Text AI watermarks will always be trivial to remove

The EU AI Act will soon require AI outputs to be detectable as artificial, but text watermarking is fundamentally flawed. Unlike images, text can't be subtly altered without degrading quality. Google's SynthID embeds a statistical fingerprint in token choices, while OpenAI and Anthropic may use Unicode homoglyphs. However, both methods are easily stripped: homoglyphs can be replaced, and SynthID can be defeated by paraphrasing with any LLM. The Act's interoperability requirement further undermines security-by-obscurity, making text watermarks a compliance checkbox rather than a real safeguard.