Replica exchange (RE) molecular dynamics (MD) simulations are frequently applied to sample the folding–unfolding equilibria of β-hairpin peptides in solution, because efficiency gains are expected from this technique. Using a three-state Markov model featuring key aspects of β-hairpin folding we show that RE simulations can be less efficient than conventional techniques. Furthermore we demonstrate that one is easily seduced to erroneously assign convergence to the RE sampling, because RE ensembles can rapidly reach long-lived stationary states. We conclude that typical REMD simulations covering a few tens of nanoseconds are by far too short for sufficient sampling of β-hairpin folding–unfolding equilibria.