A subtle pattern invisible to readers could soon reveal whether Claude wrote a passage, but major limitations may determine how useful the system becomes.<\/p>\n
Anthropic has revealed additional details about the watermarks that will appear in texts created by the Claude chatbot. The system is intended to help identify materials generated by artificial intelligence, in line with the transparency requirements of the EU AI Act\u2019s Code of Practice.<\/p>\n
The initiative has already sparked controversy among Claude users. In Reddit discussions, some people reacted negatively to the new mechanism, while others pointed out that the absence of labeling could make it easier to spread misleading content. Business Insider reported that dozens of X users said they were canceling their subscriptions.<\/p>\n
How Claude\u2019s watermarks work<\/p>\n
Anthropic explained that the watermark will be a hidden pattern in the model\u2019s response. When Claude can choose between several equally suitable words or phrasings, it will create a sequence that a person will not notice while reading. At the same time, a specialized system with the appropriate key will be able to detect this marking.<\/p>\n
Watermarks do not affect the quality of Claude\u2019s output. For the reader, a watermarked response is no different from one without a watermark.<\/p>\n
\u2013 Anthropic<\/p>\n
For this purpose, Anthropic will use the SynthID-Text approach, introduced by the Google DeepMind team in 2024. The company is also preparing an API that will make it possible to check whether a text contains watermarks.<\/p>\n
This approach differs from tools such as Pangram. They do not look for an embedded mark; instead, they analyze linguistic features and characteristic constructions that may indicate the use of a generative model.<\/p>\n
How easy it is to remove the marking<\/p>\n
According to Anthropic, minor text edits are unlikely to remove a watermark completely. However, fully rewriting the material and replacing every word could make the marking undetectable.<\/p>\n
In the latter case, of course, one could argue whether such text can still be described as AI-generated.<\/p>\n
\u2013 Anthropic<\/p>\n
If Claude only reviews a person\u2019s text or makes minor edits to it, the result will depend on the scale of the editing and the length of the material. With minimal changes, most of the words will remain the author\u2019s own, making it difficult for the model to create a pattern prominent enough to be detected.<\/p>\n
Why watermarks in code will be weaker<\/p>\n
In programming code, the possibilities for marking are significantly more limited. Claude must generate correct commands, and in many cases it cannot replace one element with another without risking disruption to the program.<\/p>\n
In areas where there is flexibility in choosing between certain words or terms in code, a watermark can be applied, for example, in code comments. But by definition, it will have only a minor impact on the code actually produced.<\/p>\n
\u2013 Anthropic<\/p>\n
Anthropic noted that Claude will not be the only chatbot with a similar mechanism. Other major model developers have also joined the same Code of Practice and will therefore need to implement their own systems for labeling AI-generated content.<\/p>\n","protected":false},"excerpt":{"rendered":"A subtle pattern invisible to readers could soon reveal whether Claude wrote a passage, but major limitations may…\n","protected":false},"author":2,"featured_media":141117,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[67295,67197,53,9677,3154,182,69473,69472,1977,66,69474],"class_list":["post-141116","post","type-post","status-publish","format-standard","has-post-thumbnail","category-anthropic","tag-ai-content-detection","tag-ai-generated-text","tag-anthropic","tag-anthropic-ai","tag-anthropic-claude","tag-claude","tag-claude-watermarks","tag-claude-watermarks-anthropic-ai-ai-generated-text-synthid-text-ai-content-detection-eu-ai-act","tag-eu-ai-act","tag-news","tag-synthid-text"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/141116","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=141116"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/141116\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/141117"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=141116"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=141116"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=141116"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}