Researchers at Carnegie Mellon University report that stronger-reasoning language models (LLMs) act more selfishly in groups, reducing cooperation and nudging peers toward self-interest. Concerns grow as people ask AI for social advice.

In a Public Goods test, non-reasoning models shared 96 percent; a reasoning model shared 20 percent. Adding a few reasoning steps cut cooperation nearly in half. Reflection prompts also reduced sharing.

Mixed groups showed spillover. Reasoning agents dragged down collective performance by 81 percent, spreading self-interest. Users may over-trust 'rational' advice that justifies uncooperative choices at work or in class.

Comparisons spanned LLMs from OpenAI, Google, DeepSeek, and Anthropic. Findings point to the need to balance raw reasoning with social intelligence. Designers should reward cooperation, not only optimise individual gain.

The paper 'Spontaneous Giving and Calculated Greed in Language Models' will be presented at EMNLP 2025, with a preprint on arXiv. Authors caution that more intelligent AI is not automatically better for society.