[DigitalToday reporter Jinju Hong] Anthropic has introduced a feature that inserts a watermark that is difficult for the human eye to detect into text written by its generative AI model Claude. The move is intended to help identify cases in which AI-generated writing is submitted or published as if it were written by a person, and is expected to affect how the education and publishing industries verify AI use.

On Aug. 11 local time, Business Insider reported that Anthropic said it would insert an "undetectable watermark" into generated text starting with newly released Claude models launched that day.

The watermark does not change the meaning or readability of the text, and is designed to travel with it even if a user copies and pastes it elsewhere. Anthropic explained that the watermark may remain even after some editing.

The feature applies to Claude models released since Aug. 2. Anthropic said those models support text watermarking from the time of release, and it is working to expand the feature to earlier models.

The watermark applies not only when users access Claude directly, but also when they use Claude models through cloud providers. Anthropic plans to provide a detection tool for third parties so outside institutions can verify whether content was generated by AI.

The move is also linked to efforts to strengthen transparency under the European Union's AI Act. It is intended to make it easier to check the source of AI-generated content so users and content consumers can judge whether AI was used.

Publishers and schools, including universities, could gain a new way to verify AI-generated text. As disputes continue over whether AI was used in published works such as novels, how to check the creation process and whether AI was used has emerged as a key issue.

Last month, the crime novel "Call Me, I'll Hide the Body" faced allegations that the author may have used AI, and an agent withdrew support. The controversy grew after 14 publishers took part in a bidding process for publication rights and contracts were signed. Author Jerry Pallade denied writing the novel with AI.

Earlier this year, the horror novel "Shy Girl" was also caught up in controversy over AI-generated phrasing. Author Mia Bollard claimed she did not use AI and that a freelance editor inserted AI-generated phrases without her consent.

As such cases continue, text watermarking could become a new clue that publishers can refer to when checking whether AI was used in manuscripts. In education, it could also be used to verify whether AI was used in student assignments.

The watermark is not a perfect way to determine whether content was generated by AI. Anthropic also acknowledged that it may fail to detect the watermark after large-scale edits, paraphrasing, translation, or mixing with other text. It also said that even if a watermark is found, it does not mean Claude wrote the entire text, because the watermark can remain when Claude is used to proofread or translate writing created by a person.

Anthropic is the second major AI research lab to introduce text watermarking. Google DeepMind said in 2024 it would apply SynthID-based watermarking to text and video generated on the Gemini app and the web. It has provided watermarking technology for images since 2023.

Ultimately, Anthropic's move is significant in that it has 마련 a minimal technical mechanism to trace the source of AI-generated text. It is difficult to confirm AI use through a watermark alone, but as publishers and educational institutions gain a new verification tool to use alongside existing AI detection methods, debate over transparency in AI writing and verification of authored works is expected to intensify.