How Large Language Models Handle the Trade-off Between Fairness and Efficiency: CogSci 2026 Study Reveals Similarities and Differences Between LLM and Human Decision-Making

A study to be published at CogSci 2026 systematically compares the decision-making patterns of large language models (LLMs) and humans in the trade-off between fairness and efficiency through task allocation scenarios, revealing their similarities and differences and providing an empirical basis for understanding the social preferences of AI systems. The core of the study is to explore whether the choice patterns of LLMs when facing conflicts between fairness and efficiency are similar to those of humans or exhibit unique preferences.

In real-world scenarios such as resource allocation and task scheduling, decision-makers often face difficult trade-offs between fairness and efficiency: complete equality may reduce overall output, while pursuing maximum efficiency may lead to excessive individual burdens. This trade-off exists not only in human policy-making but also in AI automated decision-making. With the widespread application of LLMs in decision support systems, a key question emerges: Are the choice patterns of AI in fairness-efficiency conflicts similar to those of humans?

The study uses classic experimental paradigms from behavioral economics and cognitive psychology to ensure that humans and LLMs receive consistent information (task descriptions, option presentations, outcome feedback) for direct comparison. Humans completed tasks via an online platform, while LLMs generated decisions through the same text prompts. Multiple prompt variants were tested to evaluate decision robustness, and confounding variables were controlled to ensure the comparability and ecological validity of the results.

Similarities: LLMs are sensitive to fairness and consider efficiency—they are not purely instrumental rationalists. They tend to reject highly efficient but obviously unfair solutions, which echoes findings from human behavioral economics. Differences: LLM decisions are more consistent and stable, not affected by framing effects, emotions, or fatigue; human choices have large intra- and inter-individual variations; LLMs have a lower tolerance for extreme unfairness—humans may accept a certain degree of inequality in exchange for efficiency improvements, while LLMs have a stronger fairness bottom line.

Theoretical Significance: LLMs are not value-neutral tools; they embed specific social preferences and moral judgments. Understanding their origins and characteristics is crucial for the responsible deployment of AI. Practical Implications: Developers need to carefully design LLM decision systems to ensure that fairness concepts are aligned with humans (e.g., in high-risk fields like healthcare and education); algorithmic audits are needed to evaluate whether the system's behavioral patterns in fairness-efficiency scenarios meet ethical standards and social expectations.

Limitations: The experiment is a simplified laboratory environment, which differs from complex real-world decision-making; it only focuses on single decisions and does not examine dynamic fairness in repeated interactions. Future Directions: Compare differences between LLMs with different architectures/training methods; explore fine-tuning or prompt engineering to adjust LLM social preferences; study human perception and response to AI's fair decisions and their impact on human-AI collaboration.

This study opens a new window for understanding the social decision-making behavior of LLMs. It not only shows that LLMs have a 'fairness intuition' but also reveals subtle differences in value trade-offs with humans. As AI's role in decision-making scenarios increases, the value of such comparative studies becomes prominent. Only by deeply understanding AI's values and behavioral patterns can we ensure that it serves human well-being and avoids exacerbating inequality or harming fairness.