Claude’s invisible watermark raises a very visible question: Who owns AI-generated work? 

Artificial intelligence was supposed to make writing, coding and research easier. Now, at least according to some Claude users, it may also make their work permanently identifiable. 

Anthropic, the company behind Claude, has introduced machine-readable watermarking for content generated by newer Claude models. The invisible markers are designed to survive copying and some editing, allowing systems to detect that content originated with Claude. The move comes as new transparency obligations under the European Union’s AI Act take effect from August 2, requiring providers of generative AI systems to make AI-generated content machine-detectable.  

On paper, this sounds eminently reasonable. In a world increasingly flooded with synthetic text, images, audio and video, knowing where something came from is useful. The EU itself says the rules are intended to combat deception and misinformation and protect the integrity of the information ecosystem.  

But the implementation has exposed a much messier question: what exactly counts as AI-generated work? 

Who owns AI output? 

That question is at the heart of a heated discussion on Reddit’s Claude community. One post, titled “Claude watermarking our work is unethical and disgusting”, argues that Claude is a tool rather than the author of the final product. The user points out that prompts, decisions, corrections, context and countless refinements may all come from a human, with Claude merely helping execute the process.  

The concern becomes particularly interesting when Claude is used for editing rather than creation. A writer might compose an article and ask Claude to tighten the prose. A journalist could use it to reorganise a transcript. A programmer might ask Claude to refactor existing code. A student could use it to improve grammar. In each case, the human contribution may be substantial, even dominant. Yet if the resulting text carries a machine-readable signature, a future detection system could potentially reduce that complicated collaboration to a simple conclusion: AI-generated. 

That is where watermarking risks becoming less a transparency mechanism and more a label of provenance. But provenance and authorship are not the same thing. A photograph edited in Photoshop is not necessarily “made by Photoshop”. A spreadsheet calculated in Excel does not belong to Microsoft. AI complicates the analogy because generative systems can produce substantial portions of the underlying work. But that does not mean every interaction with an AI system should be treated as equivalent to outsourcing the entire creative process. 

The Reddit debate captures this tension neatly. Some users see watermarking as sensible infrastructure for a world where distinguishing human from machine output is increasingly difficult. Others worry that it could stigmatise legitimate AI assistance and make people reluctant to use tools that are becoming as ordinary as spellcheckers and search engines.  

A need for nuance 

There is also an uncomfortable irony here. AI companies are being asked to prove the provenance of their outputs even as those systems were built by training on enormous quantities of human-created material. Users therefore understandably ask whether the relationship should work both ways: if AI companies want attribution and traceability for their outputs, how much transparency should they provide about the material and processes that shaped the models themselves? 

None of this means watermarking is inherently bad. In fact, one could go so far as to say it is essential. Deepfakes, automated propaganda and industrial-scale synthetic content create genuine problems, and invisible provenance systems could eventually help platforms and users navigate an internet where authenticity is increasingly difficult to establish. 

But the technology needs nuance. The real danger is not that Claude can say, “I helped make this.” It is that a machine-readable mark could eventually say, “A machine made this,” when the truth is considerably more complicated. 

The future of AI transparency should therefore be about disclosure, not digital guilt. Users should know when AI has materially contributed to something. But there should also be room to distinguish between generation, collaboration, editing, translation, proofreading and human authorship, and understanding where the human ends and th machine takes over the wheel.