In twelve days, on 2 August 2026, the transparency obligations of the EU AI Act become enforceable. The deadline has come up several times in our earlier posts, but now that the countdown is really running out, it is time to turn the abstract legal text into concrete action. This article explains what exactly changes, for whom, and how recent model developments are responding to it.
What Article 50 requires
Article 50 of the AI Act sets out four obligations that apply from 2 August:
- Chatbot disclosure: Any AI system that communicates directly with people must be designed so that the user knows they are interacting with an AI. This is not required where it is already obvious, but in professional use (customer service, HR portals, legal assistants) that is rarely the case.
- Machine-readable marking of AI content: Audio, images, video and text created by generative AI must be marked in machine-readable form as artificially generated. For systems already on the market before 2 August, a grace period runs until 2 December 2026, but new deployments fall under the rule immediately.
- Information provision for emotion recognition and biometric categorisation: Users must be informed when AI analyses their emotions or personal characteristics. This affects HR tools, recruitment software and some marketing platforms.
- Labelling of deepfakes and AI-generated public-interest text: Manipulated images and audio must be clearly identifiable as artificial; AI-generated text published to inform the public must carry a disclosure, unless it is under editorial responsibility.
Supervision in the Netherlands rests with the Dutch Data Protection Authority and the Rijksinspectie Digitale Infrastructuur. The maximum fine for non-compliance is 15 million euro or 3% of worldwide annual turnover.
Source: AI Act Blog – Article 50 in practice · EU AI Act – Article 50 text
Claude Sonnet 5: the most popular model for agents, now with higher stakes
Timing matters: Claude Sonnet 5, which Anthropic launched on 30 June 2026, is now the default model for agentic tasks on the Free and Pro plans. With a SWE-bench score of 72.7% and a Terminal-bench score of 76.1%, a jump of 20 points on its predecessor, many organisations are deploying Sonnet 5 in automated workflows that deal directly with customers or internal staff.
That makes the Article 50 deadline especially relevant: a Sonnet 5 agent that answers customer queries, summarises contracts or processes HR requests falls directly under the disclosure obligation. It is not the model that determines whether you are compliant, but the way you deploy it.
Introductory pricing until 31 August: $2/M input, $10/M output. Afterwards $3/$15. You can find more on the model specifications on our models page.
Source: Anthropic – Claude Sonnet 5 · TechCrunch – Anthropic launches Claude Sonnet 5
Three actions for the next twelve days
1. Inventory which systems fall under Article 50.
Establish which AI applications in your organisation communicate directly with people (chatbots, voice assistants, automated email handling) and which systems generate AI content for external publication.
2. Determine your role: provider or deployer?
If you purchase an off-the-shelf tool (for example an AI customer service platform), you are a deployer and responsible for the disclosure layer. If you integrate a model yourself via API, you are also a provider for that specific system. Both roles carry obligations.
3. Check the DPAs of your AI vendors.
The disclosure obligation is technically straightforward; labelling AI content is more complicated and requires cooperation with the provider of the model or platform. Check whether your vendor already supports watermarking or C2PA metadata. On our safety page you will find an overview of the privacy documentation for each model.
The deadline of 2 August is not a symbolic date, it is when national supervisory authorities become empowered to enforce. For most organisations, Article 50 compliance is achievable in the time remaining, but delay increases the risk of a flying start with a fine in sight.