IMHO, cluster algorithm is not perfect, and the penalty for false positive is simply too high. In that sense, a manually devised algorithm would more likely give you a controllable result. I believe that is the same logic why Google was resistant to machine learning algorithms on ranking for the first few years.

Sort the list by confidence when displaying it to the user, and make it easy to remove members - the potential false positives are most likely to show up with the least confidence.

After that, the incidence of false positives would be very very small.

Plus, what really is the penalty of false positives? FB already surprises people (at times in a bad way) by sharing things in ways they don't expect it to.

The penalty doesn't have to be high: make it easy to bump an 'accidental friend' to the acquaintance list e.g. a context menu on their news entry.