Graph neural network outputs converge to asymptotically constant functions on random graphs, limiting their expressive power.
