Evaluating SAE interpretability without explanations