From 2 August 2026, new European rules apply to anyone who offers or deploys AI systems that communicate directly with people. The legislation is not new, the EU AI Act entered into force in August 2024, but now the second phase of the rollout begins: concrete obligations for providers of chatbots, generative AI and systems that produce synthetic content.
What applies from 2 August
Article 50 of the AI Act imposes transparency obligations on five categories of AI systems:
- Chatbots and conversational AI, users must know they are communicating with AI before their first message.
- Generative AI (text, image, audio, video), output that could appear "real" must be labelled in a machine-readable way as AI-generated.
- Emotion recognition and biometric categorisation, the people concerned must be informed.
- Deepfakes, AI-manipulated images or audio must be labelled as such.
- Public information, AI that informs the public about matters of public interest also falls under the obligation.
The European Commission has published guidelines and made official EU icons available for labelling AI content and deepfakes, reports NLdigital.
Transitional period: four months
There is an exception for systems that were already on the market before 2 August. They have until 2 December 2026 to comply with the transparency obligations. Newly launched systems must be compliant immediately, writes PONT Data&Privacy.
Enforcement is now active: the European AI Office supervises providers of general-purpose AI models. They must submit technical documentation to supervisors, implement a copyright policy and publish detailed summaries of training data.
What this means for Dutch professionals
If your organisation offers a chatbot or AI assistant to customers or employees, internal or external, you will most likely fall under these obligations. In practice:
- Internal use: an AI assistant for employees that works via a chat interface falls under category 1. Add a clear notice that users are communicating with AI.
- Customer-facing AI: inform customers before the conversation, not afterwards in the small print.
- Generative content: if you publish AI-generated text, images or video, machine-readable metadata is required. The standard APIs from OpenAI, Anthropic and Google do not yet add this metadata automatically, that is the responsibility of the organisation deploying them.
For the GDPR profiles of the models you are considering, the usual criteria still apply: data residency, sub-processors, data processing agreements.
High-risk AI: deadline only in 2027-2028
Systems classified as high-risk, think AI in HR decisions, credit scoring or medical applications, get more time: the obligations for these only apply from 2 December 2027, and for integrated products carrying a CE marking from 2 August 2028.
Arena Elo: GLM rises sharply, DeepSeek falls
Alongside the regulatory developments, there is movement in the Chatbot Arena scores. In the week of 16-22 August, DeepSeek dropped from Elo 1465 to 1459 (−6 points). This runs parallel to the news about the hacked Hugging Face environment in July: users appear to be more cautious about models with known security incidents.
More striking is the rise of GLM (Zhipu AI): from 1471 on 16 August to 1487 on 22 August, a gain of +16 points in seven days. That is the strongest weekly movement in our dataset. GLM-4 has been scoring above average on multilingual tasks for months, but it now also appears to be gaining ground in general comparisons against other Chinese models such as Qwen (stable at 1474) and Kimi (stable at 1461).