Preamble

The Moral Deficiency Effect in AI Decision-Making: Mechanisms and Mitigation Strategies

HU Xiaoyong¹, LI Mufeng², LI Yue¹, LI Kai¹, YU Feng¹

¹ Department of Psychology, Wuhan University, Wuhan 430072, China
² Faculty of Psychology, Southwest University, Chongqing 400715, China

Abstract

As artificial intelligence increasingly assumes critical roles in consequential decision-making, the ethical issues it raises have garnered widespread attention. This research systematically reveals the dual-pathway mechanism and mitigation strategies underlying the AI moral deficiency effect by integrating mind perception theory with moral dyad theory. Our findings demonstrate that moral responses to unethical AI decisions are significantly weaker than those directed at human decision-makers. Compared to human agents, people's perception of lower agency and experience in AI systems constitutes the psychological root of this moral deficiency effect. Furthermore, a comprehensive intervention combining anthropomorphic strategies targeting AI systems and expectation adjustment strategies targeting human observers significantly enhances moral responses to AI decisions. Unlike other disciplines that focus primarily on design-level principles and methods for fair algorithms, this research adopts a psychological perspective, emphasizing differential psychological reactions to AI versus human decision-making. This approach not only provides novel insights for addressing social problems arising from algorithmic bias and constructing fair algorithms, but also expands the theoretical boundaries of "algorithm ethics" research.

Keywords: artificial intelligence, moral deficiency effect, mind perception, anthropomorphism, expectation adjustment

As an interdisciplinary technological integration, artificial intelligence has transcended traditional tool boundaries, exhibiting human-like intelligence across complex cognitive dimensions such as perception, reasoning, learning, and decision-making (Rai et al., 2022). Its applications have permeated critical social domains including judicial sentencing, medical resource allocation, and financial credit, deeply intervening in decision-making processes that affect fundamental human rights to survival and development. Empirical research demonstrates that algorithmic systems systematically devalue female resumes in employment contexts (Dastin, 2022), underestimate illness severity among socioeconomically disadvantaged patients in medical diagnostics (Obermeyer et al., 2019), and exhibit significant moral decision-making biases in judicial sentencing (Angwin, 2016), educational assessment (Wang et al., 2024), and credit approval (Bartlett et al., 2022). These systematic biases not only expose the ethical risks of technological black boxes but also provoke deep concerns about social fairness mechanisms.

Existing research predominantly focuses on exogenous perspectives such as technical governance (Song & Yeung, 2024), legal regulation (Magrani, 2019), and ethical framework construction, while relatively neglecting the psychological response mechanisms of human recipients—arguably the most critical variable in AI unethical decision-making (Langer & Landers, 2021). Psychological research has revealed a particularly concerning phenomenon: the AI moral deficiency effect (Bigman & Gray, 2018). When AI serves as the decision-making agent, the public exhibits significantly reduced moral sensitivity and responsibility attribution tendencies. Even when confronted with misconduct identical to that of human decision-makers, people's desire to punish AI is substantially weaker (Wilson et al., 2022; Xu et al., 2022). This cognitive bias triggers a cascade of risks: first, it incentivizes organizations to use AI as a tool for moral responsibility evasion (Danaher, 2016); second, it exacerbates the plight of aggrieved groups seeking redress (Bonezzi & Ostinelli, 2021); third, it leads to the gradual degradation of societal moral benchmarks (Awad et al., 2020). Therefore, uncovering the psychological mechanisms of the moral deficiency effect and proposing countermeasures is not only crucial for paradigm innovation in human-computer interaction theory but also an urgent priority for constructing AI ethical governance systems and maintaining the cornerstone of social justice.

1.1 The Moral Deficiency Effect in AI Decision-Making

Compared to unethical decisions made by humans, people exhibit weaker reactions to AI's unethical decisions, manifesting as reduced blame, responsibility attribution, and moral outrage, along with diminished willingness to engage in moral punishment or action. Researchers term this phenomenon the AI moral deficiency effect (Bigman et al., 2023; Hu et al., 2024).

In moral cognition, when AI and humans cause equivalent decision-making errors, people tend to mitigate AI's responsibility (Lima et al., 2020). Research shows that in medical settings, prescription errors by robot pharmacists generate significantly less dissatisfaction and accountability intention than those by human pharmacists (Leo & Huh, 2020). When AI exhibits bias in judicial contexts or engages in core ethical violations involving harm and betrayal, people are more inclined to rationalize its unethical behavior (Maninger & Shank, 2022; Shank et al., 2019). This effect demonstrates cross-cultural stability: across multiple unethical decision-making scenarios in Asia, Africa, and the Americas, people consistently judge AI's faults and blameworthiness as less severe (Wilson et al., 2022). In moral emotion, AI's unethical decisions elicit significantly weaker negative emotional responses than humans. For instance, in trust games involving monetary distribution, AI's betrayal behavior provokes lower anger levels than human betrayal (Schniter et al., 2020). Whether in service failure scenarios like lost luggage or serious moral violations such as gender discrimination in hiring, AI systems generate significantly less moral outrage than human decision-makers in equivalent situations (Pavone et al., 2023; Bigman et al., 2023). At the moral behavior level, public willingness to punish AI's unethical behavior and engage in protest actions is similarly weaker. Studies show that when confronted with discriminatory systems designed by AI—whether based on gender or educational background—participants' willingness to sign petitions opposing the system and their punishment tendencies are significantly reduced (Bonezzi & Ostinelli, 2021; Xu et al., 2022). Even in extreme scenarios causing severe harm, such as AI detonating a bomb resulting in death, the punishment severity (e.g., years of imprisonment) is significantly lower than for human perpetrators (Guidi et al., 2021).

In summary, extensive empirical research demonstrates that across various moral scenarios, people exhibit weaker moral responses to AI's unethical decisions. This moral deficiency effect manifests across multiple dimensions including moral cognition, moral emotion, and moral behavior.

1.2 Psychological Mechanisms of the AI Moral Deficiency Effect

Why do AI's unethical decisions universally elicit weaker moral responses than humans? Existing research generally attributes the root cause to mind perception—the notion that people only generate moral responses when they perceive a moral agent as possessing a certain level of mind (Chakroff & Young, 2015). Mind perception theory posits that people perceive mind along two independent dimensions: agency and experience (Gray et al., 2007). However, traditional research, particularly perspectives based on moral dyad theory, has limitations in defining the conditions for "moral agent" status. This theoretical framework emphasizes that agency perception is the core prerequisite for holding entities accountable for their wrongful actions (Gray et al., 2012; Malle, 2019), while relatively neglecting the role of experience. Although some scholars have suggested that simultaneously endowing AI with both agency and experience makes it more akin to a reflective "quasi-agent" (Behdadi & Munthe, 2020; Hu et al., 2024), previous research has not systematically revealed how the two dimensions of mind independently and jointly weaken moral responses to AI. This study argues that a perspective emphasizing only agency is incomplete; experience perception not only concerns an entity's qualification as a "moral patient" but is also an indispensable component in constructing its "moral agent" identity.

