Why does a least squares fit appear to have a bias when applied to simple data?

Why does a linear least squares fit appear to have a bias when applied to simple test data?