Text AI watermarks will always be trivial to remove
3 hours ago
- The EU AI Act, enforceable from August 2026, mandates that AI outputs must be detectable as artificially generated, requiring LLM providers to apply watermarks.
- Text watermarking is challenging because text is a compressed medium where changes are noticeable, unlike images; solutions include Google's SynthID and Unicode homoglyphs.
- SynthID works by influencing token sampling to leave a mathematical fingerprint that can be cheaply detected, while Unicode watermarks use homoglyphs to encode patterns.
- Text watermarks are easily removed through paraphrasing or replacing homoglyphs, making them unreliable for long-term detection.
- C2PA metadata signing is not a substitute for text watermarking, as it only applies to file-based content, not plain text outputs.