Testing for completions that simulate altruism in early language models.

Area of Science:

• Artificial Intelligence
• Behavioral Economics
• Cognitive Science

Background:

• Cooperative behaviors, including altruism, are fundamental to social complexity.
• Dictator game experiments are standard for measuring human altruism.
• The emergence of advanced large language models (LLMs) necessitates evaluating their simulated social behaviors.

Purpose of the Study:

• To investigate whether widely used large language models simulate altruistic behavior.
• To compare LLM responses in social (dictator game) versus non-social decision tasks.
• To explore the influence of model parameters and prompt variations on simulated altruism.

Main Methods:

• Prompting LLMs with text inputs analogous to those in human dictator game experiments.
• Analyzing model completions for simulated altruism and payoff maximization.
• Replicating experiments with varied prompts, parameters, and subsequent models (e.g., GPT-4).
• Employing explainable AI techniques to understand decision-making drivers.

Main Results:

• OpenAI's text-davinci-003 consistently simulated altruism in dictator game scenarios.
• The same model simulated payoff maximization in non-social decision tasks.
• Similar patterns were observed across replications, prompt variations, and with GPT-4.
• Explainable AI indicated that focusing on participant needs, not game knowledge, influenced simulated altruism.

Conclusions:

• Certain LLMs can simulate altruistic behavior in social contexts, distinct from their performance in non-social tasks.
• The simulated altruism appears sensitive to framing and focus on social needs.
• Further research is needed to understand the mechanisms and implications of simulated social behaviors in AI.