As an advocate of a very specific story of censoring, I do think it has some specific use cases. Particularly, hazardous information (e.g. specific research papers about bioweapons) that can’t easily be rederived from surrounding uncensored info, and which doesn’t have a bunch of common sense downstream of it which would be hurt by the censorship.
Thus, I advocate for the WMDP project’s ‘bioremove’ dataset beginning censored from training data. (I helped with this project, but the selection of papers was almost entirely done by Kevin Esvalt).
As an advocate of a very specific story of censoring, I do think it has some specific use cases. Particularly, hazardous information (e.g. specific research papers about bioweapons) that can’t easily be rederived from surrounding uncensored info, and which doesn’t have a bunch of common sense downstream of it which would be hurt by the censorship.
Thus, I advocate for the WMDP project’s ‘bioremove’ dataset beginning censored from training data. (I helped with this project, but the selection of papers was almost entirely done by Kevin Esvalt).
Do you agree with this specific use case?