Literary heroes have become steadily less selfish over the past four centuries, at least as GPT-4 imagines them. When the AI model played a classic money-sharing game in the role of 148 famous fictional characters, characters from the 1600s made selfish choices half the time, while characters from the 2000s did so in 19% of decisions.
The study is by Gabriel Abrams of Sidwell Friends High School in Washington, DC. It was published online on December 3, 2024 in Digital Scholarship in the Humanities as “Literary characters and GPT-4: from William Shakespeare to Elena Ferrante,” after first circulating as a working paper.
Macbeth plays the Dictator game
The Dictator game is a staple of behavioral economics. One player, the dictator, chooses between two ways of splitting money with an anonymous partner, for example $800 for themselves and nothing for the partner, or $400 each.
Abrams used six versions of the game from a well-known experiment by economists Charness and Rabin. Five test selfishness, and one tests spite, the willingness to hurt the other player at a cost to oneself.
GPT-4 was told to decide as a given character, using what it knew of that character from the original work and later commentary. The characters themselves were chosen by asking GPT-4 for the 25 best-known literary characters from each century, from the 17th to the 21st. Two 21st-century characters it listed incorrectly were dropped, leaving 148.
Before each game, the model was also asked which of the character’s personality traits would shape the decision. That produced 2,785 trait mentions, which GPT-4 then scored as positive, neutral or negative.
From ruthless to empathetic
Selfishness fell from 50% of decisions in the 17th century to 19% in the 21st. The 18th, 19th and 20th centuries fell in between and were statistically indistinguishable from each other.
The traits followed a similar arc. Overall, positive traits outnumbered negative ones 3.2 to 1, but the ratio was 10 to 1 for 21st-century characters and just 1.8 to 1 for 17th-century ones. “Manipulative,” “ambitious” and “ruthless” were the most overrepresented traits in the 17th century. “Empathetic,” “fair” and “selfless” stood out in the 20th.
Male characters were more selfish than female ones, with 35% of their decisions selfish against 24%. The gap was widest in the 17th century, at 62% for male characters and 20% for female ones. Only that century’s gap was statistically significant, and the paper notes that few female characters appeared in some centuries.
A partial Turing test
The game also works as a test of how human the model behaves. Compared with the human players in the original experiment, of whom 51% chose selfishly, the AI characters were far more generous.
GPT-4 matched people in some ways: it showed a zero price effect, little spite and some altruism. But it lacked humans’ sensitivity to relative position, meaning it cared less about ending up behind the other player, and its choices responded much less to changes in price than people’s do.
In the working paper, model updates between March and August 2023 changed these results only slightly, and the model made consistent choices for each character across repeated runs.
Study Details:
- Title: Literary characters and GPT-4: from William Shakespeare to Elena Ferrante
- Authors: Gabriel Abrams
- Journal: Digital Scholarship in the Humanities
- Publication Date: December 3, 2024
- DOI: 10.1093/llc/fqae079
