Moving least squares (MLS) are used as promising approximation in meshfree methods. However, known drawbacks of MLS, such as the accuracy sensitivity to the support size of the weight function and to the steep changes of the node distribution, have not yet been discussed sufficiently. In this paper, these problems are spotlighted from a standpoint of the improvement of the solution reliability and the computational efficiency on MLS. Two questions are submitted against the well-known advantages of MLS. A computational bottleneck of MLS, which is setting of the support size corresponding to the node distribution, is also presented concretely.