The labeling algorithm is just a relatively naive example. More complex "neutral" algorithms can be designed. Mitigations for abuse that are "neutral" themselves may be added before/after initial implementation. (e.g To extend the labeling example, a moderation system could be added that allows users to flag inappropriate labels.) Also, I don't think it is always possible to foresee in what ways a system/algorithm might be abused; therefore, it might be necessary to address said abuse once it reveals itself.