随着人工智能的广泛应用，人们在面对冒犯时不仅需决定是否宽恕人类，也需决定是否宽恕机器人。两项实验考察了交互对象（人类 vs. 机器人）及其行为反应（无行为、道歉、幽默自贬、自我惩罚）对宽恕的影响及其心理机制。结果发现，人们更倾向于宽恕机器人，该效应通过更高的净积极情绪实现，而净收益也可能在其中发挥作用。值得注意的是，机器人采用自我惩罚策略在促进宽恕方面展现出一定的优势。计算建模作为补充分析发现，交互对象与行为反应会影响个体在宽恕中的决策偏好。研究结果初步验证了“认知-情感”双加工路径在人机互动中对宽恕决策的解释力，并为设计更具人际友好性的社交机器人提供了启示。

Forgiveness is essential for maintaining cooperation and resolving conflicts in social interactions. As artificial agents become more prevalent in interpersonal contexts, human-robot interactions (HRI) are becoming increasingly common. Nowadays, people are confronted with the decision to forgive not only human transgressors, but also robotic ones. To address this emerging social challenge, this research investigates how different post-offense behavioral responses from humans versus robots influence forgiveness, and explores the underlying psychological mechanisms through a dual-path framework that involves cognitive and affective factors. Across two experiments (each with a total sample size of N = 320), participants engaged in scenarios in which either a human or a robot committed an offense and subsequently exhibited one of four behavioral responses: no response, apology, self-deprecating humor, or self-punishment. Both experiments systematically manipulated agent type and behavioral response types.

In Experiment 1, the offense and behavioral response were presented textually. By employing a 2 (agent type: human vs. robot)×4 (behavioral response type: no response vs. apology vs. self-deprecating humor vs. self-punishment) between-subjects design, Experiment 1 revealed significant main effects of both the agent type and behavioral response type on forgiveness. Specifically, participants were more likely to forgive robots than humans. Confirming our mechanistic hypotheses, mediation analyses showed that perceived net benefit and net positive affect significantly mediated the effect of agent type on forgiveness, highlighting the role of cognitive and affective factors in shaping forgiveness.

Experiment 2 partially replicated and extended the findings of Experiment 1 using more ecologically valid video stimuli, which made the behavioral responses of human and robot offenders visible and thus more vivid and realistic. The main effects were validated once again. Mediation analysis revealed that people tend to forgive robots more readily, primarily due to higher levels of perceived net positive affect. Additionally, Experiment 2 revealed a significant interaction effect: the same behavioral response can have different effects on forgiveness depending on whether it is performed by a human or a robot. Specifically, compared to human, robots demonstrated a relative advantage in promoting forgiveness when adopting a self-punishment strategy. Building on this significant interaction, Experiment 2 further explored the psychological mechanisms through which agent type influenced forgiveness under different behavioral responses. Mediation analysis revealed that in the self-punishment condition, both net benefit and net positive affect significantly mediated the effect of agent type on forgiveness. These results suggest that when engaging in self-punishment, robots elicited greater perceptions of net benefit and net positive affect, thereby increasing their likelihood of being forgiven. Additionally, as a supplementary analysis, we conducted computational modeling. The results indicated that agent type and behavioral response type influenced individuals’ aversion to advantageous inequity, rather than altering their underlying decision algorithm in the context of forgiveness-related decisions.

The main novel contribution of this study lies in extending the scope of forgiveness research to include robots, an area that has been largely overlooked before. Most previous studies have focused on forgiveness toward humans, typically explaining it using the dual-path framework of cognition and affect. In response to the growing relevance of artificial intelligence in contemporary society, the present study examines forgiveness toward robots. The findings provide preliminary support for the applicability of the cognition-affect dual-process framework in explaining forgiveness toward robots, thereby extending its theoretical scope to encompass human-robot interactions.

In conclusion, this study revealed two key findings. First, in the context of social offenses, individuals exhibited greater forgiveness toward robots than toward humans - a pattern primarily mediated by perceived net positive affect, with perceived net benefit also potentially contributing. Second, robots demonstrated a relative advantage over humans in eliciting forgiveness when using a self-punishment strategy.