What is the main statistical role of cross-fitting?

In double machine learning, nuisance functions are trained on K−1 folds and evaluated through an orthogonal score on the remaining fold. What is the main purpose of this cross-fitting procedure?