Anthropic has activated invisible text watermarking on every Claude model launched on or after August 2, 2026 — applied worldwide at the model level with no opt-out for users or API customers. The watermark uses Google DeepMind's SynthID Text approach and embeds an imperceptible, machine-readable signal directly into generated text that travels with copied content and may survive editing.
The change, confirmed in an Anthropic support article published August 11 and detailed further on August 14, applies to all Claude models released after the August 2 cutoff. That includes any new model versions deployed globally through Claude's web interface, mobile apps, and API. Older models released before August 2 are not yet watermarked, though Anthropic says it is working to extend coverage retroactively.
How the watermark works
Anthropic's implementation uses SynthID Text, the watermarking technique Google DeepMind open-sourced in 2024. The system subtly alters token selection probabilities during generation — the model chooses between statistically equivalent words in a way that creates a detectable pattern without changing the meaning, quality, or readability of the output.
"Watermarking does not impact the quality of Claude's output," Anthropic said in its August 14 explainer. "To a reader, a watermarked response is indistinguishable from an unwatermarked one." The company also confirmed the watermark carries no identifying information and cannot be traced to a specific person, organization, or chat session.
For file outputs, Anthropic is attaching signed C2PA provenance metadata alongside the text watermark on supported file types.
EU AI Act compliance driving the rollout
The watermarking requirement stems from transparency obligations under the EU AI Act, which took effect for general-purpose AI models on August 2, 2026. Anthropic signed onto the EU AI Act Code of Practice, which requires providers to implement machine-readable marking on AI-generated content. The company is applying the watermark globally rather than restricting it to European users.
Other major AI providers are taking similar steps. Google already uses SynthID across Gemini, and OpenAI has signaled it will implement watermarking for ChatGPT. But Anthropic's approach is notably aggressive — applied at the model architecture level rather than as an application-layer feature, meaning API users cannot disable it even in private deployments.
User backlash and subscription cancellations
The watermark has drawn sharp criticism from parts of Claude's user base. Business Insider reported on August 14 that some Claude users have canceled subscriptions in protest, arguing the watermark compromises use cases where output provenance must remain private — including creative writing, journalism, and AI companion interactions.
404 Media published a critical take on August 18 arguing the watermark "proves AI companies do not care at all about writing," characterizing the move as treating AI-generated text as interchangeable tokens rather than meaningful content.
For AI companion developers and creative AI users who rely on Claude's API, the watermark represents a structural change: every response from a post-August-2 Claude model now carries a detectable signal that identifies it as AI-generated. While the watermark is invisible to end users, detection tools — which Anthropic plans to release as a public API — could theoretically allow third parties to scan companion chat logs and flag AI-generated messages.
What this means for companion and creative AI users
The watermark applies to API responses just as it does to web chat output. Any AI companion or creative tool powered by a post-August-2 Claude model will have watermarked responses by default. Developers running older Claude models (pre-August 2) are unaffected for now, but Anthropic's stated plan to extend watermarking retroactively means those models may eventually carry the mark as well.
Anthropic has not announced a detection API release date but confirmed one is in development. When available, it would allow anyone to check whether text was processed by Claude — a capability with obvious implications for companion app privacy and content moderation.