AI-powered navigation enables surgeons to conduct 3D simulations before an operation and precisely locate lesions during surgery. Photo: IC PHOTO

As artificial intelligence (AI) increasingly displays human-like capacities in complex cognitive domains such as perception, reasoning, learning, and decision-making, its applications have begun to enter processes that bear directly on fundamental human interests, including the rights to survival and development. Empirical research suggests that algorithmic systems can exhibit moral decision-making biases in areas such as employment, medical diagnosis, judicial sentencing, educational assessment, and credit approval. These systematic biases not only expose the ethical risks associated with technological black boxes, but also raise deeper concerns about the preservation of social fairness and justice.

Existing research has predominantly approached these issues from external perspectives, including technological governance, legal regulation, and the construction of ethical frameworks, while paying comparatively little attention to a critical variable: human psychological response mechanisms. Psychological research has identified what may be called the AI moral deficit effect—a phenomenon that deserves particular scrutiny. When AI serves as the decision-making agent, public moral sensitivity and the tendency to attribute responsibility decline significantly. Even when AI engages in misconduct identical to that of a human decision-maker, individuals exhibit a markedly weaker willingness to punish it. This cognitive bias may generate a series of risks: encouraging organizations to use AI as a means of evading moral responsibility, worsening the difficulties faced by harmed groups in seeking redress, and gradually eroding society’s moral standards. Revealing the psychological mechanisms behind this effect and proposing corresponding interventions is therefore not only essential to advancing theories of human–AI interaction, but also urgent for building effective AI ethical governance systems and safeguarding social justice.

AI moral deficit effect in decision-making

Why do unethical decisions made by AI generally elicit weaker moral reactions than identical decisions made by human beings? Existing research commonly traces this phenomenon to mind perception, arguing that people generate moral responses only when they perceive a moral agent as possessing some degree of mind. According to mind perception theory, individuals perceive mind along two independent dimensions: agency and experience. In defining the conditions required for moral agency, traditional research—particularly perspectives derived from moral dualism—has emphasized perceived agency as the core prerequisite for holding an entity accountable for wrongdoing, while paying relatively little attention to the role of experience.

Our research argues that a theoretical perspective focused exclusively on agency is incomplete. Perceived experience is not only relevant to an entity’s status as a moral patient; it is also indispensable to its identity as a moral agent. The attenuation of public moral responses to AI therefore originates in a dual deficiency in mind perception: lower perceived agency and lower perceived experience. Through two parallel and independent pathways, these deficiencies jointly undermine AI’s standing as a moral agent, ultimately producing the AI moral deficit effect.

Agency refers to an entity’s capacity for intention, reasoning, goal pursuit, and communication. It is closely tied to moral responsibility: The greater an individual’s autonomy, and the clearer its intentions and motivations, the more responsibility they are expected to bear for their decisions and actions. Existing evidence indicates that people attribute a certain degree of agency to AI, but at a level significantly lower than that attributed to human beings. This perceptual gap constitutes the first psychological pathway underlying the AI moral deficit effect.

Experience refers to an entity’s capacity for emotional response, the feeling of pain, and subjective consciousness. It not only defines who can be harmed, and therefore who qualifies as a moral patient, but also profoundly shapes who can be regarded as a fully developed perpetrator, or moral agent. Its central mechanism lies in the fact that experience forms the basis of empathy and moral emotion. Only an agent capable of understanding and feeling the emotional states of others, including pain and happiness, is seen as capable of forming moral norms and anticipating the emotional consequences of its actions. This constitutes the psychological foundation of moral responsibility. Because AI is generally perceived as possessing far less experience than human beings, and as lacking the capacity to understand others’ emotions, it is often viewed as an incomplete moral actor. A deficit in perceived experience therefore constitutes the second critical psychological pathway leading to the AI moral deficit effect.

Building on these insights, our study revises and extends classical moral dualism by proposing that a complete moral agent requires the participation of both dimensions of mind. Public indifference toward unethical AI decisions arises precisely because AI is perceived as lacking both sufficient autonomous intention—reflecting an agency deficit—and the necessary emotional empathy—reflecting an experience deficit.

Dual-path psychological mechanism of AI moral deficit effect

Following an “effect–mechanism–intervention” research logic, this study conducted a systematic empirical investigation in three stages.

First, we developed culturally adapted moral scenarios tailored to the realities of Chinese society. Drawing on China’s sociocultural context, the study employed three typical forms of discrimination—educational discrimination, age discrimination, and gender discrimination—to demonstrate that individuals exhibit significantly weaker moral reactions to unethical decisions made by AI than equivalent decisions made by human decision-makers. Experimental paradigms developed within Western individualistic cultures may not fully capture the characteristics of moral cognition in collectivist cultures. Our findings therefore demonstrate the robustness of the AI moral deficit effect within China’s collectivist cultural context.

