Frey, S., Bos, M.W., and Sumner, R.W (2017) “Can you moderate an unreadable message? ‘Blind’ content moderation via human computation” Human Computation 4:1:78–106. DOI: 10.15346/hc.v4i1.5
Open access (free) here.
What’s it about?
Say I’m the mailman and you just received a letter, and you wanted to know before opening it if it has anything disturbing. You could ask me to invade your privacy and open it. Or I could respect your privacy and make you take a chance. But I can’t do both. In this sense, safety and privacy are opposed. Or are they? In certain decision settings, its possible to filter out unsafe letters without opening any of them.
In this project, I lay out two tricks I developed for determining without looking at a piece of content whether it contains inappropriate content. This is important because most kids are on the Internet. In fact, according to some reports, a third of all cell phones are owned by minors.
One of the two methods could one day work for protecting voters from intimidation, by replacing normal checkboxes on a ballot with low-resolution pictures of two generic faces. Here’s the basic idea. You have a tyrant and an upstart competing for the tyrant’s seat. Everyone wants to vote for the upstart, but everyone is afraid that the tyrant will read their ballot and seek retribution. Assume the big assumption that the winner will get to take office and there’s protection from voting fraud and all that stuff, and just focus on the mechanics of the ballot.
In my scheme, your ballot doesn’t actually name any candidate. All there is are two copies of the same generic faces, both sort of fuzzed up with noise like the snow on a TV. By chance, because of the noise, one face will barely looks slightly more like the candidate you prefer. To vote, all you do is circle that face. Every person gets a ballot with the same face, but different noise. Then after all the ballots are collected, you take all the faces that got circled, average them, and the generic face plus the averaged noise will look like the face of the upstart. But from each individual ballot it’ll be impossible for the tyrant to know who you voted for. This averaging method, called reverse correlation in social psychology, has already been shown to do all kinds of cool stuff. But never anything vaguely useful before. That’s why this paper could be considered a contribution.
I’m proud of this paper, and with how quickly it came out: just, umm, three years. Quick for me.