You wrote the text. Claude got the credit.
Ask Claude to shorten a paragraph you wrote yourself, and the result gets marked as AI-generated content. Not because Claude wrote it — but because Claude touched it, and according to Anthropic, "Claude may not be the original author."
This isn't a bug. It's a deliberate consequence of Anthropic signing the EU's Code of Practice on Transparency of AI-Generated Content this summer, which operationalizes Article 50(2) of the EU AI Act. Claude now embeds an imperceptible watermark in all generated text, plus signed C2PA metadata on supported file types (SVG, PNG, JPG) — across every Claude product, from Claude.ai to the API to Claude Code.
The question worth asking: if everything that's touched AI gets marked the same way — regardless of how much of the content is actually yours — what does that do to reader trust in what's marked?
The law is actually more nuanced than Claude's own policy
Here's the interesting part: EU AI Act Article 50(2) has a built-in exemption for exactly this scenario. The marking requirement doesn't apply where the AI system "performs an assistive function for standard editing" or "does not substantially alter the input data." In other words: the law itself distinguishes between generating content and editing it.
There's a second, narrower exemption in Article 50(4): AI-generated text published on matters of public interest is exempt from the marking obligation if a natural or legal person has taken genuine editorial responsibility for it — not just run a spell-check.
Anthropic chooses not to lean on either exemption for its own product. That's a defensible, cautious call — Claude has no way of knowing what happens to the text after it leaves the conversation. But it means Claude goes further than the law actually requires, and that's worth knowing when you're deciding what an "AI-generated" mark actually tells you.
The other major labs handle this differently
This isn't an industry standard everyone agrees on. Quite the opposite:
OpenAI has deliberately held back. They've had a text watermarking solution ready for years — around 99.9% accurate — but never shipped it. The public reasoning has two parts: internal surveys found roughly a third of users said they'd leave ChatGPT if the feature launched, and — more seriously — a peer-reviewed study found AI detectors falsely flagged 61% of TOEFL essays written by non-native English speakers as AI-generated. A tool meant to build trust ended up systematically casting suspicion on people writing in their second language.
Google watermarks by default, via SynthID across text, image, audio, and video in the Gemini family — over 10 billion marked items by May 2026. But independent reporting has found SynthID's own detector giving inconsistent results, with both false positives and false negatives within the same conversation.
None of the three major labs has landed on the same answer. That says something about how unsettled this actually is — not just technically, but about what "AI-generated" is even supposed to mean.
What's worth watching
Here's something we think is worth saying out loud — not just about this one requirement, but about how you should think about new technology generally going forward: the rules are still moving, and they're moving differently in the EU than in the US and China.
Article 50's marking obligation was originally supposed to take effect on 2 August 2026. It's already been delayed to 2 December 2026 — part of the EU's "Digital Omnibus" package, passed this summer to simplify and delay parts of the AI Act. That's not a coincidence. In July 2025, over 45 major companies — including Airbus, ASML, Mistral AI, Mercedes-Benz, and Lufthansa — signed an open letter asking for a two-year pause on implementation. Apple delayed AI feature rollouts in the EU citing regulatory uncertainty.
That doesn't mean the regulation is wrong or unnecessary — there's just as much documented pushback against weakening it (120+ civil society organizations have criticized the Digital Omnibus package as a genuine rollback of digital rights, not mere simplification). And the Norwegian picture isn't uniformly critical either: Digitalisation Minister Karianne Tung has framed AI Act adaptation as a competitiveness necessity, not a burden, and DNB has publicly welcomed harmonized rules as something that provides predictability.
But the point we think is worth taking away: when requirements shift mid-course — and they already have, for this exact provision — it means businesses making serious bets on transformative technology in the EU/EEA need to factor compliance uncertainty into the decision, not just the cost of the technology itself. If the requirements end up substantially heavier here than in markets you're competing against, that's a real competitive disadvantage, not a hypothetical one. The risk isn't just that the rules are strict. It's that many businesses hold back from adopting technology that could genuinely give them an edge, simply because they don't know what the rules will look like in six months.
What this means for you
You don't need a settled answer to the debate over whether the EU is regulating too much or too little. You need a structure that can handle the answer changing.
That's exactly why we recommend an AI committee with department-level owners, as we've written about before — not because it solves the regulatory uncertainty, but because it gives you someone who's actually tracking it, and who can adjust policy quickly when the requirements shift. The same logic applies to vendor choice: knowing that OpenAI, Google, and Anthropic have landed on three different answers to the same question is a reminder that no single vendor has "solved" this — and that's a reason not to make yourself dependent on just one.
AI doesn't start with technology. It starts with structure.
Get in touch for a free AI Ready assessment →
Read also: Who's Responsible for AI in Your Business? Here's Why You Need an AI Committee · What Does the AI Act Mean for Norwegian Businesses?