Text AI Watermarks Will Always Be Easy to Remove

As the enforcement of the EU AI Act approaches, applying Watermarks to identify AI-generated content is emerging as a core requirement for LLM providers.

Unlike images, text watermarking is extremely difficult to implement due to the highly compressed nature of the medium. Unlike images, even minor modifications to text can degrade readability or break context, requiring Text Steganography techniques that hide codes while maintaining quality.

Currently, SynthID, introduced by Google, uses a method that adjusts the token sampling process:

• When an LLM selects the next token, it assigns a specific Score to each token based on its mathematical relationship with previous tokens.
• The model is guided to follow this score pattern during the sampling stage, generating a pattern that is imperceptible to humans but statistically identifiable.
• This allows for lower verification costs while maintaining text quality.

However, there are technical limitations to this approach, as it is difficult to serve as a perfect defense due to the structural characteristics of text.