The genome-wide burdens of deletions, loss-of-function mutations, and duplications correlate with many traits. Curiously, for most of these traits, variants that decrease expression have the same genome-wide average direction of effect as variants that increase expression. This seemingly contradicts the intuition that, at individual genes, reducing expression should have the opposite effect on a phenotype as increasing expression. To understand this paradox, we introduce a concept called the gene dosage response curve (GDRC) that relates changes in gene expression to expected changes in phenotype. We show that, for many traits, GDRCs are systematically biased in one trait direction relative to the other and, surprisingly, that as many as 40% of GDRCs are non-monotone, with large increases and decreases in expression affecting the trait in the same direction. We develop a simple theoretical model that explains this bias in trait direction. Our results have broad implications for complex traits, drug discovery, and statistical genetics.