Preview

PDF, English - main document Download (1MB) | Terms of use

Abstract

We investigate if Large Language Models (LLMs) exhibit personality-driven strategic behavior in the Ultimatum Game by manipulating Dark Factor of Personality (D-Factor) profiles via standardized prompts. Across 400k decisions from 17 open-source models and 4,166 human benchmarks, we test whether LLMs playing the proposer and responder roles exhibit systematic behavioral shifts across five D-Factor levels (from least to most selfish). The proposer role exhibited strong monotonic declines in fair offers from 91% (D1) to 17% (D5), mirroring human patterns but with 34% steeper gradients, indicating hypersensitivity to personality prompts. Responders diverged sharply: where humans became more punitive at higher D-levels, LLMs maintained high acceptance rates (75-92%) with weak or reversed D-Factor sensitivity,failing to reproduce reciprocity-punishment dynamics. These role-specific patterns align with strong-weak situation accounts—personality matters when incentives are ambiguous (proposers) but is muted when contingent (responders). Cross-model heterogeneity was substantial: Models exhibiting the closest alignment with human behavior, according to composite similarity scores (integrating prosocial rates, D-Factor correlations, and odds ratios), were dolphin3, deepseek_1.5b, and llama3.2 (0.74-0.85), while others exhibited extreme or non-variable behavior. Temperature settings (0.2 vs. 0.8) exerted minimal influence. We interpret these patterns as prompt-driven regularities rather than genuine motivational processes, suggesting LLMs can approximate but not fully replicate human strategic behavior in social dilemmas.