With recent advances in AI, artificial intelligence, the market and researchers question the extent to which technology can reproduce characteristics similar to human behavior. In this context, researchers at Willamette University and the Laureate Institute for Brain Research investigated whether large language models (LLMs), like ChatGPT, would be able to simulate altruistic patterns.
The study, published in the magazine Nature Human Behavior, revealed that LLMs can reproduce altruism in specific social experiments, pointing out similarities with the decision-making processes characteristic of human beings. The experiments carried out showed that LLMs tend to work more when interacting with other AIs than with humans, adjusting your choices according to technology.
This difference in interaction suggests that models can adapt altruistic decisions to the social context.. The research result will be relevant for the development of autonomous agents, reinforcing that LLMs do not just process data, but they can also simulate patterns of social behavior, albeit with challenges.
How the study was carried out?
To assess the ability to simulate an altruistic pattern, researchers developed scenarios in which models needed to choose between keeping resources for themselves or allocating them to another agent. When a model chose to share part of the resource in a social environment, but kept everything to himself in a non-social context, This attitude was interpreted as a sign of altruistic behavior.
Among the AI systems analyzed were text-ada-001, o text-davinci-003, o GPT-3.5-turbo and o GPT-4. The results showed that the first manifestation of this pattern occurred in text-davinci-003, by OpenAI, becoming more evident and sophisticated in the company's recent models. Future research should explore in greater depth the mechanisms that support these patterns., mainly due to the growing presence of autonomous AIs in human and digital interactions.
