Anthropic Implements Watermarking for Claude-Generated Text to Adhere to EU Regulations
Anthropic has announced its new technique for watermarking text generated by its Claude AI, designed to adhere to the European Union’s recently established AI transparency regulations. According to the company, the watermarking will not be visually obvious, will be imperceptible to readers, and will not involve inserting hidden characters. Instead, Anthropic’s approach embeds a specific pattern within the text, which can only be decoded by someone possessing the appropriate key.
The firm clarified that large language models generate text by selecting one word at a time from a pool of possible choices. Words are chosen based on context rather than random selection. When watermarking is activated, Claude will determine which word to pick based on a key, rather than relying on a typical random number generator.
For example, Anthropic illustrated its method using the digits of pi as a key. If the process starts with the digit 2 from the pi sequence 3.1415926535, the next word generated would be the sixth option in the list, then the fifth, followed by the third, and so on. This technique builds on the SynthID-Text watermarking method developed by Google DeepMind, which was detailed in a study published in a scientific journal. The company maintains that watermarking does not diminish the quality of Claude’s output or slow down its performance, and it will not necessitate additional tokens or raise generation costs.
Despite the implementation of watermarking, AI-generated text often exhibits distinctive traits. For instance, models frequently use specific sentence structures like “this isn’t [X], it’s [Y].” However, these traits indicate that an AI contributed to the writing without revealing which model was employed. Anthropic plans to provide an API equipped with keys to decode the watermark, enabling the identification of whether Claude was involved in producing the examined text.
Anthropic acknowledges that the watermarking technique has its limitations. It cannot definitively ascertain whether Claude authored the text or merely edited it, meaning that if the AI is tasked with editing an existing document, that text will also carry a watermark. The company points out that its method can only suggest Claude’s involvement at some stage. Additionally, any translated content will also be tagged. If Claude only proofreads or makes light edits to a text, or if the text is very short, the watermark might be too subtle to detect. Importantly, lightly editing Claude’s original content may not fully erase the watermark, and complete rewrites may be necessary for certainty.
There are also concerns regarding how watermarking might influence code, as its copyrightability typically requires significant human input. If a complete codebase were proven to be AI-generated, it could be copied and modified. Anthropic noted, however, that code generally features “less watermarking than other types of text” because it usually demands precise output. Thus, if no choices exist in the text creation, applying a watermark becomes complicated.
Furthermore, Anthropic plans to watermark images by incorporating a cryptographically signed note in the metadata indicating that it was generated by Claude. In the announcement, the company stated that it would apply watermarking to all outputs from Claude at launch due to the lack of reliable methods for implementing region-specific changes. This initiative will impact all Claude products utilizing models released after August 2. Additionally, the company aims to introduce watermarking capabilities for older models over the next few months.
