Magnetic resonance imaging (MRI) has been extensively used to identify anatomical and functional differences in Autism Spectrum Disorder (ASD). However, many of these findings have proven difficult to replicate. One main reason is the reliance on small cohorts, insufficient to provide robust estimations. Another is the complexity of methodologies used, and the opacity of the lack of disclosure of the many analytical choices that lead to the different results. Thanks to open data initiatives, we were able to analyse with a homogeneous methodology a sample comparable in size to all the previous literature. Our analyses failed to reproduce the differences in brain volume, corpus callosum size, or cerebellar volume reported by small sample size studies, and suggested the presence of publication bias for statistically significant results. Does this mean that MRI cannot be used to detect ASD? On the contrary: I will suggest that MRI could become our best tool for studying ASD, given that the research community agrees to collect the large datasets required, and to make them openly accessible, so that they can be analysed collaboratively. I will present the results of IMPAC, the first international challenge to test the ability of MRI to predict ASD diagnosis. We found conclusive evidence of the utility of MRI: the best algorithms reliably predicted diagnostic with a sensitivity of 85% for a specificity of 53% (AUC~0.76). The accuracy achieved by MRI biomarkers is far superior to that obtained using, for example, genotyping data in cohorts 20-times larger. Appropriately sized cohorts should lead to improved performance and increased spatial localisation, revealing more precise neural correlates of ASD.