Tezzbuzz Desk- With the increasing use of Artificial Intelligence i.e. AI, it can now be easy to find out whether any content has been written by a human being or has been prepared with the help of AI. AI company Anthropic has introduced a new hidden watermarking system for text generated from its Claude model. The special thing is that this watermark will not be visible in the text. Instead, Claude will create a special pattern while selecting words, which can be identified with the help of a special detection system.
According to Anthropic, the purpose of this technique is to find out whether Claude had a role in preparing a text or not. If a user creates content with Claude and does not make too many changes in it, then in future there will be a possibility of detecting through this system that that text was generated by an AI model. According to the company, when Claude chooses one of two or more equally suitable words, the system will control its randomness in a special way. This will create a special pattern throughout the text. This pattern will not be visible to the common reader, but it can be identified with the help of a special detection system and security. Anthropic says Claude will not be forced to use strange, unusual or forced words in the process.
For this, the company has used a version of Google DeepMind's SynthID-Text technology published in 2024. Anthropic's internal testing claimed that this system did not have any significant negative impact on the quality of content, creativity or reading experience. That means the user should not feel any obvious difference between a normal Claude answer and a watermarked answer. However, this system also has some limitations. A hidden watermark may indicate that Claude had a role in the creation of a text, but it will not prove the complete origin of any content. Through this, personal information related to the user's identity, his organization or his conversations will not be revealed. It is also not necessary that the system can recognize any text that has been generated from another AI model.
Watermark detection may be less reliable in smaller text, as the system has less material to analyze word patterns. Apart from this, its capability may also be limited in matters like factual writing, proofreading and code, because the scope of choice of words is relatively less in these. According to Anthropic, there is a possibility of watermarks remaining even after light editing. But if any text is completely rewritten, the original watermark may be lost. The watermark may also be present in translations done by Claude, because in such cases also Claude chooses the words. The company is also working on a detection API, through which users and third parties will be able to check this watermark in their text.
Anthropic does not want to be limited to just text. C2PA Content Credentials will also be used for some supported image files created or processed with Claude. Cryptographically signed metadata in files such as PNG, JPG, and SVG may indicate that Claude played a role in the creation or processing of the file. User's personal information will not be included in this. Anthropic is implementing this watermarking system worldwide. According to the company, at present there is no reliable way to limit it to different areas. This step is also important towards meeting the AI content transparency requirements of the EU's AI Act. Anthropic had signed the EU Code of Practice related to transparency of AI-generated content in July 2026. The company is also working on adding watermark to Claude models launched before August 2, 2026 in the coming months.