The watermark doesn’t know who wrote it

There has been a lot of posts and discussions recently on the introduction of watermarks by Anthropic.

It was done to comply with Article 50 of the (EU) 2024/1689, the EU Artificial Intelligence Act, otherwise know as “Transparency obligations for providers and deployers of certain AI systems”

The summary of this act, as provided by the European Commission is the following:

Providers must inform users when they are interacting directly with an AI system. AI-generated or -manipulated content must be clearly marked and detectable as artificially generated. Deployers of emotion recognition or biometric categorisation systems must inform exposed person about the operation of the system. Deepfakes and AI-generated text must be disclosed as artificially generated. An exemption from these transparency obligations applies in case the systems are authorised by law to detect, prevent, investigate or prosecute criminal offences. Other specific exceptions apply per transparency obligation. Information must be provided clearly, distinguishably and accessibly.

To satisfy its duty as an AI Provider under Article 50(2), Anthropic has chosen to weave statistical watermarks directly into the text at the model level.

AI practitioners have complained in Linkedin, Reddit, and co. that their content edited with Claude, or ideas written with the help of Claude, will be “watermarked” as generated with Claude. And yes, it appears to be the case according to Anthropic explanations.

What is the issue? This technical implementation from Anthropic completely ignores how Article 50 treats the human-in-the-loop.

While Providers are required to mark AI content, Article 50(4) explicitly carves out an exception for Deployers (the users publishing the text) when editorial control is involved:

Deployers of an AI system that generates or manipulates text which is published with the purpose of informing the public on matters of public interest shall disclose that the text has been artificially generated or manipulated.

This obligation shall not apply where [...] the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.

By embedding an indelible watermark directly into the text generated during brainstorming, proofreading, or editing, Anthropic’s solution applies a blanket approach.

Whether a draft is 100% automated “AI slop” or a human-written piece that simply used Claude for minor feedback, the watermark travels with the text.

Providers like Anthropic need to engineer solutions that respect the human-in-the-loop, rather than flattening every interaction into a binary “AI-generated” tag. Or face a backlash.