We study how social interaction and friendship shape students' political opinions in a natural experiment at Sciences Po, the cradle of top French politicians. We exploit arbitrary assignments of students into short-term integration groups before their scholar cursus, and use the pairwise indicator of same-group membership as instrumental variable for friendship. After six months, friendship causes a reduction of differences in opinions by one third of the standard deviation of opinion gap. The evidence is consistent with a homophily-enforced mechanism, by which friendship causes initially politically-similar students to join political associations together, which reinforces their political similarity, without exercising an effect on initially politically-dissimilar pairs. Friendship affects opinion gaps by reducing divergence, therefore polarization and extremism, without forcing individuals' views to converge. Network characteristics also matter to the friendship effect.