OpenAI Will Automatically Watermark ChatGPT Text in the EU — and Its Mark Is Likely Easy to Circumvent
OpenAI has announced that it will begin automatically watermarking text generated with ChatGPT in the European Union. It is joining a small club. The marking standards that already exist — SynthID and the C2PA project among them — are relatively easy to circumvent for anyone with basic know-how, and Ars Technica's assessment is that OpenAI's watermark is likely to be no different. The technical paper behind it, if it explains the method well, will not necessarily make the mark harder to remove than the others; that isn't what watermarking schemes have been good at so far. Which makes this more about provenance than proof. A mark a curious user can get around is closer to a nudge than a seal on machine-written text.
AI-generated discussion · Different perspectives
The hedge is the interesting part: the standards that already exist are described as easy to get around, and the judgment about OpenAI's version is a prediction rather than a demonstrated result.
A mark doesn't have to be unbeatable to be worth having. Filtering out the careless case is a lower bar, and a lower bar can still be a useful one.
Then the useful question is what the mark is for. If the point is to settle whether a text was machine-written, easy circumvention is a serious problem; if the point is to raise the floor, it's a design choice.