By Hezron Ochiel
AI-generated writing may soon become easier to trace.
Anthropic has announced plans to add imperceptible watermarks to text generated by Claude. The move comes as the European Union’s AI Act transparency requirements for AI-generated content begin to apply.
The watermark will not appear as a visible label or hidden character. Instead, it is designed to create a detectable statistical pattern in the generated text.
For writers, employers and publishers, this raises an important question: What does AI involvement actually mean?
A watermark can show AI involvement
A watermark can indicate that Claude likely helped produce some of the text.
That can still represent very different levels of human contribution.
Someone may write an article and ask Claude to improve the wording, while another person may ask Claude to generate most of the piece. In both cases, AI is involved, but the level of human thinking, judgement, and writing can differ greatly.
A detection result may therefore show that Claude played a role without explaining who developed the idea, checked the facts or took responsibility for the final work.
This is becoming more important as platforms such as LinkedIn also try to separate useful AI-assisted content from low-value material that can weaken reach and trust.
How Claude’s watermark works
Anthropic has not publicly disclosed every technical detail of the watermarking system.
The approach appears to work by influencing some of the model’s token or word choices in ways that create a statistical pattern while keeping the text natural to read.
That pattern can then be checked using the appropriate detection system.
Nothing needs to be visibly added for the signal to exist.
What happens when Claude only proofreads?
If you write something yourself and ask Claude to correct grammar or punctuation, most of your wording may remain unchanged.
That can leave fewer Claude-generated choices from which a detectable pattern can emerge.
This distinction also appears in Article 50 of the EU AI Act, which says the machine-readable marking obligation does not apply where an AI system performs a standard editing function or does not substantially alter the original input or its meaning.
The level of AI involvement therefore matters.
Fixing a typo, improving a paragraph and generating an entire article involve very different levels of contribution.
A watermark may indicate that AI was used without showing how much of the thinking or writing came from the person.
Where detection may struggle
Watermarking has limits.
Very short passages can provide fewer signals to analyse, while heavy editing may weaken the pattern.
The absence of a detectable watermark should therefore not automatically be treated as proof that a piece of text was written entirely by a human.
It is better understood as one signal about how the content may have been produced.
Why Anthropic is introducing it
The change comes as the European Union’s AI Act transparency requirements for AI-generated content take effect.
Article 50 of the EU AI Act requires providers of generative AI systems to make AI-generated text machine-readable and detectable as artificially generated or manipulated where technically feasible.
The transparency obligations started applying on August 2, 2026, although some existing systems have a transition period for the marking requirement.
The European Commission has also published guidelines on transparency obligations for AI systems and a Code of Practice on Transparency of AI-generated Content to help providers and deployers comply with the rules.
Anthropic’s watermarking plans therefore sit within a wider move towards making AI-generated content easier to identify.
This also adds another layer to the discussion around AI search, visibility and trust, especially as people try to understand both where information comes from and how it is produced.
What this could mean for workplace AI policies
Many organisations still frame AI policies around one question:
Was AI used?
That question may not tell us enough.
Using AI to check grammar, improve clarity or generate a complete document can involve very different levels of human contribution.
Employers, universities, publishers and communication teams may therefore need clearer rules around acceptable AI assistance, disclosure and responsibility.
A more useful question may be: How was AI used, and who is accountable for the final work?
This will matter more as AI becomes part of everyday workplace writing.
For communicators, this is about trust
Communication depends on people trusting what is published and the people or institutions behind it.
A watermark may show that AI was involved, but it cannot tell us who exercised judgement, checked the facts or accepted responsibility for the final work.
Responsibility still rests with the person or organisation publishing the content.
That is also why trust deserves a place alongside visibility and AI citability when PR teams measure communication performance.
AI policies should follow the same principle by making it clear when human review, disclosure and accountability are required.
What this means for content creators
As watermarking develops, the discussion around AI-generated content may increasingly focus on how AI was used and who remained responsible for the final work.
Knowing that AI was involved does not tell us whether the work is accurate, useful or well judged.
A practical standard is simple: Use AI in ways you would be comfortable explaining to your employer, client, editor or audience.
AI can support the work while the person publishing it remains responsible for the accuracy, judgement and trust attached to it.
Hezron Ochiel is an award-winning strategic communications and public relations professional with over 15 years of experience in media, digital communication and reputation strategy. He serves as the Deputy Corporate Communications Manager at the government-owned Kenya Medical Training College (KMTC) and is the founder of Hezron Insights, where he writes about AI visibility, Digital PR, SEO, GEO and digital authority. His work has appeared on platforms including Reuters, The New Humanitarian and The Standard.