We take cohorts of entering freshmen at the United States Air Force Academy and assign half to peer groups designed to maximize the academic performance of the lowest ability students. Our assignment algorithm uses nonlinear peer effects estimates from the historical pre‐treatment data, in which students were randomly assigned to peer groups. We find a negative and significant treatment effect for the students we intended to help. We provide evidence that within our “optimally” designed peer groups, students avoided the peers with whom we intended them to interact and instead formed more homogeneous subgroups. These results illustrate how policies that manipulate peer groups for a desired social outcome can be confounded by changes in the endogenous patterns of social interactions within the group.