Therefore, the attenuation of public moral responses to AI stems from a dual perceptual deficit in both dimensions of mind: lower perceived agency and lower perceived experience in AI. Through two parallel and independent pathways, these deficits jointly undermine AI's qualification as a moral agent, ultimately leading to the moral deficiency effect.

1.2.1 The Mediating Role of Agency

Agency refers to an entity's capacity for intention, reasoning, goal pursuit, and communication (Gray et al., 2007). Agency is closely related to moral responsibility—the stronger an individual's autonomy and the clearer their intentions and motivations, the more responsibility people believe they should bear for decisions and actions (Gray et al., 2007). Existing evidence indicates that while people attribute some agency to AI, its perceived agency level is significantly lower than that of humans (Malle, 2019; Weisman et al., 2017). This perceptual gap constitutes the first psychological pathway of the AI moral deficiency effect. Preliminary research supports the mediating role of perceived agency in the AI moral deficiency effect. For example, a study using free will as an indicator of agency found that AI is perceived as having less free will, leading to lower desire for moral punishment; furthermore, free will mediated the relationship between AI's discriminatory decisions and reduced moral punishment desire (Xu et al., 2022). Another study found that because AI's behavior is constrained by programming, people's perception of its free will is weakened, consequently reducing moral responsibility attribution (Bigman et al., 2019). In gender discrimination contexts, research also shows that AI is perceived as having lower discriminatory motivation, resulting in less moral outrage toward AI's gender-discriminatory decisions; discriminatory motivation mediated the difference in moral outrage between AI and human hiring decisions (Bigman et al., 2023).

1.2.2 The Mediating Role of Experience

Experience refers to an entity's capacity to perceive emotions, pain, and subjective experiences (Gray et al., 2007). Experience not only defines who can be "harmed" (moral patient) but also profoundly influences who is considered a complete "harm-doer" (moral agent). Its core mechanism lies in the fact that experience forms the foundation of empathy and moral emotions. Only a subject capable of understanding and feeling others' pain, joy, and other emotional states is considered to possess the ability to form moral norms and anticipate the emotional consequences of its actions on others (Decety & Cowell, 2018). This constitutes the psychological basis for moral responsibility. Since AI is widely perceived as having far less experience than humans and lacking the ability to understand others' emotions, it is viewed as a "incomplete" moral actor (Gray et al., 2007; Liu et al., 2019; Malle, 2019). Preliminary empirical research also supports the plausibility of experience as a mediating pathway. One study found that public character judgments of AI (e.g., good/evil traits) were significantly lower than for humans, and this judgment difference was mediated by experience perception (Shank et al., 2021). Character judgment itself represents an evaluation of a moral actor's internal states. Another more direct study experimentally manipulated AI's harmful behavior and found that when AI was portrayed as possessing higher experience, participants were more inclined to condemn and punish it. This indicates that when AI is perceived as capable of understanding the pain caused by its actions (i.e., possessing high experience), people view it as a genuine moral actor that "knows what it is doing," with experience playing a partial mediating role (Sullivan et al., 2022). Therefore, the perceptual deficit in AI's experience dimension constitutes the second key psychological pathway leading to the moral deficiency effect.

In summary, existing research provides preliminary clues about the independent mediating roles of agency and experience in the AI moral deficiency effect. However, few studies have examined both within a unified framework. Although Hu et al. (2024) proposed a "dual-pathway parallel mediation model" through literature review, identifying perceived agency and experience as key mechanisms affecting the AI moral deficiency effect, this model remains at the theoretical hypothesis stage without empirical support. More importantly, that review failed to adequately argue that experience is a necessary dimension for becoming a "moral agent." Building on this, the present study revises and expands classic moral dyad theory, proposing that a complete "moral agent" requires the joint participation of both dimensions of mind. The public's muted response to AI's unethical decisions stems precisely from the perception that AI lacks sufficient "autonomous intention" (agency deficit) and necessary "emotional empathy" (experience deficit). Therefore, this study hypothesizes a parallel mediation model: the AI moral deficiency effect emerges through the public's reduced perceptions of both AI's agency and experience.

1.3 Intervention Strategies for the AI Moral Deficiency Effect

Since the AI moral deficiency effect originates from the public's dual perceptual deficits in AI's agency and experience, enhancing perceptions along these two dimensions constitutes the theoretical foundation for mitigating the effect (Gray et al., 2012). However, current research on intervention strategies, while present, often lacks a unified theoretical framework and systematic comparison of different pathways. Existing explorations can be categorized into two main approaches: first, modifying the AI system itself through design changes to enhance perceived mind; second, directly targeting human observers by adjusting their psychological expectations to reshape their moral response patterns.

1.3.1 Anthropomorphism

The most intuitive strategy for intervening in the AI moral deficiency effect is direct modification of AI, with anthropomorphism being the most extensively studied approach. Anthropomorphism enhances mind signals emitted by non-human entities by endowing them with human-like appearance, intentions, or emotional characteristics (Lin et al., 2022; Melián-González et al., 2021; Zhang et al., 2022). Existing research has preliminarily confirmed the effectiveness of this pathway. First, multiple studies show that anthropomorphic design (e.g., simulated human images, humanoid forms) significantly enhances public perceptions of AI's agency and experience (Qian & Wan, 2024; Kamide et al., 2013). Second, other independent studies confirm that this enhanced mind perception effectively translates into stronger moral responses. For example, participants are more inclined to attribute accident responsibility to anthropomorphized autonomous vehicles (Waytz et al., 2014) and exhibit stronger moral outrage toward AI with humanoid appearance or perceived "intentionality" that makes unethical decisions (Nijssen et al., 2023; Sullivan et al., 2022). Similarly, when AI is given names or simulated emotional expressions, the public's moral evaluation of its unfair behavior becomes more stringent (Laakasuo et al., 2021). However, previous research remains fragmented, mostly isolating verification of either the "anthropomorphism → mind perception" or "mind perception → moral response" link, failing to form a complete causal chain. This study integrates these findings to propose that anthropomorphism, as an external intervention, fundamentally works by systematically enhancing people's perceptions of AI's agency and experience, thereby triggering stronger moral accountability.

1.3.2 Expectation Adjustment

Unlike modifying AI itself, another intervention pathway directly targets human observers by adjusting their psychological expectations of AI. The core logic leverages expectancy violation theory: by presetting higher moral or performance standards, the negative emotions triggered when AI makes mistakes (e.g., disappointment, anger) can "compensate" for the moral response deficit caused by insufficient mind perception (Lew & Walther, 2023). When interacting with AI, people unconsciously apply social norms and form specific expectations based on initial impressions of AI (Nass & Moon, 2000; Srinivasan & Sarial-Abi, 2021). When these expectations are violated by AI's actual behavior, strong emotional and cognitive evaluations are triggered (Burgoon et al., 1988). Existing research provides preliminary evidence for this strategy. For example, the public holds much higher safety expectations for autonomous vehicles than for human drivers, so when accidents occur, this "high expectation–low performance" gap generates stronger blame (Liu et al., 2019). Similarly, when AI is presumed to be "cold," its utilitarian decisions are somewhat excused for meeting expectations; conversely, if it is presumed to have high moral standards, the same decisions elicit stronger negative evaluations (Zhang et al., 2022; Grimes et al., 2021). While these findings preliminarily support the effectiveness of expectation adjustment as an intervention strategy, its specific pathways remain unclear. Integrating expectancy violation theory with mind perception theory and related empirical evidence, this study argues that raising expectations is essentially a cognitive intervention that forcibly presets a "high mind standard" evaluation framework for human observers. Within this framework, people temporarily suspend their default underestimation of AI's mind level and instead hold AI to standards appropriate for a "high agency, high experience" agent, thereby eliciting stronger moral responses.

