Anthropic introduced an imperceptible watermark in text generated by Claude models starting August 2, 2026, to comply with the EU AI Act. This watermark is embedded in the model's word choices rather than hidden characters or metadata. Following this, multiple 'watermark remover' tools have surfaced, including an MIT-licensed open source project and commercial AI detection evasion services. These tools claim to remove or bypass the watermark, but their effectiveness cannot be verified because Anthropic has not published a detector or detailed technical documentation. Most tools currently only remove metadata or hidden characters, which does not affect the watermark embedded in the text. Removing the watermark would require heavy rewriting of the text using a second model, which some tools do not perform. Anthropic acknowledges that watermark detection indicates content was processed by Claude but not necessarily authored by it. The watermark can be lost through paraphrasing or translation. The proliferation of these tools, some integrated as agent skills, introduces a supply chain risk as their claims are unverified and they may be used in automated pipelines.

The watermarking by Anthropic is intended as a compliance mechanism for the EU AI Act, enabling detection that text was generated or processed by Claude models. The inability to verify watermark removal claims means that users relying on these tools may have a false sense of security about evading detection. The watermark itself is resilient to simple metadata or hidden character removal and requires substantial rewriting to remove. The presence of unverified watermark removal tools, some integrated into agent pipelines, could introduce supply chain risks if malicious or poorly implemented code is used. However, the watermark does not guarantee attribution of authorship, only that Claude processed the text. Heavy editing or paraphrasing can remove the watermark, reducing its reliability as a provenance indicator.

Anthropic has not released a public watermark detector or detailed technical documentation yet, so the effectiveness of watermark removal tools cannot be independently verified. Users should treat all watermark removal tools with caution and avoid integrating unvetted tools into automated pipelines to reduce supply chain risk. Since the watermark is embedded in word choice, only heavy rewriting using a second model can remove it, which is not currently implemented in most tools. Anthropic plans to support third-party detection as required by EU transparency rules and will publish technical details later. Until then, no definitive remediation or bypass method is confirmed. Organizations should monitor vendor advisories for updates on watermark detection and removal capabilities.