Anthropic Claude invisible text watermark triggers user cancellations; open-source unwatermark project hits 11k GitHub stars
Reports on August 17 said Anthropic has embedded SynthID-Text invisible watermarks in Claude outputs since August 2 to comply with the EU AI Act. The move has driven dozens of users to publicly cancel subscriptions on X, while an open-source unwatermark project has gained 11,000 GitHub stars but Claude refuses to integrate it.
Multiple reports on August 17 said Anthropic has embedded SynthID-Text invisible watermarks in Claude outputs since August 2, with the new approach injecting markers at the model layer during text generation via subtle lexical adjustments that survive copy-paste and partial editing. Anthropic said the move responds to the EU AI Act adopted in 2024, which requires AI companies to label system-generated or edited content; Google, Meta, Microsoft, OpenAI and Synthesia have also committed to compliance.
The disclosure triggered intense discussion on X. Some users worry AI-assisted writing will carry permanent markers and become stigmatized. Czech freelance developer Vladislav Rajtmajer cancelled his Claude Max subscription over concerns that code delivered to clients would carry AI labels triggering copyright and contract issues; Georgia-based AI consultant Richard Echols cancelled for the same reason, noting that even content he personally authored and only edited with Claude gets watermarked.
Meanwhile, an open-source watermark-removal project has gained 11,000 GitHub stars, but Anthropic has refused to integrate it into Claude. Senior technical consultant kmleon, scanning a year of projects, found roughly a dozen Claude mislabeling cases between Sonnet 4.6 and August 14, 2026, mostly concentrated in recent weeks, where Claude marked content it had only marginally participated in as 100 percent authored or co-created. In contrast, Google announced on August 15 that users may remove visible watermarks from AI-generated images, video and music created by its models, a stance opposite to Anthropic's mandatory invisible watermark approach.