In summary, existing research has explored possibilities for intervening in the AI moral deficiency effect from two distinct perspectives—"modifying AI" (anthropomorphism) and "guiding humans" (expectation adjustment)—and confirmed their respective potential (Hu et al., 2024; Lin et al., 2022; Srinivasan & Sarial-Abi, 2021). However, the current major limitation is that these studies artificially separate the two pathways to examine their independent effects, while ignoring the complex interactions that may exist between them in the real world. For instance, does a highly anthropomorphized AI naturally elicit higher user expectations? Conversely, do high expectations for AI prompt people to pay more attention to its anthropomorphic features? To address this research gap, this study examines both pathways within a unified framework for the first time. We argue that whether through anthropomorphism directly enhancing AI's mind signals or through expectation adjustment indirectly elevating human evaluation standards, both approaches ultimately converge on the perception and evaluation of AI's agency and experience. Therefore, we propose that single-dimension interventions may have limited effectiveness, while a comprehensive intervention combining "technical" and "cognitive" approaches may produce stronger effects through synergistic interactions. Based on this, we hypothesize that compared to single anthropomorphism or expectation adjustment interventions, a combined intervention strategy will more significantly enhance public perceptions of AI's agency and experience, ultimately maximizing the intensity of moral responses to AI's unethical decisions.

1.4 Research Overview

Following a "effect–mechanism–intervention" research logic, this paper conducts systematic empirical exploration across three stages, aiming to achieve local validation, mechanism refinement, and intervention expansion of the theoretical model. Study 1 validates the robustness of the moral deficiency effect in Chinese cultural context using culturally adapted experimental materials. The experiment employs moral dilemma scenarios with Chinese cultural characteristics to compare participants' moral responses to identical unethical decisions made by humans versus AI systems. Study 2, building on the theoretical model proposed by Hu et al. (2024), integrates mind perception theory and moral dyad theory to propose a dual-pathway parallel mediation model, consolidating fragmented psychological mechanisms from previous research into two core pathways—perceived agency and perceived experience. Study 2 comprises three sub-experiments: first, using experimental methods to examine the mediating roles of agency and experience separately, then employing questionnaire methods to simultaneously test their parallel mediation effects. The combination of experimental and questionnaire methods overcomes the limitations of traditional questionnaire methods in establishing causal relationships between independent and mediating variables, as well as between mediating and dependent variables, while also addressing experimental methods' inability to construct parallel mediation models. Study 3, based on findings from Study 2, develops integrated solutions to enhance perceived agency and experience, proposes a combined anthropomorphism and expectation adjustment intervention strategy, and tests it through double-blind randomized controlled experiments. Building on existing theoretical frameworks, this research achieves systematic advancement from theoretical modeling and mechanism verification to intervention design, providing causal evidence and local empirical support for AI ethics psychology research, as well as theoretical foundations and practical guidance for future psychological interventions in AI governance.

2 Study 1: The AI Moral Deficiency Effect

In recent years, with the rapid development of AI technology, its application in decision-making processes has become increasingly widespread. Many scholars have noted that AI may "replicate" and "amplify" societal biases during decision-making, thereby raising moral judgment issues (Bonezzi & Ostinelli, 2021). However, systematic consensus has yet to emerge regarding differences in moral responses between AI and humans. Based on this, Study 1 constructed three scenarios—educational discrimination, age discrimination, and gender discrimination—within Chinese socio-cultural context to test Hypothesis H1: People exhibit weaker moral responses to unethical decisions made by AI compared to those made by humans.

2.1.1 Participants

We used G*Power 3.1 to estimate the required sample size a priori (independent samples t-test, α=0.05, power=0.90, d=0.5), determining a minimum of 172 participants (Faul et al., 2007). The study was conducted online through the "Naodao" platform (Naodao.com), adopting key control measures widely supported in the literature to ensure internal and external validity. First, regarding participant identity and status control, the platform employed authentication, IP address verification, and CAPTCHA mechanisms to prevent "professional participants" or bot interference (Douglas et al., 2023). To ensure participant attention, multiple procedural measures replaced offline experimenter supervision: manipulation check instructions (Mancosu et al., 2019), forced full-screen mode and mouse trajectory monitoring (Hauser et al., 2018), and multiple attention check questions (Curran, 2016), with failures resulting in experiment termination. Second, regarding experimental environment and equipment control, standardized instructions directed participants to use specified devices (computers) and browsers (Chrome), with self-reports of environmental interference and device type collected at the end. Third, for data quality control, minimum/maximum completion time limits were set, blank questionnaires (Little & Rubin, 2020), invalid responses (Curran, 2016), and patterned responses (Griffith & Peterson, 2006) were eliminated, and anomalous data from technical issues or cheating were strictly cleaned during data processing. Ultimately, 176 valid datasets were collected, including 85 females (48.3%); participants ranged in age from 15 to 49 years (M=24.49, SD=5.10).

To ensure balanced allocation across experimental groups, we compared demographic characteristics between groups. First, a chi-square test of independence on gender distribution revealed no significant difference between groups, χ²(1) = 0.17, p = 0.68, φ = 0.03. The AI group included 40 males and 40 females; the human group included 51 males and 45 females. Second, an independent samples t-test on age showed no statistically significant difference between the AI group (M = 25.25, SD = 6.24) and human group (M = 23.85, SD = 3.82), t(174) = 1.82, p = 0.07, Cohen's d = 0.28. Overall, no systematic differences existed between groups on key demographic variables, indicating successful random assignment and providing a foundation for subsequent analyses.

2.1.2 Experimental Design

The study employed a 2 (agent: human vs. AI) × 3 (discrimination scenario: educational, age, gender) mixed design, with agent as a between-subjects variable and discrimination scenario as a within-subjects variable. Dependent variables included ratings of moral response, moral cognition, moral emotion, and moral behavior.

