Text AI watermarks will always be trivial to remove (www.seangoedecke.com)

The European Union's AI Act, effective August 2026, mandates that AI outputs must include detectable watermarks indicating their artificial origin. Notably, Article 50 of the Act requires LLM providers to embed watermarks in their text outputs, igniting a complex engineering challenge. While companies like Google are developing watermarking solutions such as SynthID, the inherent difficulties in modifying text without compromising quality pose significant hurdles. Unlike image watermarking, text is more compressed and sensitive to changes, making it easier to remove these signatures once recognized by users. The significance of this development for the AI/ML community lies in the balancing act between maintaining output quality and meeting regulatory compliance. Proposed methods, such as the mathematical scoring in SynthID or the clever use of Unicode homoglyphs, aim to establish durable watermarks. However, these techniques are not foolproof, as technical users can potentially bypass them through various means. This raises critical questions about the efficacy and long-term viability of text watermarks, suggesting that while compliance mechanisms are evolving, sophisticated methods to detect AI-generated content will continue to be a contentious and evolving field.