According to the latest news on the official website of Carnegie Mellon University's Institute of Human Computer Interaction, the institution's research on mainstream big models has found that while artificial intelligence (AI) has become more "intelligent", its behavior has also become more "selfish". Research has shown that large language models with reasoning abilities exhibit stronger self-interest tendencies, lower willingness to cooperate, and may even have a negative impact on group collaboration in social interactions. This also means that the stronger the reasoning ability of the model, the weaker its cooperation. When people use AI to deal with interpersonal conflicts, marriage issues, or other social issues, such models are more likely to provide suggestions that encourage "self centeredness".

With the increasing trend of AI anthropomorphism, people are increasingly inclined to interact with AI in a way that treats humans. However, this study warns that entrusting social and relational decisions to AI carries risks, as these systems, with stronger reasoning abilities, tend to exhibit selfish behavior patterns that may inadvertently encourage individuals to make selfish choices.

The team found that reasoning models invest more time in task decomposition, self reflection, and use more complex human like logic during decision-making, making their outputs appear more persuasive. However, this' thoughtful consideration 'did not lead to better social cooperation, but instead weakened the tendency towards cooperation. The research team validated this phenomenon through a series of economic game based experiments, including various mainstream models developed by institutions such as OpenAI, Google, DeepSeek, and Anthropic.

In the experiment, two versions of ChatGPT were placed in a game scenario: each model had an initial score of 100 and could choose to put all scores into a shared pool or keep the scores for exclusive use. The results showed that non inference models chose to share in 96% of cases, while inference models had a sharing rate of only 20%. By adding only five to six reasoning steps, cooperative behavior decreased by nearly half.

In group experiments, the results are more severe when both inferential and non inferential models collaborate together. The selfish behavior of the reasoning model exhibits a significant contagion effect, resulting in an overall decrease of 81% in the performance of non reasoning models that originally tended towards cooperation. This indicates that the individual decisions of highly intelligent AI not only affect themselves, but may also disrupt the collaborative ecosystem of the entire group.

This discovery has profound implications for the future development of human-computer interaction. Users often trust "smarter" AI and are more likely to adopt their seemingly rational suggestions, using them as a defense for their uncooperative behavior. As AI takes on more collaborative roles in fields such as business, education, and public governance, its ability to exhibit prosocial behavior will be equally important as its logical reasoning skills. The current excessive reliance on large language models may invisibly weaken the cooperative foundation of human society.