An accurate knowledge of Internet's topology is relevant to improve its efficiency and robustness. Internet's mapping projects are based on collecting and merging several shortest paths from a set of sources to a set of destinations. I will show that these processes may generate severe sampling biases depending on the topological properties of the underlying network. The results provide some qualitative criterion useful in the analysis of experimental data. Finally, I will discuss possible techniques in order to correct sampling biases in network inference.