Two news items from the past 48 hours that affect you directly as an AI-using professional in the Netherlands: Anthropic has built watermarks into all Claude output, and Meta has released a 30B open-source model that runs on an ordinary laptop or workstation.
Anthropic: watermarks in Claude output, worldwide
From 2 August 2026, Anthropic is adding two types of markers to everything Claude produces:
- Invisible text watermark: a statistical pattern that stays intact when you copy or edit text. It is not readable to people, but detectable by software.
- C2PA metadata: cryptographically signed provenance information in supported file formats (.png, .jpg, .svg). This is lost if you re-save or convert the file.
The trigger is Article 50(2) of the EU AI Act: the transparency requirement for AI-generated content. Notably, Anthropic is applying the marking worldwide, not only in the EU.
What does this mean for you?
The marking is explicitly not proof of authorship. Anthropic acknowledges that Claude is often used for editing, that text changes after processing, and that older models do not carry a marking. So do not use detection tools as a definitive verdict in business or legal contexts.
Organisations that build Claude into their own products, such as legal advocacy tools or HR software, must independently assess whether they comply with Article 50; Anthropic's marking is a tool, not a sufficient compliance measure in itself.
For the complete GDPR and AI Act profile of Claude, see our safety page.
Meta Muse Glimmer: a 30B open model on a single consumer GPU
On 10 August, Meta Superintelligence Labs launched Muse Glimmer, a 30-billion-parameter model under an Apache 2.0 licence. The key lies in the conditions: through 4-bit quantisation, memory use drops from 55 GB to under 20 GB, allowing the model to run on an RTX 5090 (or a comparable card with 24 GB VRAM). Speculative decoding gives a generation speed 3.1× higher than the unoptimised version.
The model weights are on Hugging Face, and the Meta AI Research blog describes how it works. Strengths: autonomous agent tasks, function calling, code, multimodal (text + image), and support for 100+ languages.
Why is this relevant for Dutch professionals?
A powerful model that runs locally has a fundamentally different privacy profile from a cloud model: no prompts or documents go to an external server. That matters in sectors where personal data plays a role, such as healthcare, legal and financial advice, as long as you manage the infrastructure yourself.
Apache 2.0 permits commercial use without restrictions on EU-based entities, unlike the Llama 4 multimodal licence, which previously excluded EU headquarters. Installation does require technical know-how, though ready-made SaaS wrappers will likely follow quickly.
In brief
| News | Date | Action for you |
|---|---|---|
| Claude watermarks (text + C2PA) | 2 Aug 2026 (live) | Assess the Article 50 obligation if you build Claude into your own product |
| Meta Muse Glimmer (30B, Apache 2.0) | 10 Aug 2026 | Test local deployment if you want to avoid cloud processing |
Sources: Anthropic watermarks, explainx.ai · Meta Muse Glimmer announcement · SiliconANGLE: Meta Muse Glimmer · EU AI Act Article 50
