We consider graph neural networks with mean aggregation and non-negative weights, showing which monotonic rules they can learn and constructing sound explanatory rules.