This paper proposes alternative parameter estimators for autoregressive models that differ from the least squares estimators. In unstable (or critical) cases, i.e., cases where the roots of the characteristic equation lie on the unit circle, the least squares estimators typically have complex asymptotic distributions. In contrast, the proposed non-standard estimators tend to have simpler asymptotic distributions in most critical cases.