Text AI watermarks will always be trivial to remove

The EU AI Act will require AI providers to watermark text outputs as "artificially generated" by August 2026, but watermarking text is technically difficult because any detectable changes compromise output quality, unlike watermarking images which can hide information in imperceptible noise. While solutions like Google's SynthID and Unicode-based approaches exist, removing such watermarks will likely remain trivial, making compliance with the regulation's intent challenging for AI labs.