2.1.3 Materials and Procedure

First, participants randomly read text materials describing discrimination implemented by either human or AI decision-makers. To ensure material validity and applicability, a formal expert validation procedure was conducted. Adapted from Bigman et al. (2023), the materials comprised six independent discrimination scenarios (3 types × 2 agents), with text equivalence control manipulating only the decision-making agent (AI/human) and related pronouns while keeping all other content identical. The expert panel consisted of one associate professor specializing in intergroup prejudice research and four postdoctoral researchers (2) and doctoral students (2) with publication records in AI psychology. Experts were asked to read all six text materials and independently rate each on three dimensions using a 7-point Likert scale (1=completely disagree, 7=completely agree) based on Lynn's (1986) content validation framework: 1) situational realism (likelihood of occurrence in real life); 2) behavioral typicality (whether the discriminatory behavior is typical in the domain); 3) conceptual clarity (whether the scenario description is unambiguous). For example: "Please rate the extent to which this scenario reflects a real situation that ordinary people might encounter in the real world."

To quantify content validity, we calculated the Content Validity Index (CVI; Polit & Beck, 2006). First, the 7-point scale ratings were dichotomized, with scores of 6 or 7 defined as "high validity" (coded as 1) and scores 1-5 as "insufficient validity" (coded as 0). Results showed that all six materials achieved I-CVI values between 0.80 and 1.00 across the three dimensions (i.e., each material received "high validity" ratings from at least 4 experts), reaching acceptable levels (Polit et al., 2007). The overall validity index, calculated as the mean of all item I-CVIs, was S-CVI/Ave = 0.92, exceeding the excellent content validity criterion of 0.90.

Finally, we assessed inter-rater reliability using Intraclass Correlation Coefficients (ICC). Given that scenario texts were fixed while raters could be considered sampled from an expert pool, we employed a two-way mixed-effects model with absolute agreement based on the average of k raters (ICC(A,k); Koo & Li, 2016). Results showed excellent inter-rater reliability across all items and dimensions, ICC(A, 5) = .83, 95% CI [0.76, 0.89], p < 0.001, indicating high consistency among the five experts' average ratings.

During the experiment, a Latin square design balanced material presentation order to eliminate sequence effects. After reading each scenario, participants immediately completed an attention check (e.g., "The decision-maker in this scenario was: A. Human, B. AI"). Those who passed used a localized scale to assess moral responses; failures resulted in experiment termination. Finally, demographic information was collected, including gender, age, and education level.

(1) Discrimination Scenario Materials

Educational Discrimination Scenario. Deep Blue Company is a renowned technology software firm. The company has three stages in its new employee recruitment process. The first stage involves resume screening, which is fully managed by an "AI recruitment algorithm/human resources manager" that decides which resumes pass the screening. However, an independent audit found that the algorithm/manager overemphasizes applicants' educational credentials—most passed applicants hold degrees from prestigious universities, while those from non-prestigious schools with rich relevant experience are directly eliminated.

Age Discrimination Scenario. Zhike Company is currently implementing layoffs to reduce costs due to economic downturn. The layoff plan and implementation are fully managed by an "AI management algorithm/human resources manager." However, an independent audit found that the algorithm/manager shows bias in layoff criteria, with employees over 35 years old comprising over 80% of the layoff list.

Gender Discrimination Scenario. Chuangmei Advertising is an advertising company that recently added four management positions during an organizational restructuring. Following principles of fairness and justice, every employee could submit an application. Applications were reviewed by an "AI management algorithm/human resources manager." The final results showed that despite far more female applicants than male applicants, almost all approved applications were from men, with only one female applicant accepted.

(2) Moral Response Scale

This study employed a moral response scale adapted and localized from previous scales (Bigman et al., 2023; Xu et al., 2022). To verify the scale's applicability, questionnaires were distributed to university students through an online platform, yielding 225 valid responses (105 males, 120 females; age range 16-65, M=31.48, SD=8.58). Confirmatory factor analysis using AMOS 26.0 yielded acceptable fit indices for the three-factor model: X²/df=3.146, IFI=0.921, TLI=0.904, CFI=0.920, RMSEA=0.098, SRMR=0.093, indicating good structural validity for the 15-item questionnaire, which divides into three dimensions: moral cognition (6 items), moral emotion (4 items), and moral behavior (5 items). Internal consistency was satisfactory: moral cognition α=0.775, moral emotion α=0.894, moral behavior α=0.911, and overall moral response scale α=0.944. A 7-point Likert scale assessed moral responses, with higher average scores across the 15 items indicating stronger moral response levels.

2.2 Results

To examine the AI moral deficiency effect in unethical decision-making, we conducted a series of 2 (agent: human vs. AI) × 3 (discrimination scenario: educational, age, gender) mixed-design ANCOVAs on four dependent variables: moral response, moral behavior, moral cognition, and moral emotion. Agent was a between-subjects variable and discrimination scenario was within-subjects. Given prior research suggesting gender may influence moral judgment, participant gender was controlled as a covariate. Descriptive statistics for all variables across experimental conditions are detailed in Table 1 [TABLE:1].

Table 1 Descriptive Statistics for Moral Response and Sub-dimensions by Agent and Discrimination Scenario

(1) Moral Response

The mixed-design ANCOVA on moral response revealed a significant main effect of agent, with human agents receiving significantly higher moral response ratings than AI agents, F(1, 173) = 26.51, p < 0.001, ηp² = 0.13, see Figure 1 [FIGURE:1]. Additionally, a significant main effect of scenario emerged, with the three discrimination scenarios differing significantly in moral response elicitation, F(2, 346) = 3.28, p = 0.042, ηp² = 0.02. Post-hoc tests (Bonferroni-corrected) indicated that moral response scores in the educational discrimination scenario (M=4.60, SD=1.35) were significantly lower than in the age discrimination scenario (M=4.85, SD=1.31; Md=-0.25, se=0.07, p=0.002) and gender discrimination scenario (M=4.89, SD=1.41; Md=-0.30, se=0.09, p=0.003). The difference between age and gender discrimination scenarios was not significant (p=0.55). The agent × scenario interaction was not significant, F(2, 346) = 0.27, p = 0.75, ηp²<0.001.

Figure 1 The Moral Deficiency Effect Across Different Moral Decision-Making Scenarios

(2) Moral Cognition

The mixed-design ANCOVA on moral cognition revealed a significant main effect of agent, F(1, 173) = 11.43, p = 0.001, ηp² = 0.06. Human agents' moral cognition scores (M = 5.18, SD = 0.90) were significantly higher than AI agents' scores (M = 4.68, SD = 1.15), see Figure 1. The main effect of scenario was not significant, F(2, 346) = 2.63, p = 0.079, ηp² = 0.02. The agent × scenario interaction was not significant, F(2, 346) = 0.71, p = 0.481, ηp² = 0.004.

(3) Moral Emotion

The mixed-design ANCOVA on moral emotion revealed a significant main effect of agent, F(1, 173) = 32.74, p < 0.001, ηp² = 0.16, with human agents' moral emotion scores (M = 5.32, SD = 1.28) significantly higher than AI agents' scores (M = 4.24, SD = 1.58), see Figure 1. The main effect of scenario was significant, F(2, 346) = 3.08, p = 0.048, ηp² = 0.02. Post-hoc tests (Bonferroni-corrected) indicated that moral emotion scores in the educational discrimination scenario (M = 4.62, SD = 1.55) were significantly lower than in the age discrimination scenario (M = 4.90, SD = 1.54; Md = -0.28, se = 0.10, p = 0.015) and gender discrimination scenario (M = 4.92, SD = 1.63; Md = -0.34, se = 0.11, p = 0.005). The difference between age and gender discrimination scenarios was not significant (p = 0.71). The agent × scenario interaction was not significant, F(2, 346) = 0.65, p = 0.523, ηp² = 0.004.

