Sometimes people have something they want to say without it being
traceable back to their main identity, and on internet forums it's
common for people to use multiple accounts ("alts") for this. As
machine learning software gets better, however, it becomes
increasingly practical to link a person's accounts.
A few months ago someone ran a simplestylometry tool across Hacker
News comments, and identified many alts, including one for the
site's founder. To further demonstrate that this isn't just an
academic consideration, I recently did the same for LessWrong and the
EA Forum. I'm not going to share the code or the probabilities it
generated, and I've only looked at the output enough to be pretty sure
it's working. Trained on half of the data and tested on the other
half it was consistently able to link accounts, however, and it was
also able to identify at least one non-public alt account I already
knew about.
This is an example of a general problem with privacy: even if
something seems careful enough now, you can't trust the future to
keep things private.
(If you do want somewhat more protection now, however, I think best
practice is running your alt comments through an LLM to change the
style.)
I'm pretty okay with this having been done - seems broadly consistent with cyber disclosure norms - we know that there are people who have disputes with this community and a track record of looking to de-anonymize accounts, so relying on security through obscurity doesn't seem that compelling - seems reasonable to view this post as helpful to would-be anonymous posters - better to find out about the possibility this way than to be surprised by someone publicly attempting to deanonymize you in future, and the suggestion of using an LLM to change the style seems constructive
Sometimes people have something they want to say without it being traceable back to their main identity, and on internet forums it's common for people to use multiple accounts ("alts") for this. As machine learning software gets better, however, it becomes increasingly practical to link a person's accounts.
A few months ago someone ran a simple stylometry tool across Hacker News comments, and identified many alts, including one for the site's founder. To further demonstrate that this isn't just an academic consideration, I recently did the same for LessWrong and the EA Forum. I'm not going to share the code or the probabilities it generated, and I've only looked at the output enough to be pretty sure it's working. Trained on half of the data and tested on the other half it was consistently able to link accounts, however, and it was also able to identify at least one non-public alt account I already knew about.
This is an example of a general problem with privacy: even if something seems careful enough now, you can't trust the future to keep things private.
(If you do want somewhat more protection now, however, I think best practice is running your alt comments through an LLM to change the style.)
Comment via: facebook, mastodon