Behavioral Economics and ChatGPT: From William Shakespeare to Elena Ferrante

We prompted GPT-4 (an artificial intelligence large language model) to use literary fictional characters to play the Dictator game, a classic behavioral economics experiment. We prompted GPT-4 with 148 characters from the 17th century to the 21st century. Their 888 decisions were used to compare characters over time as well as characters to human players. There is a general and mainly monotonic decrease in selfish behavior over time in literary characters. 41% of the decisions of 17th century characters are selfish compared to just 21% of the decisions of characters from the 21st century. In the Human-AI comparison, Humans are much more selfish than the AI characters with 51% of humans making selfish decisions compared to 28% of the AI characters. 

Here is the full (short) paper by Gabriel Abrams (a junior in high school).