The context: transparency as a pillar of the AI Act
Yesterday, the European Commission published the final version of the Code of Conduct on the marking and labelling of AI-generated content, a strategic tool within the European regulatory framework for AI. The Code represents the operational response to the transparency obligations set out in Article 50 of the AI Act, which will come into force on 2 August 2026.
The AI Act, moreover, has identified transparency as one of the most urgent systemic risks to be addressed. When machine-generated content is indistinguishable from human-generated content — journalistic texts, deepfake videos, synthetic voices — it creates fertile ground for disinformation, the manipulation of public opinion and a loss of trust in the information ecosystem. Consider, for example, an apparently authentic video of a political leader circulated during an election campaign, a voice recording used to impersonate a company executive and authorise a bank transfer, or images of events that never took place circulating on social media during crisis situations. In all these cases, the ability to identify the artificial origin of the content plays an essential role in preserving user trust.
The Code of Conduct is the practical tool through which the Commission seeks to translate these regulatory principles into concrete actions.
What the Code provides for: obligations and target audience
The Code is voluntary in nature, but carries significant legal weight: companies that sign up to it will be able to rely on shared standards to more easily demonstrate compliance with the obligations set out in the AI Act in the areas covered by the Code itself. This mechanism makes it particularly attractive to industry.
There are two target groups.
Providers, i.e. those who develop (generative) AI systems. They are under an obligation to ensure that the content produced (audio, images, video, text) is marked in a machine-readable format and detectable as artificially generated or manipulated. The technical solutions adopted must be effective, interoperable, robust and reliable, in line with the state of the art and implementation costs.
Deployers, i.e. those who make generative AI systems available to end users. They are subject to disclosure obligations in two specific cases:
There is also a requirement to inform users when they interact with an interactive AI system, such as a chatbot or a virtual assistant.
The process: how the Code was developed
The Code is the result of a participatory process launched in September 2025 by the AI Office, involving a public consultation and a call for expressions of interest. The process involved two thematic working groups (one for providers and one for deployers), led by independent chairs and vice-chairs, and included a wide range of stakeholders: developers of detection technologies, trade associations, civil society organisations, academics and major online platforms.
Over seven months — from November 2025 to June 2026 — three interim drafts were produced and put out for consultation before the final version was finalised. In parallel, the Commission published interpretative guidelines to clarify the scope of regulatory obligations and cover aspects not addressed by the Code.
Practical implications: what changes for businesses and users
For technology companies, signing up to the Code means adopting technical standards for content labelling and traceability, such as digital watermarking, signed metadata and provenance standards such as C2PA, as well as implementing appropriate disclosure systems for end users.
For example, an image generated by an AI model could contain metadata that allows platforms and verification tools to automatically identify its artificial origin, even if the label visible to the user has been removed.
Companies that do not sign up will still have to comply with the obligations of the AI Act from 2 August 2026, but without the benefit of being able to refer to the shared standards identified by the Code.
For the public, the expected impact is significant: deepfakes and AI-generated texts on matters of public interest will have to be clearly identified, making it easier to recognise when one is dealing with synthetic content.
The question of technical feasibility remains open. Watermarks and metadata can be removed or altered, and the detection of synthetic content in distributed environments still poses a significant technological challenge. The Code itself acknowledges this limitation, requiring the adoption of ‘technically feasible’ solutions in light of the state of the art.
Outlook: a model for the rest of the world?
The European initiative comes at a time of intense global regulatory competition over artificial intelligence. Whilst the United States is proceeding with a more fragmented approach and China has adopted specific rules on synthetic content, the European Union aims to build a model based on transparency and the accountability of operators.
A scenario that is far from unlikely is one in which a global platform chooses to apply European labelling standards to all users, rather than developing different systems for each national market. This is a phenomenon already observed in the past with European data protection legislation.
The challenge will be to maintain a balance between transparency and innovation, ensuring that excessive compliance burdens do not put European businesses at a disadvantage compared to their global competitors. The Code of Conduct on represents a significant attempt to strike this balance, with August 2026 serving as the first real-world test.
