Sean Walsh
Sean is a Director at Intelligency heading up our digital marketing and client services operations. Sean has 15+ years experiencing working both in-house and agency with brands including Lloyds, Alstom, Hitachi, Lufthansa, Viaplay, DFDS Seaways and Mercedes-Benz.
Anthropic announced this week that it is adding invisible, machine-readable watermarks to text generated by new Claude models. The rollout is global, covering Claude, the Claude API, Claude Code, Claude Cowork and Claude Tag. The trigger is the European Union’s AI Act transparency code, but Anthropic has confirmed it is not limiting the feature to Europe – it is being enabled everywhere Claude is available.
For anyone using Claude as part of their content workflow, whether to draft, edit, refine or produce marketing material, this is worth understanding. It does not change what the tool does or how it reads. But it changes how that content can be identified, and that has implications that are already beginning to ripple through discussions about disclosure, authenticity and AI detection.
What the watermark actually is
The watermark Anthropic is embedding is not a visible label or a footer. It is an imperceptible, machine-readable signal woven into the text itself. Anthropic says it does not change the meaning, quality or readability of the output. To the person reading it, nothing looks different. The mark is invisible to the human eye.
What the watermark does is allow a detection tool to confirm, with high probability, whether a given piece of text was generated by a supported Claude model. Think of it as a fingerprint in the language patterns rather than a stamp on the page.
Anthropic is also working to apply watermarking retrospectively to Claude models released before 2 August 2026, during the AI Act’s transition period. So this is not just about new models going forward – it is being applied more broadly over time.
What it does not do
Two caveats Anthropic has been explicit about are worth flagging because they change the practical implications considerably.
- It does not prove authorship of ideas. The watermark confirms that a Claude model generated the text. It does not mean Claude originated the ideas, research or creative direction. A human could have provided all of that and used Claude only to write up the output. The watermark captures the generation tool, not the intellectual source.
- Absence of a watermark is not proof of human authorship. Older models, heavily edited text and files with stripped metadata may not carry a detectable watermark. Someone could also remove or corrupt it deliberately. The absence of a watermark does not confirm that content was written by a person.
These two points matter because a lot of the commercial and legal conversation around AI content disclosure assumes a binary: AI-generated or human-generated. The reality is more layered. Most professional use of AI tools involves significant human input, direction and editing. The watermark detects the generation surface, not the level of human involvement.
Why this is happening now
The immediate driver is the EU AI Act, which came into force in stages through 2025 and 2026. The transparency code it establishes requires AI providers to make it technically possible to identify AI-generated content. Anthropic’s watermarking commitment is a direct fulfilment of that requirement for systems deployed in the EU.
The decision to roll it out globally rather than just in the EU is significant. It reflects a broader industry direction that has been building for several years. In 2023 major AI companies including OpenAI, Google, Meta and Amazon made voluntary commitments to develop watermarking and provenance technologies. Google ran a large-scale trial of its SynthID watermarking technology across 20 million Gemini responses in 2024. Research into text watermarking has been underway since the 1990s, but practical large-scale deployment has only become viable with the current generation of AI systems.
The timing also reflects growing pressure from publishers, regulators and educators who want tools to identify AI-generated content more reliably. AI detection tools have faced significant criticism for accuracy problems, including false positives that have wrongly flagged human-written text. Watermarking embedded at the generation stage is a more structurally reliable approach, though it still has the limitations Anthropic has acknowledged.
What this means if you use Claude for content
Claude is widely used across marketing, content production and communications. AI tools sit at the centre of most agency and in-house content workflows now, and Claude specifically is used for drafting, editing, tone adjustment, research summaries and strategy documents across the industry. The watermark does not change any of that. It does change the transparency landscape around it.
The practical implications depend on your context:
- For client-facing content. If a client, publication or platform uses a watermark detection tool and your content triggers it, you will want to have a clear position on how you use AI in your workflow. This is not a new conversation – disclosure expectations have been developing across industries for two years – but watermarking makes the question technically answerable in a way that was not previously possible at scale.
- For SEO and publishing. Google has been consistent that AI-generated content is not inherently penalised – quality and helpfulness are what matter. Watermarking does not change that position. But it does give Google, publishers and platforms more technical capability to understand the scale of AI content in any given context. How that capability is used over time is an open question.
- For edited and hybrid content. Anthropic explicitly notes that heavily edited text may not carry a detectable watermark. If your workflow involves significant human editing of AI drafts, the practical footprint of the watermark in your final output may be limited. This is consistent with the direction most good AI-assisted content workflows have been heading anyway – AI as a drafting and structuring tool, human judgment as the final layer.
The broader direction this points to
Anthropic’s move is the most significant practical step yet toward AI content provenance becoming a standard layer of the web. The question of how brands build recognition and trust online is increasingly bound up with questions of authenticity and transparency. Watermarking is one part of that picture.
The technology is not yet universal – other major AI providers are at different stages of deployment, and there is no single interoperable standard. But the direction is clear. Over the next two to three years, it is likely that most major AI-generated text will carry some form of machine-readable provenance signal, and that platforms, regulators and clients will increasingly have tools to detect it.
The most straightforward response for any business using AI in its content workflow is not to panic, and not to try to strip or circumvent watermarks – that path leads nowhere good. It is to be clear internally about how AI is used, at what stages and with how much human oversight, and to develop a consistent, honest position on disclosure that you would be comfortable defending if asked. That position should already exist. If it does not, the arrival of invisible watermarks in every Claude response is a reasonable prompt to develop one.
