On the underestimation of model uncertainty by Bayesian K-nearest neighbors