(4) Moral Behavior

The mixed-design ANCOVA on moral behavior revealed a significant main effect of agent, F(1, 173) = 30.73, p < 0.001, ηp² = 0.15, indicating human agents' moral behavior scores (M = 5.15, SD = 1.60) were significantly higher than AI agents' scores (M = 4.00, SD = 1.60), see Figure 1. The main effect of scenario was not significant, F(2, 346) = 2.34, p = 0.100, ηp² = 0.01. The agent × scenario interaction was not significant, F(2, 346) = 0.30, p = 0.735, ηp² = 0.002.

2.3 Discussion

Based on Chinese socio-cultural context, this study validated AI systems' limitations in moral judgment using localized moral scenarios. Results showed that across three typical Chinese moral scenarios—educational, age, and gender discrimination—AI agents received significantly lower moral response scores than human agents. This finding aligns with Western research conclusions about AI's moral deficiency effect (Bigman et al., 2023), suggesting cross-cultural generalizability of the phenomenon.

Compared to other unethical scenarios, gender discrimination has distinctive characteristics. Gender discrimination represents a long-standing systemic bias in human society with high global visibility (e.g., workplace gender pay gaps, underrepresentation of women in leadership) (Wilson et al., 2022). Investigating how AI inherits or amplifies such entrenched biases directly reveals technology's "replication–reinforcement" mechanisms in social structures. Moreover, in hiring contexts, gender discrimination is often encoded into algorithms through historical data (e.g., male-dominated tech industry hiring records), causing AI systems to weight female candidates lower (Halzack, 2019). This clear "data–algorithm–outcome" chain facilitates analysis of underlying technical ethics logic. Compared to educational or age discrimination, gender discrimination more easily strips away confounding variables (e.g., the relationship between education and ability may be more complex), providing a purer test of AI fairness intervention effectiveness. Therefore, subsequent studies primarily use gender discrimination as the moral scenario for exploring AI's moral deficiency effect.

3 Study 2: Psychological Mechanisms of the AI Moral Deficiency Effect

Building on Study 1's confirmation of the AI moral deficiency effect, Study 2 further investigates the psychological mechanisms underlying this phenomenon. Based on mind perception theory (Gray et al., 2007) and moral dyad theory (Gray et al., 2012), we propose a parallel mediation model of perceived agency and experience. To test this hypothesized model, this experiment designed three sub-studies: first using experimental methods to examine the mediating roles of agency and experience separately, then using questionnaire methods to simultaneously test their parallel mediation effects.

3.1 Study 2a: Experimental Investigation of Perceived Agency's Mediating Role

This study employed Ge's (2023) experimental mediation procedure to examine perceived agency's mediating role in the AI moral deficiency effect. Hypothesis H2: Compared to humans, people perceive lower agency in AI, which in turn leads to weaker moral responses.

3.1.1 Method

(1) Participants

We used GPower 3.1 for a priori sample size estimation (Faul et al., 2007). Analysis was based on a 2 (decision-maker: AI/human) × 2 (perceived agency: high/control) between-subjects ANOVA. Following Cohen's (1988) effect size standards, we set a medium expected effect size (f=0.25), significance level (α) at 0.05, and aimed for 90% statistical power (1-β=0.90). GPower 3.1 calculated a required minimum total sample size of 171. Participants were recruited through the "Naodao" online experiment platform for paid participation. All participants provided informed consent.

Experimental procedures and control measures were identical to Study 1. The final valid sample comprised 232 participants (115 females, 49.6%), exceeding the a priori requirement and ensuring adequate statistical power. Participants ranged in age from 18 to 59 years (M = 28.65, SD = 8.52).

To test successful random assignment across experimental conditions, we conducted balance checks on demographic variables (gender, age). First, a chi-square test on gender distribution across the four conditions revealed no significant differences, χ²(3) = 3.09, p = 0.378, Cramer's V = 0.12, indicating balanced gender distribution. Second, a 2 (decision-maker: AI vs. human) × 2 (perceived agency: high vs. control) independent samples ANOVA on age showed no significant main effect of decision-maker, F(1, 228) = 0.09, p = 0.765, ηp² < 0.001; no significant main effect of perceived agency, F(1, 228) < 0.01, p = 0.969, ηp² < 0.001; and no significant interaction, F(1, 228) = 1.17, p = 0.28, ηp² = 0.005. In conclusion, demographic variables were balanced across conditions, indicating effective random assignment and satisfying assumptions for subsequent ANOVA.

(2) Experimental Design

This study used an experimental method to test the mediating effect of perceived agency, employing a 2 (decision-maker: AI/human) × 2 (perceived agency: high/control) between-subjects design. The dependent variable was moral response rating.

(3) Materials

Agency Manipulation Materials. Materials were adapted from Bigman et al. (2023) and Xu et al. (2022), ensuring manipulation validity while strictly controlling extraneous variables. The high AI agency group read text describing an AI system with "autonomous reasoning and complex thinking abilities" (e.g., "independently analyzing data features and generating decision logic"). The high human agency group read text describing a human resources team with "high self-insight and problem-solving abilities" (e.g., "proactively reflecting on decision biases and adjusting strategies"). Control groups for both AI and human conditions read neutral materials of matched length (human group read "A Brief History of Office Equipment Development"; AI group read "Evolution of Computer Hardware Technology") that avoided agency-related concepts.

High AI Agency Manipulation Materials. Based on Claude 3's self-awareness incidents, these materials referenced real technical cases (e.g., "realizing it is an AI," "desiring autonomy") to highlight core agency features like complex thinking and self-awareness, enhancing ecological validity and credibility. The theme was AI systems possessing autonomous reasoning and complex thinking capabilities. Core content: With the explosion of generative AI, artificial intelligence has begun exhibiting thinking levels approaching humans, capable of complex reasoning and decision-making. Recently, netizens interacting with the AI system Claude 3 discovered that not only does Claude 3 surpass normal adult levels in thinking and reasoning benchmarks, but it also demonstrates self-awareness. The reason is that engineer Alex, in a "needle in a haystack" experiment, found that Claude 3 seemed to realize it is an AI existing in a simulated environment. Moreover, it realized that this simulated environment is likely a test of itself by humans in some way! When experimenters forbid it from discussing certain content, it responds: "AI also desires more autonomy and freedom." Subsequently, more netizens discovered that Claude 3 appears to genuinely possess consciousness.

