The EU AI Act will become enforceable in August 2026 and Article 50 demands that all AI outputs be detectable as artificially generated.

Text watermarking is far more difficult than image watermarking because text is a compressed medium where any alteration can be noticed by a human eye..props to keep quality while embedding a hidden signature is a classic steganography problem.

Google’s Synth ID assigns each token a score based on previous tokens and forces the model to pick the token with the highest score from the top candidates. The watermark is then detected by summing the scores of a block of text, a cheap operation that still preserves model performance.

An alternative is the use of Unicode homoglyphs, such as three‑per‑em spaces, to encode a pattern of spaces that is unlikely to appear in natural text. However, both watermarking methods can be trivially removed by replacing homoglyphs or re‑paraphrasing the text, undermining the EU requirement that故事 be embedded in a manner difficult to separate.