Yesterday, 2 August 2026, the transparency obligation of the EU AI Act came into force. That sounds big, and for a small share of organisations, it is. But for most Dutch professionals the immediate impact is limited, as long as you know what actually applies now and what will take years to arrive.
What is mandatory from 2 August
The obligation stems from Article 50 of the AI Act and applies to every organisation that provides or deploys AI systems for EU users, regardless of company size or country of establishment.
Chatbots and voice assistants must disclose that they are AI at the start of a conversation, unless this is already evident. A short opening sentence is sufficient in most cases.
AI-generated media, audio, images, video, must be marked as artificially generated, including machine-readable metadata where technically feasible. This also applies to deepfakes, except for clearly creative or satirical material.
Emotion recognition and biometric categorisation: if you use these, you must inform the people concerned.
AI-generated text on matters of public interest must be labelled as AI-generated, unless a human takes editorial responsibility for it.
Fines for non-compliance can reach €15 million or 3% of worldwide annual turnover (whichever is higher). Lower maximums apply to SMEs.
What does NOT take effect in August
This is at least as important: the requirements for high-risk AI, medical devices, recruitment systems, credit assessors, have been postponed via the Digital Omnibus to December 2027 and August 2028. The major compliance task therefore still has time.
Practical implications for your organisation
Do you use a chatbot on your website? Add a sentence such as "You are chatting with an AI assistant." Do you create marketing images with tools like Midjourney or Flux? Label them as AI-generated. At present, nothing more is required of the average knowledge worker or SME.
Ten Dutch supervisory authorities divide AI oversight by sector, from the ACM for consumer markets to the AP for personal data. Earlier this year the AP already published a guide on generative AI and the GDPR for anyone wishing to test their systems.
The role of your AI model
Which model you choose directly affects how easily you can comply with these rules. Models that run through European data centres, such as Mistral on its own Paris infrastructure, simplify the data flow. Models where data is processed outside the EU require additional contractual safeguards (SCCs or an adequacy decision). On our safety page you can find how this is arranged for each model.
OpenAI Astra solves open mathematical problems
Separately from the AI Act: yesterday OpenAI also announced that an internal version of its next model, Astra, solved ten open mathematical and theoretical computer science problems with formal Lean proofs, published on GitHub. Mathematician Timothy Gowers (Fields Medal) called one of the proofs worthy of publication in a top journal. Whether Astra heralds a new era or was a well-chosen demonstration remains uncertain, but it indicates that the reasoning boundary of language models is shifting quickly.
For model comparisons and benchmarks: see the model overview and the sector-specific tools per field.
Sources: Score Agency – EU AI Act obligations 2026 · WerkenmetAI – transparency obligation for SMEs · CFO.nl – new phase of the AI Act · BuildFastWithAI – AI news 2 Aug 2026