This research examines the potential of datasets generated by Large Language Models (LLMs) to support Natural Language Processing (NLP) tasks, aiming to overcome challenges related to data acquisition and privacy concerns associated with real-world data. Focusing on negative valence text, a critical component of sentiment analysis, we explore the use of LLM-generated synthetic news headlines as an alternative to real-world data. A specialized corpus of negative news headlines was created using tailored prompts to capture diverse negative sentiments across various societal domains. The synthetic headlines were validated by expert review and further analyzed in embedding space to assess their alignment with real-world negative news in terms of content, tone, length, and style. Key metrics such as correlation with real headlines, perplexity, coherence, and realism were evaluated. The synthetic dataset was benchmarked against two sets of real news headlines using evaluations including the Comparative Perplexity Test, Comparative Readability Test, Comparative POS Profiling, BERTScore, and Comparative Semantic Similarity. Results show the generated headlines match real headlines with the only marked divergence being in the proper noun score of the POS profile test.