AI Act: transparency obligations start this month, oversight moves to the AI Office

In Tech, the morning edition looked at model price lists and company releases. This analysis changes subject: it concerns the regulatory framework within which those models must operate in the European Union, and it is a story the morning edition did not cover.

The transition under way is twofold. On one hand, the transparency rules set out in the European regulation on artificial intelligence come into force this month, August 2026. On the other, as of 2 August 2026, the AI Office and the authorities of the member states are responsible for the implementation, oversight and enforcement of the AI Act, with powers over general-purpose models — the so-called GPAI, meaning models not designed for a specific task — concentrated in the AI Office (European Commission). The news currently comes from a single source (the European Commission’s official documentation on the AI Act); no independent confirmation available, as this is a primary institutional source.

The division of powers is the detail that matters more than the dates. In the regulation’s structure, oversight of systems falls largely to national authorities, each within its own jurisdiction, while general-purpose models — those on which the applications intended for the public are then built — answer to a single European counterpart. It is a choice that reduces the risk of diverging interpretations among national authorities on the segment of the supply chain from which public-facing applications originate: for those supplying these models there is a single counterpart, and this is as much a constraint as a simplification.

On the transparency front, the operational tool published by the Commission is a voluntary Code of Conduct on marking and labelling AI-generated content, accompanied by a set of icons to flag artificially produced images, audio and text, including falsified audiovisual content. Two adjectives should be kept in mind: the code is voluntary, and the icons are a signalling tool, not a detection tool. No label on its own identifies synthetic content: it declares what whoever publishes it chooses to declare. The value of a system of this kind therefore depends on how many platforms and how many providers adopt it, and on this point the Commission’s documentation does not, in the material available, provide a list of adoptions.

There is also an element that tempers the reading of the regulation as a blanket constraint: the AI Act does not introduce obligations for systems classified as minimal or no risk, a category in which — according to the Commission itself — the vast majority of AI systems used in the Union fall. The regulation’s architecture, in other words, concentrates requirements on a minority of high-risk cases and on general-purpose models, leaving most ordinary applications outside its scope. This is a useful clarification for reading the public debate on the topic, where the regulation is sometimes described as an obligation extending to any use of these tools.

The predictable point of friction concerns the synchronisation between regulatory and industrial timelines. The transparency rules come into force in a month in which providers announce new versions of their models just weeks apart from one another: every documentation and marking obligation applies to a target that moves faster than the verification cycles.

For readers wishing to gauge the effectiveness of this phase, there is one concrete indicator: which providers of general-purpose models will join the voluntary marking code, and how quickly the AI Office will make public its first oversight actions on GPAI.