{
  "id": 1520663,
  "title": "Claude to start watermarking AI-generated text – but will it make quality worse?",
  "url": "https://urgent.news/2026/08/17/claude-to-start-watermarking-ai-generated-text-but-will-it-make",
  "topic": "ai",
  "section": "AI",
  "published": "2026-08-17T16:52:04.000Z",
  "source": {
    "name": "Guardian Technology",
    "slug": "guardian-technology",
    "url": "https://www.theguardian.com/technology/2026/aug/17/claude-watermark-ai-text-quality-worse"
  },
  "original_language": "en",
  "account": "Anthropic, the creators of the Claude AI model, have announced plans to introduce watermarks to its text generation process. This change aims to comply with an upcoming EU regulation that mandates watermarks for all AI-generated text starting in December. The update will affect the way Claude makes small, random decisions in constructing sentences, but the alterations will be imperceptible to the average reader.\n\nJohn Gruber, a well-known tech blogger, expressed concerns that this watermarking would constrain Claude's capabilities, leading to poorer word choices and less precise writing. However, Steven Murdoch, a professor of computer science at University College London, suggested that these changes would likely have no noticeable impact on the quality of Claude's output. He argued that randomness is already a key component in the functioning of large language models, and removing it could potentially hinder the model's ability to generate diverse content.\n\nMurdoch emphasized that the current randomness in these models helps prevent them from getting stuck in repetitive loops and continually repeating the same phrases. The recent retraction of a paper from a leading chemistry journal serves as an example of how AI-generated text can lead to inaccuracies and misleading information. To combat such issues, Murdoch pointed out that watermarks may also serve an important function in preventing \"model collapse.\" This phenomenon occurs when AI models trained on AI-written content become confused and start mixing up concepts, ultimately deteriorating the quality of their outputs.\n\nWhile some critics may view the watermarking as an adulteration of the writing process, Murdoch maintained that there would be no discernible difference in the quality of Claude's output due to these changes. The EU regulation requires all AI companies operating within the region to implement these watermarks within months, which could potentially make it more difficult for students, lawyers, and professors to claim authorship of chatbot-generated content.",
  "summary": "Anthropic says it will change way chatbot makes small, random choices, to comply with EU regulation The world is familiar by now with the usual tropes of machine-generated text: overuse of the word “delve”, an excess of em dashes, and the chirpy, relentless construction of “it’s not X but Y” . But is it about to get even worse? Continue reading...",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "9to5Mac",
        "title": "AI-generated text should be detectable, but Apple needs to avoid Anthropic’s huge error",
        "url": "https://urgent.news/2026/08/17/ai-generated-text-should-be-detectable-but-apple-needs-to-avoid",
        "published": "2026-08-17T13:30:47.000Z"
      }
    ]
  },
  "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."
}