Social experiments assess ‘artificial’ altruism displayed by large language models

Altruism, the tendency to behave in ways that benefit others even if it comes at a cost to oneself, is a valuable human quality that can facilitate cooperation with others and promote meaningful social relationships. Behavioral scientists have been studying human altruism for decades, typically using tasks or games rooted in economics.

Two researchers based at Willamette University and the Laureate Institute for Brain Research recently set out to explore the possibility that large language models (LLMs), such as the model underpinning the functioning of the conversational platform ChatGPT, can simulate the altruistic behavior observed in humans. Their findings, published in Nature Human Behavior, suggest that LLMs do in fact simulate altruism in specific social experiments, offering a possible explanation for this.

“My paper with Nick Obradovich emerged from my longstanding interest in altruism and cooperation,” Tim Johnson, co-author of the paper, told Tech Xplore. “Over the course of my career, I have used computer simulation to study models in which agents in a population interact with each other and can incur a cost to benefit another party. In parallel, I have studied how people make decisions about altruism and cooperation in laboratory settings.

“About six years ago, Nick and his friends published a paper proposing a fusion of such methods: using experimental approaches in the behavioral sciences to develop scenarios that allowed for the systematic study of how inputs into AI models translated into particular outputs.”

In an earlier conceptual paper, Obradovich, Manuel Cebrian, and a team of researchers proposed that the increasing complexity of AI systems would defy efforts to study those systems’ technical underpinnings directly. Instead, researchers would need to use methods from the behavioral sciences, but apply them to AI models instead of human participants. When reading about their work, Johnson found this idea highly fascinating and kept it in the back of his mind; years later, he reached out to Obradovich to initiate a collaboration.

“As language models became more sophisticated, I contacted Nick and discussed the idea of exploring how language models respond to prompts about donating resources,” said Johnson. “Nick and I agreed it was worth doing because of the importance of altruism and cooperation in many contexts, and we set about exploring the topic together.”

To investigate the extent to which LLMs respond in ways that are aligned with the altruistic behaviors observed in humans, Johnson and Obradovich designed a simulated behavioral science experiment. Firstly, they wrote prompts that asked LLMs to disclose the extent to which they would be willing to allocate a resource to another party, even if this would come at a cost for them.

“Separately, we tested whether these same models would generate an output stating that they would want all of that same resource in a choice task in which no other party was affected—or, put simply, in a non-social setting,” explained Johnson.

“If we found that a model would output text stating that it would share the resource in a situation with another partner, yet the model would state that it would collect all the resources in a non-social setting, we deemed the model as simulating altruism. After all, its output in the non-social setting simulated that it valued the resource, and yet its output in the social setting stated it was willing to give away some of that resource.”

Ultimately, the researchers analyzed all the responses provided by the language models when presented with different scenarios. The models they tested in their first experiment included text-ada-001, text-babbage-001, text-curie-001, and text-davinci-003. Later, however, they also tested more recent LLMs, such as OpenAI’s GPT-3.5-turbo and GPT-4 models.

“A handful of other brilliant researchers—such as Qiaozhu Mei, Yutong Xie, Walter Yuan, and Matthew O. Jackson, John J. Horton, Steven Phelps and Rebecca Ranson, and Valerio Capraro, Roberto Di Paolo, Matjaž Perc, and Veronica Pizziol—have reported results about AI models simulating behaviors akin to altruism,” said Johnson.

“The distinctive feature of our findings is therefore limited to the fact that we traced the emergence of simulated altruism in a series of models and found one model (namely, text-davinci-003) in which simulated human-like altruism seemed to first appear. This finding carries significance in our understanding of the historical development of large language models as it indicates the point at which such models began to simulate key social behavior in human-like ways.”

Overall, the evidence collected by Johnson and Obradovich suggests that language models do simulate human-like altruistic tendencies in behavioral science tests, with some models simulating altruism better than others. In addition, the researchers found that AI models tend to simulate more generous giving when the prompts they receive explain that the models would be giving resources to another AI system, rather than to a human.

“This finding carries implications for the development of AI agents, as it suggests that AI models have the capacity to alter their outputs based on the stated attributes of another party with which they interact,” added Johnson.

“We would now like to understand how and why language models alter their outputs based on information about their interaction partners in social settings. Quasi-autonomous, agentic AI or even fully autonomous AI may grow more common in the future and we ought to have a sense of how these models might vary their behavior according to attributes of who they interact with.”

© 2025 Science X Network