Among the three forms of discrimination examined, gender discrimination is particularly visible worldwide. Studying how AI inherits or amplifies such deeply entrenched biases can directly reveal the “replication–reinforcement” mechanism through which technology reproduces social structures. In recruitment settings, for example, gender discrimination is often encoded into algorithms through historical data, such as hiring records from male-dominated technology industries, causing AI systems to assign lower weights to female candidates during evaluation. This clearly traceable chain linking data, algorithms, and outcomes provides a particularly suitable context for analyzing the underlying logic of technological ethics.

Second, by integrating mind perception theory with moral dualism, this study systematically revealed the dual-path psychological mechanism behind the AI moral deficit effect. Previous research has largely explained this phenomenon through isolated psychological pathways, such as beliefs about free will or prejudice-related motivations, resulting in fragmented accounts. Starting from the concept of moral agency, our study clarifies the mediating roles of agency and experience in the AI moral deficit effect, thereby advancing both mind perception theory and moral dualism.

Unlike previous studies that examined isolated variables such as free will or autonomy, this research is the first to demonstrate that perceived agency and perceived experience function as parallel mediators in the AI moral deficit effect. This finding challenges the traditional dichotomy that links agency with moral agents and experience with moral patients. Existing studies have already documented the influence of perceived agency on moral evaluations of AI. Our findings further indicate that merely enhancing perceptions of agency—for example, by increasing decision-making transparency—is insufficient. Coordinated optimization of the experiential dimension, such as through emotional interaction design, is equally critical. Experimental data show that when perceptions of AI’s experiential capacity are strengthened, evaluations of AI as a moral agent rise significantly. This provides strong support for emerging forms of sentimental rationalism, which hold that experience is not merely a necessary condition for moral patients, but an intrinsic component of moral agency itself.

Third, through the combined use of experimental mediation methods and survey methodologies, this study achieved causal inference regarding the mechanisms underlying the AI moral deficit effect for the first time. A 2 × 2 factorial design revealed an interaction between decision-maker type and mind dimension: Under AI conditions, agency manipulation significantly strengthened moral responses, whereas no such effect appeared when the decision-maker was human.

This pattern of agency attribution bias supports a core assumption of mind perception theory: Human beings are inherently endowed with complete mental schemas, while AI occupies a socially constructed quasi-agentic status that can be modified through the representational design of mental attributes. By simultaneously manipulating perceived agency, perceived experience, and decision-maker type, the study demonstrated that in the low-agency control condition, decision-maker type significantly predicted levels of perceived agency and experience. Within the AI condition, increasing agency perceptions significantly improved moral responses, thereby establishing a complete causal chain from agent type to mind perception to moral response.

This design overcomes the inherent limitations of traditional survey methods in causal inference and moves beyond the correlational constraints of previous research. In addition, the survey component examined the parallel mediating roles of agency and experience simultaneously, revealing the synergistic effect between the two dimensions. These findings illuminate the deeper mechanism behind the AI moral deficit effect: individuals perceive algorithms as lacking both autonomous intentions (low agency) and the capacity to understand emotional harm (low experience), resulting in insufficient moral responses. This integrated framework resolves the fragmentation of previous studies, which have emphasized different mediating variables, and provides a unified explanatory account.

Limitations, future paths

Research on algorithmic ethics in disciplines such as computer science, philosophy, law, and sociology has primarily focused on proposing principles and technical pathways for fair algorithm design. By contrast, this study adopts a psychological perspective, emphasizing the differences in individuals’ psychological reactions when confronting AI versus human decision-makers. This perspective offers new theoretical insights for mitigating social problems arising from algorithmic bias and promoting the construction of fair algorithms, while also opening a new direction for research on algorithmic ethics.

Notwithstanding these advances, several limitations should be acknowledged. First, the experimental design relied on a limited range of scenarios. Future research should incorporate a broader array of moral contexts, including environmental protection, educational equity, and judicial sentencing. Second, this study primarily used scenario-based experimental paradigms, presenting moral decision-making situations through textual descriptions and video materials, which inevitably limits ecological validity. Future research could strengthen the real-world explanatory power of these findings through field experiments, longitudinal tracking, and the integration of multimodal data collection techniques.

Hu Xiaoyong is an associate professor from the psychology department at Wuhan University. This article has been edited and excerpted from Acta Psychologica Sinica, Issue 1, 2026.