The case for text watermarking

Text watermarking takes advantage of randomness already present when a model selects its next token. A secret key determines that randomness; a detector checks whether a sufficiently long passage fits its pattern.

Zvi Mowshowitz strongly favors adoption. His argument is that well-designed schemes can provide useful provenance at negligible practical cost or quality loss. This is different from guessing that a passage sounds machine-written: the signal was deliberately introduced when it was generated.

Rewriting weakens it, short passages may offer too little evidence, and an absent signal does not prove human authorship. Those limitations narrow what detection establishes without making the signal useless. The argument is for a workable indication of origin, not an infallible detector of all AI-assisted writing.