Our approach to EU text provenance rules

Commenters discussed OpenAI’s EU-focused text-provenance measures - an opt-in API watermark and an “invisible” watermark on eligible EU outputs. Some argued the watermark is easy to defeat: one cited an evaluation where replacing 10-25% of words with synonyms dropped detection dramatically, and others said simple paraphrasing or using the model itself could produce watermark-friendly rewrites. Several commenters explained technical details as held views: that the watermark is added during sampling by biasing token probabilities, that variance in sampling explains benchmark shifts, and that watermarked models can be designed to preserve the mark when rephrasing.

Opinion splits on purpose and impact. One camp sees watermarking as legally useful and a deterrent - several suggested removed watermarks could be evidence of intent in court and that regulations can help attribute harmful content. The opposing camp called it malicious compliance or a “cookie consent”-style cost that mainly burdens smaller players, argued it’s trivial to bypass, and worried it enables regulatory capture and distraction from building better models. Others suspected evaluation gaming or doubted meaningful benefits, so the debate centers on technical robustness versus legal utility and the political economy of regulation.