Emerging forms of human–AI collaboration enable AI models to act as surrogates for human decision-makers, raising important questions about the attribution of responsibility. While prior research shows that human decision-makers use AI to mitigate responsibility for harmful outcomes, little is known about the role of intentions in responsibility attribution of human-AI decision-making. In an online experiment using the “trembling hand” game (Cushman et al., 2009), we examine participants' punishment behavior across types of decision-maker (human vs. AI), intentions (good vs. bad), and outcomes (fair vs. unfair). Preliminary results indicate that punishment is primarily driven by outcomes, while intentions have only a marginal effect on human decision-makers. For unintentional outcomes, the type of decision-maker does not significantly influence punishment. This study contributes to the research on moral judgment of human-AI collaboration by analyzing the relative roles of intentions and outcomes in responsibility attribution to AI.