{
  "id": 4936438,
  "title": "Claude Fable 5.1 watermark: It has a blind spot developers can’t ignore",
  "url": "https://urgent.news/2026/09/01/claude-fable-5-1-watermark-it-has-a-blind-spot-developers-cant-ignore",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-01T20:59:09.000Z",
  "source": {
    "name": "The New Stack",
    "slug": "the-new-stack",
    "url": "https://thenewstack.io/fable-5-1-watermark/"
  },
  "original_language": "en",
  "account": "Anthropic introduced Claude Fable 5.1 on Tuesday, incorporating a watermark system into its generated text. Developers should not expect this watermark to be equally strong across all generated content. The system alters Claude's randomness in choosing tokens to create a statistical pattern that indicates the model's involvement in writing or processing the text. This watermark works better with natural language than with code, as changing a variable, operator, or function in code can alter the program's behavior. Anthropic does not apply the watermark when a token is necessary for accuracy. Developers must also adjust to a new limitation on preserving thinking blocks. With Fable 5.1, Anthropic limits when new API accounts can carry these blocks into a modified conversation. Over a sufficiently long response, token choices create a statistical pattern detectable with the right key. Since the watermark is part of the text, copying it will not remove it. The watermark can survive some editing but will eventually be erased if the text is sufficiently rewritten. Code's low-entropy nature poses a problem for watermarking, as changing a token could make an answer incorrect or break the code. In such cases, Anthropic does not apply the watermark. The watermark can still appear in less-constrained parts of the output, such as comments. Anthropic is rolling out the watermark worldwide as it has been signed by the EU Code of Practice on Transparency of AI-Generated Content. The watermark is embedded in the text itself, so simply copying a response elsewhere won't remove it. Anthropic can detect the watermark with an API in private preview, initially available to regulators, law enforcement, media organizations, fact-checkers, researchers, and enterprises needing AI Act compliance. The watermark is expected to be added to older Claude models in the coming months.",
  "summary": "Anthropic launched Claude Fable 5.1 on Tuesday with a statistical signature embedded in its generated text, but developers shouldn’t expect The post Claude Fable 5.1 watermark: It has a blind spot developers can’t ignore appeared first on The New Stack .",
  "key_points": [
    "Claude Fable 5.1 introduces watermark system in generated text.",
    "Watermark works better with natural language than code.",
    "Developers face limitation on preserving thinking blocks."
  ],
  "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."
}