The dynamics of evolution depend on an organism’s fitness landscape: the mapping from genotypes to reproductive capacity. Knowledge of the fitness landscape can help resolve questions, such as how quickly a pathogen will acquire drug resistance or by what pattern of mutations. However, direct measurement of a fitness landscape is impossible because of the vast number of genotypes. Here, we critically examine regression techniques used to approximate fitness landscapes from data. We find that such regressions are subject to two inherent biases that distort the biological quantities of greatest interest, often making evolution seem less predictable than it actually is. We discuss methods that may mitigate these biases in some cases.