Unbiased: average w_MLE over many hypothetical datasets, and it equals the true w. Linear regression's w_MLE is unbiased.

But unbiased isn't everything. An unbiased estimator can still have huge variance — very different answers from dataset to dataset, even though they average out correctly. This is the bias-variance tradeoff, and it's why we're about to accept a little bias in exchange for a lot less variance.