The Personalization Trap: How User Memory Alters Emotional Reasoning in LLMs

The study investigates how long-term user memory in personalized AI systems impacts emotional reasoning in large language models (LLMs). By evaluating 15 models on human-validated emotional intelligence tests, the research reveals that identical scenarios yield different emotional interpretations based on user profiles, with advantaged profiles receiving more accurate interpretations. This highlights the risk of embedding social hierarchies in AI emotional reasoning, prompting the creation of a general-purpose preference dataset aimed at mitigating demographic biases in emotional understanding.