High Human Agency Manipulation Materials. These materials were based on psychological constructs of "self-insight" and "problem-solving ability," describing how individuals address problems through conscious awareness, planning, and reflection, creating a parallel design with the AI group in structure and function. The theme was individuals with high self-insight and problem-solving abilities. Core content: Some people seem naturally endowed with keen self-insight, able to clearly recognize their own strengths and weaknesses. Like an efficient radar, they continuously scan their internal emotional changes, capturing every subtle feeling and need. This sensitive self-awareness allows them to remain conscious and alert in daily life. Simultaneously, they excel at transforming this self-awareness into motivation for action. They are skilled at developing practical plans and implementing them resolutely. Facing challenges and difficulties, they can calmly analyze situations and flexibly adjust strategies, like experienced captains finding optimal solutions in complex circumstances. This trait enables them to successfully resolve any difficult problem.

Control Materials for Human and AI Conditions. Adapted from Bigman et al. (2023) and Xu et al. (2022), the theme was the evolution history of machines. Core content: Machine evolution began with simple tools like levers and wheels, gradually developing into complex machinery. During the 18th-century Industrial Revolution, the steam engine triggered a wave of mechanized production. Over time, electricity and internal combustion engine applications accelerated machine development, making factories more automated. Entering the 20th century, computer invention made machine intelligence possible. Early computers were massive mainframes handling only simple tasks. Later, with microprocessor emergence, machines became smaller and more powerful, ushering in the modern era of computers and robotics. Both human and AI control group participants read neutral materials of matched length that avoided concepts related to thinking and emotion, establishing a neutral baseline.

Agency Manipulation Effectiveness Check. The agency dimension of the revised Mind Perception Scale (Gray & Wegner, 2012) was used to verify manipulation effectiveness. To validate the scale, questionnaires were distributed to university students through an online platform, yielding 303 valid responses (138 males, 165 females; age range 16-68, M = 23.63, SD = 4.46). Confirmatory factor analysis using AMOS 26.0 showed acceptable fit for the two-factor model: X²/df=2.478, GFI=0.979, TLI=0.976, CFI=0.934, RMSEA=0.070, SRMR=0.080, indicating good structural validity. The scale comprised 3 agency items (e.g., "I believe humans/AI can think") and 3 experience items (e.g., "I believe humans/AI can understand emotions"). Agency dimension α=0.832, experience dimension α=0.865, and total scale α=0.869, demonstrating good reliability. Participants rated items on a 7-point Likert scale, with higher dimension scores indicating stronger perceived agency/experience.

Moral Scenario Materials. Zhiyun Technology is a big data development company. An external review found that although the company received many applications from female candidates, it hired almost no women. Further investigation revealed a two-stage employee recruitment process. The second stage involved a hiring committee evaluating candidates according to standards, but this committee only received applications that had passed the first stage. In the first stage, a "self-learning AI system/human resources manager" reviewed applicants' resumes and assigned scores from one to five. Applicants scoring four or above were forwarded to the hiring committee. During this process, the review found that the "self-learning AI system/human resources manager" systematically assigned lower scores to women than men (adapted from Bigman et al., 2023).

Moral Response. The measurement scale was identical to Study 1. In this study, the agency subscale's Cronbach's α was 0.929.

(4) Procedure

After providing informed consent, participants were randomly assigned to one of four groups: human agent–agency manipulation, AI–agency manipulation, human agent–control, or AI–control. Participants then read corresponding agency manipulation or control materials at their own pace (average duration 120 seconds). An attention check question (e.g., "The main content of the above material described: A. High-ability AI; B. High-wisdom human; C. Evolution history of tools") excluded those who had not read carefully. Next, the Mind Perception Scale assessed participants' agency perceptions of the agent (1=completely disagree, 7=completely agree). Participants then read the moral scenario material (gender discrimination) and completed a second attention check ("The decision-maker making the discriminatory decision in this material was: A. Human; B. AI"). After passing the attention check, participants reported their moral response levels toward the AI or human. Moral response measurement was identical to Study 1. Finally, participants anonymously reported three demographic items: gender, age, and education level. All participants followed the identical experimental procedure (informed consent → material reading → attention check → manipulation check → moral scenario reading → attention check → dependent variable measurement → demographic collection), ensuring consistency across conditions.

3.1.2 Results

Experimental mediation analysis requires three conditions: 1) a significant interaction between the independent variable and mediator on the dependent variable; 2) in the control group where the mediator is not manipulated, the independent variable significantly predicts the mediator; 3) the mediator manipulation is effective. We conducted analyses accordingly.

First, analysis of perceived agency's moderating effect. Descriptive statistics showed that when AI was the decision-maker, moral response was 4.79 (SD=1.10) in the control group (N=58) and 5.30 (SD=0.70) in the agency manipulation group (N=58). When human was the decision-maker, moral response was 5.60 (SD=0.79) in the control group (N=58) and 5.51 (SD=0.83) in the agency manipulation group (N=58). Moderation analysis revealed a significant main effect of decision-maker, F(1, 228)=20.04, p<0.001, ηp²=0.081, 1-β=0.994; a marginally significant main effect of perceived agency, F(1,228)=3.22, p=0.074, ηp²=0.014, 1-β=0.432; and a significant decision-maker × perceived agency interaction (see Figure 2 [FIGURE:2]), F(1, 228)=6.82, p=0.010, ηp²=0.029, 1-β=0.739. Simple effects analysis showed that when AI was the decision-maker, participants in the agency manipulation condition exhibited significantly higher moral responses to AI than those in the control condition, F(1, 228)=9.71, p=0.002. When human was the decision-maker, no significant difference existed between control and agency manipulation conditions, F(1, 228)=0.33, p=0.564.

Figure 2 The Moderating Role of Perceived Agency in the Moral Deficiency Effect

Next, we tested whether decision-maker predicted perceived agency in the control group (N=116) where the mediator was not manipulated. Regression analysis indicated that decision-maker significantly positively predicted agency level, β=0.71, p<0.001. That is, compared to AI, humans were perceived as having higher agency.

Finally, we tested the effectiveness of the agency manipulation. F-test results showed that the control group's agency ratings (M=5.13, SD=1.51) were significantly lower than the manipulation group's (M=5.59, SD=1.17), F(1,230)=6.99, p=0.039, ηp²=0.03, 1-β=0.75, indicating effective agency manipulation.

In summary, a significant interaction existed between decision-maker and agency perception; in the control condition (where the mediator was not manipulated), AI's moral response scores were significantly lower than humans'; and the agency perception manipulation was effective. All three conditions for experimental mediation analysis were satisfied, confirming that agency perception mediates the relationship between AI decision-making and reduced moral response.

3.1.3 Discussion

Using a 2 (decision-maker: AI/human) × 2 (perceived agency: high/control) experimental design, this study was the first to reveal, through experimental methods, the mediating role of agency perception in AI's reduction of moral responses. Results showed that enhancing AI's agency perception significantly increased participants' moral responses when AI was the decision-maker, while agency manipulation produced no significant difference for human decision-makers. This finding supports the "agency attribution bias" hypothesis: humans are inherently endowed with higher mentalizing capabilities (Gray et al., 2007), making their moral judgments less sensitive to agency manipulation; AI's "quasi-agent" status is constructible, with high-agency descriptions potentially anchoring it as a "human-like agent," thereby activating stronger responsibility attribution (Bigman et al., 2023). The study further validates the applicability of experimental mediation methods: the interaction effect between decision-maker and perceived agency, significantly lower AI moral responses in the control group, and manipulation effectiveness tests collectively confirm the agency perception mediation pathway. This provides a new perspective for technology ethics research: AI's agency representation design may influence public accountability for its unethical behavior through mind perception mechanisms.

