We performed GWAS on 2514 complex traits from the UK Biobank using a linear mixed model, identifying 40,620 independent significant associations (p−8). We estimate that winner’s curse incurs substantial overestimation of effect sizes in a mean of 35% of discovered associations per trait. We use these results to estimate that the polygenicity of most complex traits is below 10000 common causal variants. We evaluated the impact of winner’s curse on causal effect estimation and hypothesis testing in Mendelian randomization analyses. We show that winner’s curse substantially amplifies the magnitude of weak instrument bias, though any inflation of false discovery rates tends to be low or modest. We designed a process of pseudo-replication within the UK Biobank data to generate GWAS estimates that minimise bias in MR studies using these data. Our resource is integrated into the OpenGWAS platform and enables a convenient framework for researchers to minimise bias or maximise precision of causal effect estimates.