EU AI Act: Content labelling obligation

Since 2 August, Article 50 of the EU AI Act has been in force, requiring content generated or modified with AI to be marked and labelled in a machine-detectable way, in line with the transparency obligations set out in the regulation (Articles 2, 4 and 5).

To support this, the European Commission and the EU AI Board drafted and put in place a voluntary code of practice, which the major AI providers have already signed up to. You can download it here: https://ec.europa.eu/newsroom/dae/redirection/document/129555

• OpenAI: early this month it introduced a text watermark called textGrain, initially applied only to EU users over the coming weeks. It has committed to open-sourcing the marker. For now, the detector is available only to approved organisations and researchers.
• Google: a pioneer in this space, it created SynthID to watermark video, audio, images and text, and released it as open source. Gemini can check with a simple query whether content was generated by Google's ecosystem.
• Anthropic: everything Claude generates worldwide is watermarked, but detection is still in private preview. Access can be requested through the Claude watermark detector access form: https://forms.gle/9tGA33hPJJwtHsMk9

In addition, the EU has published a set of icons that deployers of generative AI systems can use to label content generated or manipulated by AI. Exceptions and limitations apply in some cases. You can access them here: https://digital-strategy.ec.europa.eu/en/policies/eu-icons-labelling-ai-generated-content

What is the impact?

If a company publishes AI-generated text to inform the public on matters of general interest, it must disclose it, unless there is genuine human review and someone takes editorial responsibility. Deepfakes must always be labelled, unless they are evidently satirical, creative, fictional or authorised by law, among other exceptions.

If a company deploys a chatbot, it must make clear to users that it is AI-based; otherwise it becomes a liability and puts its reputation at risk.

Where does it get complicated?

Watermarking is still fragile and produces false positives, and outside the providers themselves nobody can reliably verify that the marks work as intended in every case. Auditing, not the framework, is the real issue (https://arxiv.org/abs/2609.09604).

Finally, remember that penalties apply for failing to meet the transparency obligations:

• For companies in general: administrative fines of up to €15 million or 3% of total worldwide annual turnover for the previous financial year, whichever is higher.
• For SMEs and startups: the regulation provides specific moderation rules, usually favouring the lower of the two amounts to soften the economic impact on small organisations.
• For EU institutions, agencies and bodies: administrative fines are capped at €750,000.
• The EU AI Act is not limited to legal entities (companies and institutions): it can also directly affect individuals, such as freelancers or independent developers.

Next steps

• If you run a chatbot, make sure its first message states that it is an AI.
• If what you publish (news, press releases, reports) was drafted with AI, record who reviews and signs it. This exempts it from labelling.

As a starting point for reviewing your processes, and without constituting legal advice, you can run this prompt about your brand or company:

Act as an EU AI Act compliance consultant (Article 50, transparency). Here is how we use AI in my organisation: [DESCRIBE: customer service chatbot, blog posts, press releases, social media images, etc.]. For each use, give me a table showing: 1) whether any transparency obligation applies and which, 2) whether any exception applies (human review with editorial responsibility, assisted editing, artistic or satirical content), 3) what concrete change we would need to make, and 4) what question a lawyer should review. Be specific and do not invent obligations that are not in Article 50.

The Chief AI Officer's role

A Chief AI Officer takes part in this process as the owner of overall compliance verification, but is not responsible for each published piece. We classify borderline cases, check that the record is complete before publishing, run regular sampling and escalate to legal review when needed.

Editorial responsibility does not lie with the CAIO; a failure of control eventually does. At CAIO International we implement frameworks to secure this control in your company. Contact us to book a meeting: your organisation may be at risk of non-compliance.