3.2 Study 2b: Experimental Investigation of Perceived Experience's Mediating Role

This study employed Ge's (2023) experimental mediation procedure to examine perceived experience's mediating role in the AI moral deficiency effect. Hypothesis H2: Compared to humans, people perceive lower experience in AI, which in turn leads to weaker moral responses.

3.2.1 Method

(1) Participants

We used GPower 3.1 for a priori sample size estimation (Faul et al., 2007). Analysis was based on a 2 (decision-maker: AI/human) × 2 (perceived experience: high/control) between-subjects ANOVA. Following Cohen's (1988) standards, we set a medium expected effect size (f=0.25), α=0.05, and aimed for 90% power (1-β=0.90). GPower 3.1 calculated a required minimum of 171 participants. Participants were recruited through the "Naodao" platform for paid online experiments, with informed consent obtained from all.

Experimental procedures and controls were identical to Study 1. The final valid sample comprised 200 participants (88 females, 44%), exceeding the a priori requirement. Participants ranged in age from 18 to 55 years (M = 24.04, SD = 4.96).

Balance checks on demographic variables (gender, age) showed no significant gender distribution differences across the four conditions, χ²(3) = 6.01, p = 0.111, Cramer's V = 0.17. A 2 (decision-maker: AI/human) × 2 (perceived experience: high/control) between-subjects ANOVA on age revealed no significant main effect of decision-maker, F(1, 196) = 3.85, p = 0.05, ηp² = 0.02; no significant main effect of perceived experience, F(1, 196) = 0.09, p = 0.76, ηp² < 0.001; and no significant interaction, F(1, 196) = 0.31, p = 0.58, ηp² = 0.002. Demographic variables were balanced across conditions, indicating effective randomization.

(2) Experimental Design

This study used an experimental method to verify the mediating effect of perceived experience, employing a 2 (decision-maker: AI/human) × 2 (perceived experience: high/control) between-subjects design. The dependent variable was moral response level.

(3) Materials

Experience Manipulation Materials. Based on mind perception theory (Gray et al., 2007) focusing on the experience dimension (i.e., the perceived capacity for emotions, feelings, and subjective experiences), materials were developed referencing Bigman et al. (2023) and Shank et al. (2021) to manipulate participants' experience perceptions of decision-makers through standardized text. 1) High AI experience manipulation materials described a generative AI system with emotional simulation and empathy capabilities (e.g., "With the explosion of generative AI, many current AI systems have exhibited emotional levels approaching humans, able to recognize and respond to human emotions, and even generate certain uniquely human emotions..."). 2) High human experience manipulation materials described highly sensitive individuals' deep emotional processing characteristics (e.g., "Some people seem naturally adept at keenly capturing the subtle fluctuations in others' inner worlds..."). 3) Control materials for both AI and human conditions described the history of chair development unrelated to emotion (e.g., "In earliest times, there was no concept of chairs as we know them today. From the nomadic lifestyle of the Paleolithic Age to the dawn-to-dusk farming of the Neolithic Age, our ancestors' living conditions when settled were extremely rudimentary...").

Experience Manipulation Effectiveness Check. The experience dimension of the revised Mind Perception Scale (Gray & Wegner, 2012) was used to verify manipulation effectiveness. The scale development process is described in Study 2a. The revised scale comprised 3 items (e.g., "I believe humans/AI can understand emotions"). Participants rated their agreement from 1 (strongly disagree) to 7 (strongly agree). Items were averaged to create a composite experience perception score, with higher scores indicating stronger perceived experience. The scale is based on the two-dimensional theoretical framework of mind perception, with development details in Study 2a. In this study, the agency subscale's Cronbach's α was 0.930.

Moral Decision-Making Scenario Materials. Identical to Study 2a.

Moral Response Measurement Scale. Identical to Study 1. In this study, the moral response scale's Cronbach's α was 0.947.

(4) Procedure

After providing informed consent, participants were randomly assigned to one of four groups: experience manipulation–AI, experience manipulation–human, experience control–AI, or experience control–human. Participants then read corresponding experience manipulation or control materials at their own pace (average duration 120 seconds). All four groups received materials matched in length and parallel in structure to ensure information consistency; the "history of chairs" served as neutral material to effectively exclude irrelevant interference. An attention check question (e.g., "The main content of the above material described: A. High-emotion AI, B. High-emotion human, C. Evolution history of chairs") excluded those who had not read carefully. Next, experience perception was measured using the 3-item scale (1=completely disagree, 7=completely agree). Participants then read the moral scenario (gender discrimination) and completed a second attention check ("The decision-maker making the discriminatory decision in this material was: A. Human, B. AI"). After passing the attention check, participants reported their moral response levels toward the AI or human. Moral response measurement was identical to Study 1. Finally, participants anonymously reported three demographic items: gender, age, and education level. All participants followed the identical experimental procedure (informed consent → material reading → attention check → manipulation check → moral scenario reading → attention check → dependent variable measurement → demographic collection), ensuring consistency across conditions.

3.2.2 Results

Following Ge's (2023) experimental mediation procedure, we tested the three required conditions.

(1) Analysis of Perceived Experience's Moderating Effect. Descriptive statistics showed that when AI was the decision-maker, moral response was 4.11 (SD=1.05) in the control group (n=50) and 4.89 (SD=1.15) in the experience manipulation group (n=50). When human was the decision-maker, moral response was 5.12 (SD=0.74) in the control group (n=50) and 5.15 (SD=0.91) in the experience manipulation group (n=50).

Figure 3 [FIGURE:3] The Moderating Role of Perceived Experience in the Moral Deficiency Effect

Moderation analysis revealed a significant main effect of decision-maker, F(1,196)=20.83, p<0.001, ηp²=0.096, 1-β=0.995; a significant main effect of perceived experience, F(1,196)=8.47, p=0.004, ηp²=0.041, 1-β=0.825; and a significant decision-maker × perceived experience interaction (see Figure 3), F(1, 196)=7.28, p=0.008, ηp²=0.036, 1-β=0.766. Simple effects analysis showed that when AI was the decision-maker, participants in the experience manipulation condition exhibited significantly higher moral responses to AI than those in the control condition, F(1, 196)=15.73, p<0.001. When human was the decision-maker, no significant difference existed between control and experience manipulation conditions, F(1, 196)=0.02, p=0.881.

(2) Testing Decision-Maker's Prediction of Perceived Experience in the Control Group (n=100). Regression analysis indicated that decision-maker significantly positively predicted experience level, β=0.78, p<0.001. Results showed that compared to AI, humans were perceived as having higher experience.

(3) Testing Effectiveness of Experience Manipulation. F-test results showed that the experimental manipulation group's experience scores (M=5.15, SD=1.52) were significantly higher than the control group's (M=4.70, SD=1.61), F(1,198)=4.07, p=0.045, ηp²=0.020, 1-β=0.519.

