Anonymity is hard – the posts themselves, e.g. word choice, phrasing, etc., are all extremely vulnerable ‘side channels’ for breaking anonymity. Defending against those kinds of attacks are probably both impractical and extremely difficult, if not effectively impossible.
The most important step in defending against attacks is (at least roughly) determining a threat model. What kind of attacks do you expect? What attacks are worth defending against – and which aren’t?
But as the OP notes, we need only defend against mobs and hard evidence. If, e.g., Scott writes pieces with throwaway accounts that are obviously in his style, he can always claim that’s an imitator. The mob lives for easy outrage, so as long as we decrease that factor we have a partial solution.
I don’t think that would work because a definite identity is needed for people to follow Scott. I don’t think I could possibly track ‘Scott’, or anyone, and notice that there was a specific identity, if I couldn’t track a named identity, i.e. a specific account.
Part of who Scott is to a lot of us is someone, i.e. a specific identity, that’s worth following, tracking, or paying attention to over long periods. Using throwaway accounts makes that harder, even ignoring the kind of fuckery that someone could pull with something like GPT-3 at hand – maybe it’s even (effectively) impossible now.
And to the degree that we all could continue to track Scott, even with him posting via throwaway accounts – and even assuming otherwise perfect opsec from Scott too – so could people who feel malice towards him track him similarly. We’d see ‘likely Scott posts’ lists. We wouldn’t be able to prevent people from speculating as to which posts were his. Scott would only be protected to the degree that we couldn’t tell it was him.
There’s probably a math/compsci/physics theory that covers this, but there’s no way for Scott to be able to maintain both his pseudonymity and his identity – to the degree he wants (or needs) – if his (full) legal name is linked to his pseudonym on the first page of web search results for the legal name.
The safer option would be for Scott to create a new pseudonym and try to maintain (very) plausible deniability to any connection to Scott Alexander. But that would mean starting over and would probably be very hard – and maybe impossible anyways. It might be impossible for anyone to write in a different enough style, consistently, and avoid detection, especially at the lengths he typically posts. It’s probably even that much harder (or impossible) for people that have written as much as he has publicly.
I’ve seen some presentations about how to do style-matching off of GitHub repos to pretty-confidently ID anonymous coders. While set-up requires a sizable amount of compute and data, the results have gotten quite impressive. There are ways to work against this (stuff that deliberately obscures your coding style, usually by rewriting your code), but they’re not that well known. And a similar thing can be done with writing style and writing samples.
Staying anonymous against high-effort attempts to discern your identity has gotten very hard, and is only likely to get harder.
At some point, all you can do is guard against the low-effort ones.
Anonymity is hard – the posts themselves, e.g. word choice, phrasing, etc., are all extremely vulnerable ‘side channels’ for breaking anonymity. Defending against those kinds of attacks are probably both impractical and extremely difficult, if not effectively impossible.
The most important step in defending against attacks is (at least roughly) determining a threat model. What kind of attacks do you expect? What attacks are worth defending against – and which aren’t?
But as the OP notes, we need only defend against mobs and hard evidence. If, e.g., Scott writes pieces with throwaway accounts that are obviously in his style, he can always claim that’s an imitator. The mob lives for easy outrage, so as long as we decrease that factor we have a partial solution.
The mob doesn’t care if you have plausible deniability.
I don’t think that would work because a definite identity is needed for people to follow Scott. I don’t think I could possibly track ‘Scott’, or anyone, and notice that there was a specific identity, if I couldn’t track a named identity, i.e. a specific account.
Part of who Scott is to a lot of us is someone, i.e. a specific identity, that’s worth following, tracking, or paying attention to over long periods. Using throwaway accounts makes that harder, even ignoring the kind of fuckery that someone could pull with something like GPT-3 at hand – maybe it’s even (effectively) impossible now.
And to the degree that we all could continue to track Scott, even with him posting via throwaway accounts – and even assuming otherwise perfect opsec from Scott too – so could people who feel malice towards him track him similarly. We’d see ‘likely Scott posts’ lists. We wouldn’t be able to prevent people from speculating as to which posts were his. Scott would only be protected to the degree that we couldn’t tell it was him.
There’s probably a math/compsci/physics theory that covers this, but there’s no way for Scott to be able to maintain both his pseudonymity and his identity – to the degree he wants (or needs) – if his (full) legal name is linked to his pseudonym on the first page of web search results for the legal name.
The safer option would be for Scott to create a new pseudonym and try to maintain (very) plausible deniability to any connection to Scott Alexander. But that would mean starting over and would probably be very hard – and maybe impossible anyways. It might be impossible for anyone to write in a different enough style, consistently, and avoid detection, especially at the lengths he typically posts. It’s probably even that much harder (or impossible) for people that have written as much as he has publicly.
I’ve seen some presentations about how to do style-matching off of GitHub repos to pretty-confidently ID anonymous coders. While set-up requires a sizable amount of compute and data, the results have gotten quite impressive. There are ways to work against this (stuff that deliberately obscures your coding style, usually by rewriting your code), but they’re not that well known. And a similar thing can be done with writing style and writing samples.
Staying anonymous against high-effort attempts to discern your identity has gotten very hard, and is only likely to get harder.
At some point, all you can do is guard against the low-effort ones.