Maybe binary content propagates differently. Text that meets certain criteria is replicated indiscriminately, but binary content is only replicated when a user votes on it.
Edit: you could apply game theory to this problem. Model the network as a graph and write an agent-based modeling rule set for... say... CP-propagators and non-CP-propagators. Run iterative simulations of different propagation rule sets and weightings/parameters. Now introduce bad actors in the form of, say, government agents trying to suppress political discourse. The difference is that average-joe will cooperate in pushing out CP but will "defect" in a game with the other kind of bad-actor. You're looking for rule-sets and parameters where the CP gets pushed to the margins of the network or excluded but where the other kind of bad-actor is also excluded.
Could Bayesian classification be implemented through a homomorphic cypher?