Journal article: Observing that most social networks are clustered, the literature often argues that agents are more willing to form links that close triangles. We challenge this idea by proposing a simple model of new collaboration formation that shows why network clustering may arise even though agents do not “like” network closure. We address empirically this question on the longitudinal evolution of the French co-invention network, and find that two inventors are less likely to form a first research collaboration when they have common partners. Our findings further reveal the preferences of inventors towards forming non-redundant connections.