Two weeks after the EU AI Act's enforcement began on 2 August, Anthropic has become the first major provider to introduce a concrete technical measure: all future Claude output will carry an invisible text watermark. At the same time, the AI Office is now actively requesting documentation from providers of general-purpose AI (GPAI). Time for a concrete update.
How Anthropic's text watermark works
Anthropic's approach works through what the provider calls "low-stakes choices": whenever the model has to pick between two near-equivalent words or phrasings, it chooses according to a hidden pattern. That pattern is statistically detectable, but not visible to the reader. Key points:
- No loss of quality: the watermark produces no extra tokens and, according to Anthropic, has no effect on the quality of the output.
- Not traceable to the user: the mark identifies neither who nor which organisation generated the text, only that it was Claude.
- Detection API on its way: a separate API that lets you verify whether text came from Claude will be released soon.
- Images and files: images and supported files will receive C2PA content credentials, an industry standard for provenance marking.
For Dutch organisations this is directly relevant: Article 50 of the AI Act obliges providers of AI systems to ensure that AI-generated content is detectable as such (machine-readable marks). Anthropic is on schedule with this. The question is what will reach the market from providers that have not yet taken this step.
Source: Anthropic release notes, August 2026
GPAI enforcement: what the AI Office is now requesting
Alongside the transparency obligation for end users (chatbots must make clear they are AI, deepfakes must be labelled) there is a less discussed obligation that directly affects providers such as OpenAI, Google, Anthropic and Mistral: the GPAI documentation requirement.
From 2 August 2026, the AI Office can request from every provider of a general-purpose AI model:
- Technical documentation of the model
- A summary of the training data, detailed enough to allow copyright claims to be assessed
- Evidence of an active copyright policy
Providers of models with systemic risk, that is, models trained with more than 10^25 FLOP, face additional requirements on cybersecurity and incident reporting on top of this. GPT-5.6, Claude Opus 5 and Gemini 3.7 Ultra likely fall into this category. 180 organisations have already signed the Code of Practice for GPAI transparency.
More than 180 organisations have signed the voluntary Code of Practice for GPAI transparency. Enforcement can lead to fines of up to €15 million or 3% of global annual turnover.
What this means if you use AI today
If you use AI as a professional or in an organisation in the Netherlands, there are three concrete actions:
1. Check your provider. Has the provider of your tool published a GPAI summary or signed the Code of Practice? Major providers such as Anthropic and Google now do; for niche providers this is far from always the case. On our models page you can see which providers offer a data processing agreement and privacy documentation.
2. Look at your output processes. If your organisation publishes AI-generated text, such as newsletters, reports or social media posts, you are obliged under Article 50 to label it as such where the text is informative in nature and aimed at the public. This also applies to text that Anthropic has watermarked: the watermark is no replacement for the labelling obligation, but a supplement to it.
3. Take stock of which tools process which data. The GPAI training data obligation also has downstream effects: if you supply data to a model as a processor, you share responsibility for the legal basis. See our safety page for a GDPR profile per provider, and the sector pages for industry-specific risks (e.g. legal, healthcare, financial).
Over the coming weeks, more providers will release watermarking and detection APIs, or come under pressure from regulators to do so. The technical bar for compliance is now becoming concretely visible.
Sources: EU AI Act enforcement August 2026 (AI Office) · Anthropic release notes · AI Act article 50 explained · Dutch regulatory framework (RDI/AP)