Two stories from the past day that matter to you as a Dutch professional. OpenAI will invisibly watermark text from ChatGPT in the EU, because the AI Act requires it. And two of the biggest users of AI coding assistants are steering their staff away from Claude, mainly on cost. What does that mean for you?
Invisible watermark in ChatGPT and Codex text
Over the coming weeks, OpenAI will add an invisible watermark to text that ChatGPT and Codex produce for users in the EU, on all subscriptions. That is according to TechCrunch. The technique, textGrain, subtly steers the model's word choice. You won't notice anything as a reader, but a detector can recognise the pattern. The trigger is the transparency rules of the AI Act, which have applied since 2 August 2026.
A few details that matter:
- According to OpenAI, the watermark does not identify the user. It only shows that an OpenAI system was involved in the text, not how much of it you wrote yourself.
- Via the API you can switch it on worldwide for a number of models. It is off by default.
- For now, the detector is only available to approved researchers and expert organisations.
- Detection is not foolproof. Anyone who replaces 10% of the words with synonyms lowers detection in OpenAI's own test from 92% to 66%. Short texts, arithmetic answers and translations are harder to recognise. And the absence of a watermark does not prove that a human wrote the text.
If you work in education, that last point matters: a "clean" result is not proof, and you cannot use the detector yourself yet. If you work in media or marketing, bear in mind that texts you take unchanged from ChatGPT are in principle recognisable as AI text. Your GDPR assessment does not change: the watermark says nothing about who you are or what happens to your input. You can read that per model at /veiligheid/ and at /chatgpt/.
Meta and Microsoft steer staff away from Claude
According to The Information, summarised by PYMNTS, Meta and Microsoft are restricting internal use of Claude. At Meta, the number of internal Claude Code users fell from about 60,000 to 30,000, in favour of in-house tools such as Muse Code and MetaCode. Microsoft was on course to spend about $1 billion a year on Anthropic internally and has cut that by a third. Developers are being directed to GitHub Copilot, although engineers may still choose other models.
The article also cites a Wall Street Journal study: only 11% of nearly 400 companies correctly predicted their AI spending. Usage-based pricing makes the bill hard to plan.
For a Dutch development team, the lesson is not that Claude is worse, but that cost and vendor dependency matter. Set a budget limit on your API keys, review your usage monthly, and choose tools where you can switch models. At /claude/ and /copilot/ you can see how both score and what they cost. The coding tools we have checked are at /voor/softwareontwikkeling/. To compare models side by side, go to /modellen/.
In brief
- ChatGPT and Codex text will receive an invisible watermark in the EU over the coming weeks. It does not identify you, and detection is not foolproof.
- A text without a watermark is no proof that a human wrote it.
- Meta and Microsoft are curbing internal use of Claude, mainly because of cost. Keep an eye on your own usage and don't tie yourself to a single vendor.
