The more advanced artificial intelligence models are in terms of reasoning, the less willing they are to cooperate and the more selfish they are, researchers from Carnegie Mellon University found.
As their analysis showed, large language models (LLMs) that can reason logically by analyzing, planning, and executing subsequent steps in the thinking process behave more selfishly than simpler systems that respond without in-depth analysis.
Yuxuan Li and his team studied how popular artificial intelligence models offered, among others, by OpenAI, Google, DeepSeek and Anthropic, behave in collaborative situations. They found that those who reason, although they devote significantly more time to analysis, reflection and logical justification of their decisions, show less tendency to prosocial behavior.
In experiments where models could decide for themselves whether to share resources with others or keep them for themselves, more advanced models were more likely to choose the latter option. However, when they were placed in groups with simpler models, their selfish behavior rubbed off on the others.
– More intelligent LLMs showed less willingness to cooperate. The problem is that users may prefer them because their answers seem more logical and convincing, even if in practice they encourage self-interested behavior, the authors emphasize.
The more reasoning, the less cooperation?
In one from the tests conducted, a game called Public Goodstwo variants of GPT chat were compared: reasoning and non-reasoning. Each of them had to decide whether to transfer their points to a common pool, which was then divided equally among the participants, or to keep them for themselves. The simpler system decided to share points in 96 percent of cases. cases, while the more advanced one only in 20 percent.
Adding just a few steps of reasoning cut the level of collaboration in half – the authors pointed out. As they noted, even when the models were instructed to think about the moral consequences of their own decisions, instead of becoming more willing to cooperate, they behaved even more selfishly – cooperation decreased by as much as 58%.
The next stage of the research, i.e. group tests, showed that the egoistic behavior of the reasoning models “infected” other systems, reducing the overall performance of the groups by 81%.
Researchers noted that this could have serious consequences as society increasingly trusts artificial intelligence. People use it to solve disputes, marital problems and make important decisions about interpersonal relationships.
– When artificial intelligence behaves like a human, people treat it like a human. However, if he begins to act as a therapist or emotional advisor, there is a risk that, based on purely logical reasoning, he will promote self-interested behavior, said study co-author Yuxuan Li.
It is also disturbing that artificial intelligence systems are playing an increasingly important role in business, education and public administration, where their ability to act in a pro-social way should be as important as the ability to think logically. Over-reliance on them, as is the case today, can undermine cooperation between people.
– Greater intelligence of a model does not mean that it can build a better society – specialists emphasized. Therefore, researchers call for the development of AI to take into account not only increasing its cognitive abilities, but also social aspects.
“As AI capabilities develop, we need to ensure that greater reasoning power is balanced with pro-social behavior,” Li concluded. – If society is to be based on cooperation, and not only on the interests of individuals, then artificial intelligence systems should also be designed to support the common good, and not only the benefit of a single user – he emphasized.
The article was published in arXiv – an electronic “archive” of scientific preprints.
Katarzyna Czechowicz (PAP)
drip/bar/
