AI Act: transparency obligations come into force, conversational systems must identify themselves

On 2 August 2026 the part of the European AI Act devoted to transparency came into application. This is not a statement of intent: the European Commission announced the start of enforcement of the rules, and the related note from the DG Communication was published the same day.

What the rules require. Two distinct obligations. The first concerns interaction: conversational systems and other interactive systems must inform users that they are engaging with an artificial intelligence and not with a person. The second concerns content: images, videos and audio that are artificially generated or manipulated — so-called deepfakes — must be labelled, and generated content must carry machine-readable markings.

The difference between the two requirements is not merely formal. The visible label is aimed at those who view the content; the machine-readable marking is aimed at the systems that index, redistribute or collect that content to train other models. It is the second obligation, not the first, that affects platforms’ technical infrastructure, because it requires that information about synthetic origin survive the circulation of the file.

Who enforces the rules. The enforcement structure has three levels: national market surveillance authorities, the European Artificial Intelligence Office, and the European Data Protection Supervisor with regard to EU institutions. This is an architecture that distributes oversight among the authorities of individual member states, with central coordination. The point worth watching, for anyone wishing to gauge its effectiveness in the coming months, is potential unevenness: the same rule applied by national authorities with differing resources and priorities.

The regulatory context. On 7 May 2026 the Council and the European Parliament had agreed to simplify the rules, while retaining the framework of the Commission’s proposal on high-risk systems, as reported in the statement from the Council of the European Union. The sequence of events matters: first the easing agreed in May, then the start of enforcement in August. Anyone reading the second piece of news without the first risks describing a tightening; anyone reading the first without the second risks describing a retreat. The two events took place three months apart and concern different chapters of the same regulation.

What the available documents do not say. None of the three sources cited reports the scale of penalties envisaged for failure to comply with the transparency obligations, nor any concrete cases already opened by supervisory authorities, nor mandatory technical standards for machine-readable marking. This last point is the one destined to determine the rule’s practical effectiveness: without a shared standard, each provider can mark its own content in a format that other systems fail to recognise, and the obligation is reduced to the formal act of affixing a marking that no one reads.

The real scope. The rules apply to systems offered on the Union’s market, regardless of where their producers are based. Since the main widely used conversational systems are developed outside Europe and distributed globally, the European obligation to disclose the non-human nature of the interlocutor affects products designed elsewhere. This is the mechanism by which a regional rule ends up shaping global design choices, when adapting a product to a single market costs more than standardising it.

The first measurable test will be the appearance — or absence — of markings in generated content circulating on platforms accessible from the Union, given that the rule has been in application since 2 August and provides for no further transitional periods in the documents cited.