lol this is awesome. savvy operators will automate ways to remove the watermarks, but this will catch everyone below a certain level of sophistication with far more irrefutable proof than Pangram
@tsarina-anadyomene and other grad students, unclear exactly when this will be rolled out globally but this could be of interest to y'all
Good news. I do wonder if this will just lead to OpenAI monopolizing undergrads unless they sign on too. I looked into it a little and they avoid saying they've signed onto anything and instead have just stated their "support for the Code and the ambition behind it" which is weaker language, and haven't made any mention of text watermarking except as an area of research in the other document I found. But if they're serious about it this could mean that soon it'll be a lot easier to detect outright cheating with AI. Which (if any undergrads are reading this) protects both the educative and the economic value of the degree in the long run.
I'm also curious though if it would be possible to reverse engineer the text watermark pattern by manipulating the model — perhaps by tricking it into minimally processing human written inputs so that the pattern can be detected by the attacker in the outputs? — to produce dewatermarking tools that anyone could use, savvy or unsavvy. Certainly those tools would be in demand for the same reason watermarking is in demand.
this is based on research done originally at OpenAI by Scott Aaronson iiuc
