{
  "id": 1447632,
  "title": "Anthropic releases more info about Claude AI watermarking, amid user confusion",
  "url": "https://urgent.news/2026/08/17/anthropic-releases-more-info-about-claude-ai-watermarking-amid-user",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-17T07:47:54.000Z",
  "source": {
    "name": "The Hindu - Sci-Tech",
    "slug": "the-hindu-sci-tech",
    "url": "https://www.thehindu.com/sci-tech/technology/anthropic-releases-more-info-about-claude-ai-watermarking-amid-user-confusion/article71355224.ece"
  },
  "original_language": "en",
  "account": "Following the announcement from Anthropic that Claude AI-generated text would bear an invisible watermark, the tech company has now provided additional details regarding the mechanics of this watermarking technology. Anthropic firmly denied the notion that textually detectable characters were being embedded within the AI-generated content, instead clarifying that the watermark is generated by the AI responding with a distinct pattern. This pattern is discernible by anyone who holds a key to decode the watermark.\n\nAnthropic revealed that Claude's watermarking process is fundamentally based on the SynthID-Text method, which was originally proposed by Google DeepMind in a paper published in Nature two years ago. The company emphasized repeatedly that the watermarking procedure would not compromise the quality, creativity, readability, speed, or pricing of Claude's operations. They also mentioned that the watermark may become more apparent over time as Claude generates more content, hence more \"space\" for the watermark in longer passages.\n\nIn terms of implementation, the watermarking effect is relatively lesser on code snippets compared to other types of text. However, translations are expected to carry the Claude watermark since every word is chosen by the AI. Anthropic clarified in its blog post that a watermark can only indicate that Claude was involved in the generation of content, not specifically stating whether Claude wrote or heavily edited the text.\n\nRegarding the potential for removal of the watermark, Anthropic stated that light editing \"probably\" wouldn't be sufficient to eliminate the watermark completely. However, if the entire text is replaced, the watermark might be removed. Importantly, Anthropic clarified that the watermark cannot be traced back to a specific individual, organization, or chat session where the text was generated.\n\nIn the months to come, Anthropic plans to extend the watermarking feature to its older models as well. The company reiterated its commitment to implementing the watermark as a measure to comply with the EU AI Act, noting that several other significant AI model providers had signed the EU Code of Practice on Transparency of AI-Generated Content in July.",
  "summary": "Anthropic noted that Claude’s text watermark was a version of the SynthID-Text approach published by Google DeepMind in a paper two years ago",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}