Water freezing is ubiquitous and affects areas as diverse as climate, the chemical industry, cryobiology and materials science. Ice nucleation is the controlling step in water freezing1-5 and has, for nearly a century, been assumed to require the formation of a critical ice nucleus6-10. But there has been no direct experimental evidence for the existence of such a nucleus, owing to its transient and nanoscale nature6,7. Here we report ice nucleation in water droplets containing graphene oxide nanosheets of controlled sizes and show that they have a notable impact on ice nucleation only above a certain size that varies with the degree of supercooling of the droplets. We infer from our experimental data and theoretical calculations that the critical size of the graphene oxide reflects the size of the critical ice nucleus, which in the case of sufficiently large graphene oxides sits on their surface and gives rise to ice formation behaviour consistent with classical nucleation theory. By contrast, when the graphene oxide size is smaller than that of the critical ice nucleus, pinning at the periphery of the graphene oxide deforms the ice nucleus as it grows. This gives rise to a much higher free-energy barrier for nucleation and suppresses the promoting effect of the graphene oxide11. The results provide experimental information on the existence and temperature-dependent size of the critical ice nucleus, which has previously only been explored theoretically and through simulations12-16. As pinning of a pre-critical nucleus at a nanoparticle edge is not specific to the ice nucleus on graphene oxides, we expect that our approach could be extended to probe the critical nuclei in other nucleation processes.