Anthropic is embedding an imperceptible watermark into Claude-generated text, and it’s a bigger deal than it might sound. In this episode, I break down how the watermarking system actually works, what it can and cannot prove, and why the EU AI Act is driving this change globally — not just in Europe.
What Is Anthropic’s Text Watermarking System?
This isn’t a visible label, a disclaimer at the end of a response, or hidden Unicode text. Anthropic’s watermark works at the model level, using a version of SynthID-Text — originally developed by Google DeepMind. As Claude generates a response, it makes slightly different but still valid choices between possible words or tokens. Across enough text, those choices create a statistical pattern that can be detected with the correct method or key.
It applies across Claude’s consumer app, API, Claude Code, Claude Cowork, Claude Tag, and supported deployments via AWS, Google Cloud, and Microsoft. Models launched on or after 2 August 2026 support marking from launch, and Anthropic is working to add it to earlier models too.
What the Watermark Actually Proves
Here’s where it gets really important for those of us working in public sector comms. A detected Claude watermark is evidence that Claude processed the text — it is not proof that Claude wrote every word. The mark can remain when Claude has translated, summarised, proofread, or lightly edited human-written material.
So please don’t treat it as a plagiarism verdict or conclusive proof of authorship. It’s a provenance signal — useful evidence, but not a complete picture of how a document was created.
Key Limitations You Need to Know
- It’s probabilistic. Detection indicates the likelihood Claude was involved, not a definitive yes or no.
- Short text is harder to assess. Statistical signals are more reliable over longer passages — a short email or social post offers weaker evidence than a long article.
- Editing can weaken the signal. Rewriting, paraphrasing, translating, or combining outputs can reduce detectability.
- It identifies a process, not a person. The watermark contains no information identifying a particular user, organisation, or conversation.
- It’s not Claude-exclusive. It provides a broader AI-provenance signal, not confirmation that Claude specifically produced the text.
Why the EU AI Act Matters Here
This story is fundamentally a European regulatory story with global consequences. Anthropic says it’s implementing watermarking to comply with the EU AI Act — and crucially, it’s deploying it globally, not just for European users. That’s exactly how regulation shapes product design worldwide, even for organisations outside the EU.
What This Means for Public Sector Communicators
My advice is straightforward: build a clear AI-use policy, disclose meaningful AI assistance where it matters, retain drafts and source material for high-stakes work, and treat detection tools as one piece of evidence, not a final judgment. Anthropic’s watermark is not a lie detector for AI writing — it’s a provenance signal that can make the web more transparent, but only if we understand both what it reveals and what it cannot prove.
If you want to build your AI literacy and skills, I can help
Grab AI Playbook on Amazon, the perfect podcast companion.
Subscribe to Human in the Loop, my LinkedIn Newsletter
Take the CPD Accredited AI Playbook Diploma
Book a call at digitaltraining.ie
Sources
Anthropic: How Anthropic’s Watermarking Works. https://www.anthropic.com/news/claude-text-watermark
TechCrunch: Anthropic says it will watermark text generated by AI models. https://techcrunch.com/2026/08/11/anthropic-says-it-will-watermark-text-generated-by-its-ai-models/
Key Statistics
– Models launched on or after 2 August 2026 support watermarking from launch.
– Watermarking is deployed globally, not solely for EU users, to comply with the EU AI Act.
– The system covers Claude’s consumer app, API, Claude Code, Claude Cowork, Claude Tag, AWS, Google Cloud, and Microsoft deployments.
– The watermark uses a version of SynthID-Text, originally developed by Google DeepMind.
About the Podcast Host
AI and Communications Specialist | Author | Podcaster | International Speaker
Joanne Sweeney is a recognised leader in digital marketing and digital transformation, specialising in the practical application of AI to enhance communication strategies. The founder of Digital Training Institute, Joanne, empowers leaders and communications professionals to leverage AI responsibly and effectively, ensuring transparency, trust, and tangible results.
She completed the Harvard Kennedy School Executive Programme in AI and Public Policy this summer before launching her fourth book AI Playbook, in September. Joanne has also published two editions of Public Sector Marketing Pro, which were preceded by her first title, Social Media Under Investigation Law Enforcement and the Social Web.
Her books have become go-to resources for governments, public sector and policing agencies, as well as marketing and communications professionals, all navigating the challenges of the changing Internet.
Joanne is the host of the AI SIX Podcast, a daily six-minute podcast that provides actionable insights, expert interviews, and practical advice to help professionals navigate the new world of work.
