I think people are hesitant to call ML “statistical modeling” because traditional statistical models approximate the underlying phenomena; e.g., a logarithmic regression would only be used to study logarithmic phenomena. ML models, by contrast, seldom resemble what they’re actually modeling.