LLM Emotional Alignment Varies by Model and Demographic Group

Evaluation of seven LLMs against 3,011 human participants shows emotional alignment for news-based sympathetic framing varies widely. GPT-5.2 achieved a high correlation of 0.789 with human judgment, while Mistral Large 2512 reached a medium 0.4, though leading models remained broadly aligned across demographic subgroups.