Large Language Models and Empathy: Systematic Review.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Computational Linguistics

Background:

• Empathy is crucial in human interaction and healthcare, traditionally considered a uniquely human trait.
• Large language models (LLMs) are advanced AI systems whose capacity for empathy has been questioned.

Purpose of the Study:

• To review existing literature on the ability of large language models (LLMs) to demonstrate empathy.
• To synthesize findings on LLM performance in empathy-related tasks across various contexts.

Main Methods:

• A comprehensive literature search was conducted across multiple academic databases (MEDLINE, Google Scholar, PsyArXiv, medRxiv, arXiv) from December 2022 to February 2024.
• Included were English-language full-length publications evaluating empathy in LLM outputs, excluding studies not specifically focused on empathy.

Main Results:

• Twelve studies from 2023 met the criteria, primarily evaluating ChatGPT-3.5 and comparing it with other LLMs like GPT-4 and LLaMA.
• LLMs demonstrated elements of empathy, including emotion recognition and emotional support, with some outperforming humans in specific tasks.
• Evaluation methods varied, including automatic metrics (e.g., BLEU, ROGUE) and human subjective evaluations, with noted limitations such as repetitive phrasing and prompt sensitivity.

Conclusions:

• LLMs exhibit cognitive empathy, recognizing emotions and providing supportive responses, moving them closer to human-like interaction.
• Advancements in LLM empathy expand their potential applications in areas requiring emotional intelligence.
• Further improvements in LLM performance and evaluation strategies for soft skills are necessary.