Text AI watermarks will always be trivial to remove
The EU AI Act requires AI text outputs to be watermarked by August 2026. Methods like Google's SynthID and Unicode homoglyphs exist, but all text watermarks can be trivially removed via paraphrasing or character replacement, making them fundamentally different from image watermarks.
Background
The EU AI Act, enforceable from August 2026, requires AI-generated text to be detectable via watermarks. Text watermarking is fundamentally harder than image/video watermarking because text has no "noise" to hide a signature in — any change is noticeable. Existing approaches: Google's SynthID alters the model's token-selection math to leave a statistical fingerprint; OpenAI/Anthropic may use Unicode homoglyphs (visually identical characters with different code points). Neither is robust — SynthID is defeated by paraphrasing (even with a weak unwatermarked model), and homoglyphs are removed by normalizing Unicode characters. The Act also requires interoperability (publishing watermark methods), which undermines security-by-obscurity. C2PA metadata signing doesn't help with plaintext chat output, which has no file format to attach metadata to. The practical outcome: watermarks will be trivially stripped by any motivated user.