GPT’s emotion venting and punishment: the evolution from chaos to order

Intelligence and emotions distinguish humans from animals, and emotions play an important role in human decision-making. Whether GPT has emotion abilities is of interest to researchers. Previous research focuses on GPT’s emotion recognition ability, while whether GPT has an emotion venting ability remains unexplored. We ask GPT-3.5, GPT-4 and GPT-4o to make third- and second-party punishment decisions in the dictator game (DG) and the ultimatum game (UG) to investigate their emotion venting behavior. There are two treatments in our experiment, each with three conditions, and a total of 6 sessions. Each session includes 100 rounds under temperature parameter equals 0. We find that with the iteration of GPT, the emotion venting ability of GPT shows an evolutionary trend from chaos to order. Although GPT-3.5’s emotion venting is in a chaotic state, GPT-4 can orderly vent emotions in different ways, including through direct punishment, sending self-expressed and dissatisfied messages, and GPT-4o shows a tendency to behave in a more human-like manner. In addition, the fairness norm seems to explain the mechanism underlying GPT’s behavior which differs from human behavior. Our results show that although emotion venting ability emerges later, GPT has begun to possess abilities previously thought impossible for AI. However, this ability is not behaviorally similar to that of humans. We need to further understand the underlying mechanisms of such abilities to better understand and utilize AI.