumes. Free and unrestricted access is guaranteed in every case to authorities, the media, fact-checkers, independent researchers and civil society organisations. Placing on the market or promoting tools whose purpose is to circumvent the markings is expressly prohibited, and 2 February 2027 is set as the deadline for implementing an interoperability solution between detection mechanisms.<\/p>\n

Section 2 is the one that affects the majority of companies, because very few develop generative AI systems but almost all of them use such systems. Its centrepiece is the EU icon included in Annex 1, available in three variants that are free to use and require no attribution: one for content that is entirely AI-generated, another for partially modified content, and a third, basic one designed to be supplemented with an interactive layer. The user testing carried out across several Member States showed that the variants bearing an explicit textual label, of the generated<\/em> or modified<\/em> type, perform significantly better in terms of perception and clarity than the icon on its own. An equivalent proprietary icon may be used, provided that it complies with the design and placement specifications.<\/p>\n

The placement rules are considerably more specific than the wording of the Regulation would suggest. The label must be perceivable without requiring user interaction or sustained attention, at a point where no overlay elements exist, such as the top right corner of an image or a video. In video it must be displayed at the beginning and, as far as possible, at regular intervals and at a minimum after each advertising break, precisely in order to cover screenshots and clipped fragments. In published text it is placed above the content, near the headline or in the opening colophon. Where visual disclosure is not possible, as is the case with audio, a short audible disclaimer in plain language at the beginning is called for, supplemented by periodic reminders. All of this must additionally meet the accessibility requirements of Directives (EU) 2019\/882 and 2016\/2102, with reference to standard ETSI EN 301 549 and to the WCAG 2.1 Level AA guidelines.<\/p>\n

Two specific regimes deserve attention. The first is that of artistic, creative, satirical or fictional works, where disclosure must be made without hampering the enjoyment or the normal exploitation of the work; for example, in the accompanying notes, in the credits, in an exhibition leaflet or on the entrance ticket. The second, and probably the most underrated, is the exception of human review and editorial control<\/a> for published text. That exception does not operate automatically by virtue of the mere fact that someone has read the text before publishing it: the Code requires identifying the natural or legal person assuming editorial responsibility, with name, role and contact details, describing the organisational and human resources allocated to that review, and publishing those contact details where they are not already available. If your organisation intends to rely on this route, you need to be able to evidence it in documentary form.<\/p>\n

The relationship between the Code of Practice and Article 50 of the AI Act<\/h2>\n

It is worth putting the three pieces in order, because they are easily confused. Regulation (EU) 2024\/1689 contains the legal obligation. The Commission Guidelines on Article 50, adopted on 20 July 2026, interpret its scope: they clarify which providers and deployers it reaches, which content falls within the scope of application and how the exceptions apply. They are not binding, and only the Court of Justice may give an authoritative interpretation, but in practice they will be the reference followed by the national market surveillance authorities. The Code, for its part, does not interpret the rule: it offers a catalogue of concrete measures that make it possible to demonstrate compliance with it.<\/p>\n

There is a limit to its coverage that is worth underlining. The Code encompasses paragraphs 2, 4 and 5 of Article 50, but neither paragraph 1 nor paragraph 3. That is to say, the obligation to inform a person that they are interacting with an AI system, unless this is obvious from the context, and the obligation to give notice of the use of emotion recognition or biometric categorisation systems remain fully enforceable and must be complied with by other means. Signing the Code does not cover those fronts. It is a frequent mistake to assume that adherence resolves Article 50 as a whole.<\/p>\n

As regards the timeline, Article 50 applies as from 2 August 2026. Regulation (EU) 2026\/1744, known as the Digital Omnibus on AI, was published in the Official Journal on 24 July and has been in force since 27 July, and it does not alter that date: what it defers is the regime governing high-risk systems. It does introduce a transitional period for systems placed on the market before 2 August 2026, which have until 2 December 2026 to bring their marking into line. The Guidelines further specify that there is no retroactive labelling obligation: for images, audio and video the relevant date is that of the generation of the content, whereas for text on matters of public interest the decisive factor is the date of publication, so that text generated before 2 August but published thereafter must indeed be labelled.<\/p>\n

Failure to comply with the transparency obligations under Article 50 may be penalised with fines of up to EUR 15 million or 3% of the worldwide annual turnover for the preceding financial year, whichever is higher, subject to the proportionality regime of Article 99(6) for SMEs and start-ups, which applies the lower of the two amounts. And one final clarification on the decision whether to sign: adhering to the Code does not exempt anyone from complying, just as not signing it does not amount to non-compliance. The real difference lies in the burden of proof. The signatory demonstrates its conformity by reference to a framework already validated by the Commission and the AI Board; those who do not sign must build that demonstration on their own and defend it before the authority having jurisdiction over them.<\/p>\n

How Letslaw can help you<\/h2>\n

At Letslaw we are specialists in digital regulation and artificial intelligence. If your company generates, edits or publishes content using AI tools, we can help you determine whether you are acting as a provider or as a deployer, review your labelling and editorial control processes, and assess whether adhering to the Code of Practice is advisable for you. Do not hesitate to contact us<\/a>.<\/p>\n

Contact Us<\/h2>\n
\n

<\/p>

    <\/ul><\/div>\n
    \n
    \n<\/fieldset>\n
    <\/span><\/div>\n
    <\/span><\/div>\n
    <\/span><\/div>\n
    <\/span><\/div>\n