{
  "id": 1516864,
  "title": "What Claudes AI text watermark actually does",
  "url": "https://urgent.news/2026/08/17/what-claudes-ai-text-watermark-actually-does",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-17T15:59:09.000Z",
  "source": {
    "name": "Mashable",
    "slug": "mashable",
    "url": "https://mashable.com/tech/what-claude-ai-text-watermark-actually-does"
  },
  "original_language": "en",
  "account": "Anthropic, the company behind Claude, has started adding a watermark to text generated by upcoming Claude models, in order to identify if a piece of writing was possibly produced by its AI. This step is in compliance with EU rules and the watermark is intended to be invisible to human eyes, without altering Claude's normal writing output. A post on Anthropic's website explains how the watermark works and its purpose.\n\nThe watermark is based on Claude's numerous small decisions while generating text. Instead of using a random number, the watermarked Claude uses a cryptographic key in combination with the preceding text to make those decisions. This generates a subtle statistical pattern within the response that can be detected only by someone with the matching key, allowing them to estimate the probability that Claude generated the text.\n\nThe watermark does not affect the output quality or speed of Claude. Internal tests found no significant difference in creativity, accuracy, or readability between watermarked and unwatermarked responses. Additionally, the feature does not increase the cost or token usage of Claude. However, the watermark might not be effective when Claude has little room for variation, such as with factual statements, precise code, or mathematical answers.\n\nThe watermark also struggles to detect heavily edited or proofread human writing, as most of the original wording remains unchanged. In such cases, it becomes debatable whether the text is still AI-generated. Furthermore, the watermark cannot identify a specific user, account, or conversation, and it does not establish authorship or ownership over content. It only indicates the likelihood that Claude was involved in producing or editing the text.\n\nAnthropic implemented the watermark to comply with EU regulations, and it will roll out the feature globally, extending it to older Claude models over the coming months. The company also plans to offer a separate API allowing anyone to check for Claude's watermark in text.",
  "summary": "Anthropic explains how Claude's new text watermarking works, why it's rolling out, and what it can and can't prove.",
  "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."
}