A company wants to detect harmful or toxic language in social media comments but will not use labeled data to train a model. Which approach should the company use to identify toxic content?

Amazon Comprehend toxicity detection is the correct choice because it is a pre-trained natural language processing (NLP) service specifically designed to identify harmful or toxic language in text. Since the company does not want to use labeled data for training, a pre-trained service is ideal. Amazon Rekognition moderation is for image and video analysis, not text. Amazon SageMaker built-in algorithms would require labeled data for training a custom model, which contradicts the problem statement. Amazon Polly is a text-to-speech service and has no functionality for content moderation or toxicity detection.