The European Commission has published a detailed guide for labelling AI-generated content – a practical tool aimed at companies building and using generative AI. The guide was released on 10 June 2026 as a voluntary code of practice, but it is closely tied to legal requirements under the EU's AI Regulation (AI Act) that take effect from 2 August this year, according to AI News.

For Norwegian companies with operations in the EU market, the deadline is already close.

What the law actually requires

The AI Act mandates clear labelling of content that is substantially generated by artificial intelligence. This applies in particular to three categories: deepfakes that appear to be genuine recordings of real people or events, AI-generated text published to inform the public on matters of public interest, and interactions with chatbots and other AI systems where the user does not necessarily know who – or what – they are communicating with.

The code of practice was developed with input from more than 180 stakeholders, including technology providers, academia, and civil society. It recommends the use of invisible digital watermarks and accessible detection tools, as well as common EU icons for AI labelling.

6%
Maximum fine of global turnover
3 hours
Deadline for removal of unlawful synthetic content
EU law forces Norwegian AI companies to label content from August - Bilde 1

Loopholes concern experts

Despite the ambitions, there is widespread expert criticism of the regulation's practical reach. A central objection is that the requirements are relatively narrow: labelling is primarily mandatory for deepfakes and text of public interest, not for AI-assisted content in general. An earlier classification intended to distinguish between "AI-generated" and "AI-assisted" content was removed from the code of practice precisely because the boundaries were deemed too unclear.

The regulation also contains explicit exemptions. AI-generated text is exempt from labelling requirements if it undergoes "genuine human review or editorial oversight" and someone assumes editorial responsibility – but this requires documented procedures that can demonstrate such oversight. Artistic, satirical, and fictional content is likewise exempt.

The requirements presuppose a baseline level of AI literacy in the population that does not yet exist evenly across society

In addition, transparency requirements relating to the public sector are limited: law enforcement and migration authorities are not required to publish summaries of fundamental rights impact assessments, and information about AI use in these sectors is stored in a non-public database. Private actors using high-risk AI are also not required to register their use in the EU's database.

Technical limitations challenge implementation

The labelling technology itself is not without problems. Metadata and digital watermarks are vulnerable – they can be destroyed by basic image editing or compression. Researchers point out that diverging labelling standards across countries and platforms make interoperability difficult to achieve, particularly since AI tools and digital content are effectively borderless in practice.

An AI label can actually make unlabelled images more credible – not just make labelled ones less credible

A study involving German Instagram users, cited in the research background, shows that AI labels have a measurable effect: they reduced the perceived authenticity of labelled images by approximately 9.1 percentage points. However, the study also revealed an unintended side effect – exposure to some labelled content led users to perceive unlabelled images as marginally more credible (an increase of approximately 1.9 percentage points). This suggests that sporadic labelling can backfire, and that full coverage under the regulation is essential for it to function as intended.

What Norwegian companies should do now

Norway is an EEA member state and continuously implements EU regulations. Although the timing of the national incorporation of the AI Act is still being clarified, there is little doubt that Norwegian companies providing services to the EU market must comply with the rules directly.

The Commission's code of practice provides concrete recommendations that companies can already begin to follow: implement watermarking technology, establish documented procedures for editorial oversight, and ensure that users are clearly informed when they are interacting with AI systems.

The deadline is 2 August 2026 – just under seven weeks away.