This chapter evaluates transfer learning approaches in three experiments for detecting harmful content. The first focuses on fine-tuning vs. feature extraction for hate speech detection using models pretrained models (i.e., LLMs) like BERT, RoBERTa, and XLNet in Sect. 4.2. The second explores the performance of different pretrained models in feature extraction settings in Sect. 4.3. Results suggest that domain-aware configurations, like leveraging multiple pretrained layers, significantly improve results. The third experiment frames claim extraction as a sequence-to-sequence task, testing T5 and PEGASUS in zero-shot, fine-tuned, extractive, and query-based summarization settings (Sect. 4.4). Fine-tuned abstractive models perform best, indicating that output-generating models benefit more from domain-specific training. Overall, the chapter highlights the adaptability of TL for text classification and generation tasks and its power in data-scarce contexts.