By 
Anthropic has begun embedding invisible, machine-readable watermarks into text generated by its newer Claude models, alongside signed provenance metadata for supported file types such as PNG, JPG and SVG images. The company confirmed the rollout in an updated Claude Help Center article published this week, and multiple international outlets, including Euronews and Business Standard, reported on the change after it took effect.
The marking system applies to any Claude model launched on or after August 2, 2026, and it is not restricted to Europe. Anthropic has said the watermarking works everywhere Claude is used, covering Claude.ai, the Claude Platform (API), Claude Code, Claude Cowork and Claude Tag, as well as Claude models accessed through cloud partners AWS, Google Cloud Vertex AI and Microsoft Foundry. That means a Kenyan developer building on Claude through Bedrock, or a user chatting with Claude.ai from Nairobi, receives the same marked output as someone using it in Berlin.
Why Anthropic Made the Change Now
The trigger is Article 50(2) of the European Union's AI Act, which requires providers of generative AI systems to mark synthetic content so it can be identified as machine generated. Anthropic has signed the EU's Code of Practice on Transparency of AI-Generated Content, and the transparency obligations under Article 50 became enforceable on August 2, 2026. Non-compliance under the AI Act can trigger fines running into tens of millions of euros or a percentage of global turnover, whichever is higher, which explains the urgency behind Anthropic's timeline.
Rather than limit the change to European users, as some companies have done with other compliance measures, Anthropic opted to apply the marking globally. The company's help documentation states that the watermark does not change the meaning, quality or readability of a response, and that it is woven directly into the token selection process rather than appended as separate metadata.
How the Watermark Actually Works
Anthropic uses two distinct techniques depending on the type of output.
Text watermarking: When a supported Claude model generates text, it applies a statistical bias to how it selects the next token in a sequence. This bias is not visible to a reader and does not alter grammar or meaning, but it creates a detectable pattern across the text. Because the signal is embedded in the words themselves rather than in a metadata wrapper, it travels with the content when copied and pasted into another document, CMS or chat window, and can survive light to moderate editing.
File provenance metadata: For supported generated files such as .png, .jpg and .svg images, Claude attaches digitally signed metadata built on the Coalition for Content Provenance and Authenticity (C2PA) open standard. A valid C2PA manifest can show that a file passed through Claude and can flag whether that metadata was later altered.
The approach mirrors what Google DeepMind has done with SynthID on its Gemini models, applying a comparable statistical biasing technique to mark AI output at generation time rather than relying solely on after-the-fact labelling. Anthropic's move makes it one of the first major frontier labs to deploy this kind of text watermarking across its entire product line at once, a step that, according to reporting from TechTimes, OpenAI has acknowledged it has not yet managed to do at the same scale.
What the Watermark Does Not Prove
Anthropic has been explicit about the limits of the system, and those caveats matter for anyone treating a detected watermark as definitive proof of AI authorship.
A watermark showing up in a piece of text only indicates that the content passed through Claude at some point. It does not confirm that Claude originated the ideas or the underlying material. Someone who writes their own article and asks Claude to proofread, translate or summarise it will still produce output carrying the watermark, even though the substance of the work is human.
The reverse also holds. The absence of a watermark does not guarantee a text was written without AI assistance. Anthropic has identified several ways the signal can be lost:
The content came from a Claude model released before August 2, 2026, since older models were not built with the marking system, though Anthropic says it is working to retrofit earlier models during the AI Act's transition period.
The text was heavily paraphrased, machine-translated, or blended with large amounts of unmarked text.
The passage is too short to carry enough token choices to establish a reliable statistical pattern.
File-level metadata was stripped through format conversion, re-saving or a screenshot, a limitation that applies equally to C2PA-based systems generally and not just Anthropic's implementation.
Anthropic has also said it plans to publish technical documentation explaining how third parties can detect the watermark and verify C2PA metadata, but that documentation, and any public detection tool, had not shipped as of this week. Until it does, institutions such as schools, employers or newsrooms have no independent way to confirm a Claude watermark's presence themselves.
A Regulatory Trend Kenya Is Also Moving Toward
The EU AI Act is not the only regulatory effort pushing AI providers toward content labelling. In Kenya, the Ministry of Information, Communications and the Digital Economy published the draft Kenya Artificial Intelligence and Other Emerging Technologies Policy, 2026 for public participation in late July, with the comment window closing on August 4. The draft policy, developed by a multi-stakeholder Technical Working Group that included the Kenya ICT Action Network (KICTANet), academia, government and industry representatives, proposes a national framework covering the governance, development, deployment and use of AI across all sectors of the economy.
Among its proposals is the classification of AI systems by risk level, mandatory registration requirements for high-risk applications, and the creation of a National AI and other Emerging Technologies Council to oversee compliance, supported by directorates covering policy and standards, compliance and risk, and safety and security. The policy does not yet carry binding legal force, and it stops short of the kind of prescriptive content-marking mandate found in the EU's Article 50, but it signals that Kenyan regulators are watching the same transparency questions that pushed Anthropic to act.
For Kenyan newsrooms, universities and businesses already relying on Claude for drafting, translation or research support, Anthropic's watermarking rollout is a preview of a compliance direction that local policy may eventually formalise. Whether Kenya's eventual AI framework adopts a similar content-marking requirement will depend on how the draft policy evolves once the National AI and other Emerging Technologies Council, if established, begins issuing binding rules rather than guidance.
Comments