On the 2nd of August 2026, Article 50 of the EU AI Act came into force, setting out mandatory transparency requirements for certain AI systems.
This includes letting users know when they interact directly with AI, applying machine-readable watermarks to synthetic text, images, audio, and video, disclosing deepfakes, and labelling things considered ‘public-interest AI text’.
In response, frontier AI firm Anthropic put a hidden watermark in all of the written content generated by its Claude LLM models.
This news has been met with mixed responses from the public – some deeming the move as significant for AI governance, while others suggesting it risks painting all AI-assisted writing with the same brush – regardless of how much, or how little, human input actually went into it.
Content Created Equally
An interesting byproduct of Claude’s regulatory move is that all watermarked content will likely be judged the same – even though it perhaps shouldn’t.
On one end, there’s the complete non-writer, endlessly prompting until something usable emerges. On the other, a skilled writer running finished copy through Claude for a light edit or a bit of restructuring. Both get the same watermark. However, neither gets credit for the very different work behind it.
Dr Alan Finkel, Australia’s former Chief Scientist and co-founder of Proudly Human – a human-created content certification – sees this as a larger issue around authorship and trust: “Knowing that the author or artist is human and imagining what motivated their work is part of the pleasure for the reader, viewer or user. Currently, AI-generated content masquerading as human-created content is denying our right to choose human-composed music, human-authored books or human-created art.”
A Shift for LLM Usage
This shift might point to something bigger – GenAI tools being used in an entirely new way. Internet content pre-GenAI was grounded in an awareness of content ownership online, and so internet users generally had a recognition and mindfulness of plagiarism when using online content to inform their own work. Not copying and pasting content directly, but instead putting already existing ideas into their own words.
Claude’s watermark could usher in an era where users are deterred from a simple ‘copy and paste’ to writing original content with the help of AI models like Claude as tool.
Finkel welcomes Anthropic’s move, but with caveats: “It is an important initiative in a flood of AI generated content. There will be unethical companies that will produce programmes to strip the watermark, and there will be AI text generators that do not participate.The result will be that Anthropic’s initiative will be important and will help but there will still be gaps.”
Where this Leaves Writers
Article 50 was never designed to help distinguish between AI being used for a first draft or a light sub-edit – it only asks for disclosure. In practice, that blanket approach is already reshaping how content is perceived, and perhaps how it is made too.
Watermarking won’t stop AI assisted writing, and as Finkel notes, it won’t catch everything either. What it may do is push writers to think more carefully about how they use tools like Claude – not as a shortcut to finished copy but as one part of a process that still has a human hand in it.
