3 ms·
Policing at scale is difficult, as YouTube and Facebook have learned. Small communities are far easier to police. But to do it pre-emptively with AI is an open
by EGreg 7y ago
Policing at scale is difficult, as YouTube and Facebook have learned.
Small communities are far easier to police. But to do it pre-emptively with AI is an open problem. My interest is how can we make a system that will reliably converge in most cases to what’s true, like Wikipedia? It would be GREAT to have it for making sense collaboratively of news, politics, religion. People submit claims/evidence and each one has its own page.
https://en.m.wikipedia.org/wiki/Wikipedia:Wikiality_and_Other_Tripling_Elephants https://en.m.wikipedia.org/wiki/Wikipedia:Wikiality_and_Othe...
What is the incentive model? One I thought of is that you gain reputation points linearly but lose them exponentially in high-stakes challenges. The question is who can resolve a challenge one way or the other. And whether the result is immutable or can swing back if it turns out eg that the challenge was swayed by a coordinated effort.
If it’s a static system there are always going to be increasingly sophisticated attacks built on coordination (sybil attacks, sleepers). The most basic attacks require a net gain in credits to be sustainable. Namely, coordinated attacks (marauders) have to win challenges (and net positive points). But later on, attacks can be paid for by trading outside currency (eg BTC or USD) in exchange for using internal credits (“aged accounts”) in “vandalism” actions (attacks which do not gain credits but may even lose them, but are sustained by outside gain).
This is an issue for the source of truth for cryptocurrency as well, but at least the data is very well defined, and there doesn’t seem to be a huge gain from vandalism (except shorting). National elections in the US also have similar things for ballot counting. On Amazon and the sites I’m talking about, we are referencing outside events instead of internal processes. Wikipedia seems to have the most successful model w human editors of various kinds watching over the result.
Is there published research on the game-theoretic aspects of maintaining integrity of collaboratively edited information about the external world?