Anthropic Will Watermark Claude-Generated Text and Files to Comply With EU AI Act
Anthropic will begin watermarking text and files produced by its AI models, including Claude, to comply with the European Union's AI Act Transparency Code. The company confirmed the change on August 11, 2026, through an updated support page describing how Claude marks AI-generated content (TechCrunch).
The EU AI Act's Transparency Code took effect on August 2, 2026. It requires AI companies to tag AI-generated or edited content in a machine-readable way, meaning the marking can be detected by software rather than relying on human inspection. All Anthropic models released after that date will include watermarking automatically, applied to both text output and generated files. Models launched before August 2 are still being updated to add the capability (India Today). Anthropic has stated it will extend watermarking support to its older models as well.
For files, Anthropic is adopting the C2PA open standard, a provenance specification backed by a coalition of media and technology organizations. Think of C2PA as a digital certificate of origin attached to a file, recording where it came from and whether AI was involved in producing it. The text watermark works differently: it is embedded at the model level, meaning it is present regardless of which Claude product or interface the output comes from. According to Anthropic's support documentation, the watermark travels with the text when copied and pasted elsewhere and may survive some degree of editing.
The watermarking will apply across Anthropic's product surface area, including the Claude platform API, Claude, Claude Code, Claude Cowork, and Claude Tag.
Anthropic is not alone in aligning with the EU's transparency framework. Black Forest Labs, Google, Meta, Microsoft, OpenAI, and Synthesia have all committed to adhering to the EU's code of practice on transparency of AI-generated content (European Commission).
The broader context here matters for developers building on Claude's API. The EU AI Act's Transparency Code obligations create a compliance surface that extends to downstream applications. If a developer integrates Claude's API into a customer-facing product, the watermark will be present in the output by default. That has implications for any pipeline that processes, transforms, or re-transmits model output: the provenance signal is designed to survive copy-paste and at least some editing, which means it could surface in contexts where its presence is unexpected. Developers should not assume that model output consumed via API and then reformatted will shed the watermark.
There is also a historical dimension worth noting. In July 2023, OpenAI, Google, and other companies pledged at a White House event to develop systems for watermarking AI-generated content across text, images, audio, and video (Reuters). That voluntary commitment, made in a fundamentally different regulatory environment, produced little in the way of durable, interoperable watermarking infrastructure. The EU AI Act changes the calculus by attaching legal obligation to what was previously a voluntary pledge. Whether the C2PA-based approach Anthropic is taking for files and the model-level text watermarking approach will hold up under adversarial conditions, where someone is actively trying to strip or fool the watermark, remains an open technical question.
Anthropic's Transparency Hub, updated on July 23, 2026, lists technical watermarking of AI-generated content among the company's voluntary commitments related to public awareness of AI capabilities and limitations (Anthropic). The watermarking announcement follows a period of active product and research cadence at Anthropic. The company announced Claude Sonnet 5 on June 30, 2026, and Claude Opus 5 on July 24, 2026, both emphasizing improvements in coding, agentic workflows, and professional work at scale (Anthropic News). On July 1, 2026, Anthropic announced the global return of Fable 5 alongside a proposed industry-wide framework for scoring jailbreak severity, developed with Amazon, Microsoft, Google, and other Glasswing partners. On July 9, 2026, the company launched a public initiative asking for difficult questions about AI and committing to transparent responses.
In practical terms, what this creates is a provenance layer that travels with AI-generated content rather than depending on the platform of origin to disclose it. For organizations building compliance-aware pipelines, model-level watermarking shifts a portion of the transparency burden upstream to the model provider. For the broader ecosystem, it creates a testable claim: can a watermark embedded during generation survive the messy, multi-hop reality of how text actually moves through systems? The answer to that question will determine whether this approach becomes a durable standard or a compliance checkbox.
