Anthropic has disclosed comprehensive details about its invisible digital watermarking system for AI chatbot Claude, as reported by ITmedia on August 16. The method is based on Google DeepMind's SynthID-Text technology and replaces random-number generation with a secret key and the immediately preceding context to leave statistical patterns. According to Anthropic's explanation, the watermarking is part of the company's compliance with the EU AI Act's Transparency Code, which requires AI companies to implement systems that make AI-generated content identifiable. The company explained that when Claude makes 'low-stakes choices' — such as selecting between words like 'overcast' and 'grey' to describe weather — it can embed a pattern in its responses that is 'undetectable to the reader, but is detectable to anyone who has a key that encodes it.' Anthropic emphasized that watermarking does not affect output quality: 'To a reader, a watermarked response is indistinguishable from an unwatermarked one.' The company will use the SynthID-Text approach developed by Google DeepMind in 2024 and plans to release a watermark detection API.
Cardano founder Charles Hoskinson announced on X on August 16 (UTC+8) the launch of the open-source project 'anthropies,' aiming to develop a comprehensive set of tools compatible with most large language models to detect watermarks and markers in Anthropic/Claude outputs. According to Odaily News, the project appeared on GitHub on Sunday, five days after Anthropic switched on invisible marking across every Claude product. Hoskinson shipped the tool under an Apache 2.0 license, which is a permissive open-source term that lets anyone copy, modify, sell, or bundle the code, provided they keep the original notices and grant a patent license. As reported by ITmedia, Anthropic plans a global rollout tied to an EU AI Act transparency code of practice, with the scope extending across models and products worldwide, including South Korea. About 190 organizations, including Anthropic, signed the code of practice, and other major AI developers also plan to implement their own watermarks. The measures apply to models launched in the European Union after August 2, 2025, with Anthropic implementing the marking technology wherever those models are offered worldwide.
The latest 'anthropies' project distinguishes three types of information: statistical watermarks in text, C2PA content credentials in images, and 'Co-Authored-By: Claude' signatures in Git commits, as detailed in the project's README. However, the project explicitly notes that it cannot guarantee complete evasion of Anthropic's undisclosed detection mechanisms. This represents an evolution from Hoskinson's previous tool, which focused primarily on removing Claude watermarks through different layers of the output generation process. Anthropic acknowledged that while light editing probably won't remove the watermark completely, 'a complete rewrite where every word is replaced will.' For text that Claude only proofreads or lightly edits, the watermark's presence depends on 'the length of the text and how heavily Claude has edited it.' If only lightly edited, 'nearly all the words' come from the human author, leaving 'very little (if anything) for the watermark to attach to.'
The repository ships three modes: Clean strips trailers deterministically, humanize rewrites prose elsewhere, and orchestrate blocks the job when the host model is Claude. As reported by CoinDesk, code barely carries a signal at all due to syntax limitations, so the watermark has nowhere to sit. However, the tool carries a blunt limitation as running the rewrite inside Claude or Gemini would stamp the mark straight back on, so the skill refuses to do so. Anthropic noted that code will carry less watermarking than other text because the model must produce functional code, limiting its freedom to choose between equally valid options. In areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, but it will have a 'negligible effect' on the actual code produced. The latest 'anthropies' project acknowledges these limitations while expanding detection capabilities beyond previous tools.
Hoskinson framed the release on X as a warning rather than a utility, arguing that the watermark and Claude co-author line could cause issues down the road. According to CoinDesk, roughly half the repository is not code, and Hoskinson attacks the assignment clause that assigns ownership of output to users, which applies 'subject to your compliance with our Terms.' He reads that phrase as a condition precedent, meaning a hurdle users must clear before ownership passes, with a breach of terms potentially meaning ownership never transferred. The controversy has triggered significant backlash, with 'dozens' of users on X claiming to cancel their subscriptions, as reported by Business Insider. Some users on Reddit have called it a 'conspiracy against innocent Claude users,' while others argue that opposition suggests a desire to deceive. Anthropic noted that existing models released before August 2, when application of Article 50 of the EU AI Act began, are subject to a legal transitional measure, with watermark additions applied sequentially over the coming months. The deadline for legacy systems to comply is December 2, 2026, giving other major AI providers several months to reveal their identification approaches.
The tool launch comes as Anthropic prepares a listing above $2 trillion and has already withheld its strongest model from the public. According to CoinDesk, the watermark rollout answers the European Union AI Act transparency code, which took effect on August 2, 2025. The controversy highlights ongoing tensions between AI transparency and user autonomy, with Anthropic stating that watermarking aims to meet regulatory requirements while minimizing disruption to users. The EU framework distinguishes between technical marking by AI providers and disclosure obligations for specific types of content, including certain deepfakes and AI-generated material concerning matters of public interest. Anthropic noted that other major AI developers have signed the same Code of Practice and will implement their own watermarks, indicating industry-wide adoption of watermarking systems. For images, Claude generates cryptographic signatures using the C2PA metadata approach, following the same standard used by camera makers and photo-editing software, which can be checked with C2PA-compatible tools. The push is also spreading to platforms, with Substack partnering with AI-detection company Pangram and Snap prioritizing original content on Spotlight.
