Since 2 August 2026, one of the most practical obligations under the EU AI Act has been in force: the transparency requirement for interactive AI systems and AI-generated content (Article 50). The first enforcement actions by the AI Office of the European Commission have started. This is what you as a Dutch professional or business now need to arrange in practice.
Three obligations that already apply
1. Chatbot disclosure at first contact
Are you deploying a chatbot, virtual assistant or AI telephone service? Then you must actively inform the user at the first point of contact that they are speaking with an AI, not tucked away in the small print. A short notice such as "you are chatting with an AI assistant" at the start of the conversation is sufficient. This follows directly from Article 50 of the AI Act and applies immediately.
2. Deepfake labelling for AI imagery
Realistic AI-generated image, video or audio material must be labelled as such. The definition is broad: alongside manipulated video, it also covers AI-generated images of existing buildings, products or people where they relate to public matters or could cause confusion. Want an AI image of a fictional product in an advert? Then a label is required.
3. Postponed: machine-readable watermarks
The obligation to also make AI-generated content machine-detectable (watermarks, metadata) applies to systems already on the market before 2 August only from 1 December 2026. New systems are subject to this requirement straight away.
What this means for you
For most SMEs, compliance here is an afternoon's work, not a major project: adjust the chatbot disclosure, check AI imagery, and record who is responsible for compliance when new AI tools are introduced. The AI Office's enforcement priority lies with providers (OpenAI, Anthropic, Google), but as a deploying organisation you can also be held accountable.
For non-compliance, the maximum fines are €15 million or 3% of worldwide annual turnover. Those limits apply to deployers; for providers of prohibited AI practices they rise to €35 million or 7% of turnover.
Latest model releases and implications
Alongside the regulation, there are also two recent business news items relevant to the safety assessment of AI tools:
Anthropic (10 September): Anthropic published a misuse report on five cases in which Claude was used for biological weapons research, cyberattacks and large-scale model extraction. The company also reported on a Russia-linked hacking group. The report is the most extensive Anthropic has published to date and makes clear that frontier models are being actively used for state-sponsored misuse.
OpenAI (12 September): OpenAI launched ChatGPT for Financial Services, aimed at market research and financial modelling. The system combines live financial data with GPT-6 Astra. For financial professionals in the EU, this type of tool may be classified as high-risk AI once it is used in decisions on loans or investment advice.
What you can do now
- Check all customer-facing chatbots for an AI disclosure.
- Inventory AI imagery in marketing communications and add labels where needed.
- Record who in your organisation oversees AI compliance; this is a requirement for DPIAs for high-risk applications.
For an overview of the GDPR profile of commonly used AI models, see our safety page. Sectors with the most high-risk AI applications are listed on our sector overview pages.
Sources: ICTRecht, AI Act transparency obligations · PONT Data&Privacy · WerkenmetAI, SME guide · AI Weekly, Anthropic misuse report