“Using AI models like Claude for automated interpretability often results in descriptions that are generally true but lack the specific nuance of the feature being analyzed.”