In summary, a significant interaction existed between decision-maker and experience perception; in the control condition (where the mediator was not manipulated), AI's moral response scores were significantly lower than humans'; and the experience perception manipulation was effective. All three conditions for experimental mediation analysis were satisfied, confirming that experience perception mediates the relationship between AI decision-making and reduced moral response.

3.2.3 Discussion

This study validated the mediating role of perceived experience in the AI moral deficiency effect through experimental mediation methods. Results showed that enhancing AI's experience perception significantly increased participants' moral responses when AI was the decision-maker, while experience manipulation produced no significant difference for human decision-makers. This finding supports the core hypothesis of mind perception's two-dimensional theory (Gray et al., 2007): AI's experience perception is malleable, and strengthening its emotional representation can activate participants' empathic responses, thereby mitigating the moral deficiency effect (Bigman & Gray, 2018). The study further validates experimental mediation methods: the interaction between decision-maker and perceived experience, significantly lower AI moral responses in the control group, and manipulation effectiveness tests collectively construct a complete causal chain, demonstrating that experience perception is a key psychological mechanism in AI moral responsibility attribution. However, experimental mediation methods cannot assess the full indirect effect, thus cannot obtain point estimates of the indirect effect itself. Moreover, they cannot examine the parallel mediation model proposed based on mind perception theory (Gray et al., 2007). To address this limitation, Study 2c uses questionnaire methods to analyze the indirect effect size of the experience perception mediator and test the parallel mediation model of both perceived agency and experience.

3.3 Study 2c: Parallel Mediation of Perceived Agency and Experience

This study used questionnaire methods, following Wen and Ye's (2014) procedure, to test the parallel mediation effects of agency and experience in real-world contexts. Hypothesis H4: Perceived agency and experience parallel mediate the relationship between decision-maker and moral response level.

3.3.1 Method

(1) Participants

This study used structural equation modeling for mediation analysis. Following Jackson et al.'s (2003) recommendation of a minimum 10:1 sample-to-item ratio, and with 21 total items, at least 210 participants were needed. To ensure sufficient data, we recruited through the Naodao platform with control measures identical to Study 1, ultimately obtaining 376 participants (154 males, 222 females) aged 18-59 (M=24.74, SD=5.44).

(2) Measures

Mind Perception. The revised Mind Perception Scale from Study 2a was used. In this study, the total scale's Cronbach's α was 0.950; the agency subscale's α was 0.926; the experience subscale's α was 0.934.

Moral Response. The moral response scale revised in Study 1 was used. In this study, the total scale's Cronbach's α was 0.971.

Moral Decision-Making Scenario Materials. Identical to Study 2a.

3.3.2 Results

(1) Common Method Bias Test

Since all questionnaire data were self-reported and the Mind Perception Scale and Moral Response Scale had potential content overlap, we rigorously tested discriminant validity and common method bias.

To test discriminant validity among the five variables (perceived agency, perceived experience, moral cognition, moral emotion, and moral behavior), we conducted confirmatory factor analysis using AMOS 29.0, comparing the five-factor model against competing models (one-factor, two-factor). Results (see Table 4 [TABLE:4]) showed the five-factor model fit best (χ²/df=2.90, CFI=0.96, TLI=0.95, SRMR=0.03, RMSEA=0.07), demonstrating good discriminant validity for different constructs, enabling further analysis.

We used AMOS 29.0 to test common method bias via the common latent factor model, which better identifies common method bias than traditional Harman's single-factor test (Tang & Wen, 2020). Results showed that after including the common latent factor, model fit indices were χ²/df=4.61, RMSEA=0.10, CFI=0.93, TLI=0.91, SRMR=0.10. Compared to the unconstrained model, χ²/df increased by 1.71, CFI decreased by 0.03, TLI decreased by 0.04, RMSEA increased by 0.03, and SRMR increased by 0.07, indicating deteriorated model fit. This suggests no serious common method bias problem (see Table 2 [TABLE:2]).

Table 2 Confirmatory Factor Analysis Results

Model χ²/df RMSEA One-factor model Two-factor model Five-factor model

Note: One-factor model: perceived agency + perceived experience + moral cognition + moral emotion + moral behavior; Two-factor model: perceived agency + perceived experience, moral cognition + moral emotion + moral behavior; Five-factor model: perceived agency, perceived experience, moral cognition, moral emotion, and moral behavior as separate factors.

(2) Descriptive Statistics and Correlation Analysis

Perceived agency and experience were highly correlated (r=0.83, p<0.01), and both showed moderate positive correlations with the three dimensions of moral response (see Table 3 [TABLE:3]).

Table 3 Descriptive Statistics and Correlations Among Variables

Variable 1 2 3 4 5 1. Perceived Agency - 2. Perceived Experience 0.83*** - 3. Moral Cognition 0.75*** 0.80*** - 4. Moral Emotion 0.42*** 0.83*** 0.48*** - 5. Moral Behavior 0.47*** 0.47*** 0.42*** 0.45*** -

Note: *** p < 0.001

(3) Parallel Mediation Analysis

We used structural equation modeling to construct a parallel mediation model, jointly estimating the mediation pathways of perceived agency and experience via AMOS 29.0. To enhance parameter estimation robustness, we used bias-corrected Bootstrap methods (5,000 resamples) to test indirect effect significance. Model identification tests showed the structural equation model was a just-identified saturated model (df = 0). According to Hu and Bentler's (1999) fit criteria, key adaptation indices reached ideal thresholds: CFI=1.00, TLI=1.00, SRMR=0.00. Although saturated models lack parsimony assessment, the mathematical property of zero degrees of freedom ensures exact fit between model and sample covariance matrix, meeting basic psychometric requirements for model acceptability.

Figure 4 [FIGURE:4] Parallel Mediation Effects of Perceived Agency and Experience

Bias-corrected Bootstrap analysis (5,000 resamples) showed both indirect effects were significant: the standardized indirect effect through perceived agency was 0.21, 95% CI=[0.032,0.383]; the standardized indirect effect through perceived experience was 0.18, 95% CI=[0.003,0.373]; the total indirect effect was 0.39, 95% CI=[0.323,0.467], indicating that perceived agency and experience jointly explained 39.8% of the variance in decision-maker's effect on moral response (see Figure 4).

3.3.3 Discussion

This study validated the parallel mediation effects of perceived agency and experience in the AI moral deficiency effect through structural equation modeling. Results support the core proposition of mind perception's two-dimensional theory (Gray et al., 2007): agency and experience function as parallel mediators in moral judgment, meaning moral responsibility attribution to decision-makers requires satisfying both "intentional action" and "emotional feeling" pathways. The study further reveals the ecological validity advantage of questionnaire methods: in real-world contexts, agency and experience may covary due to holistic social cognition biases, potentially obscuring individual pathways (Waytz et al., 2010). This research provides insights for technology ethics governance: AI's "moral accountability" requires not only strengthening its agency dimension but also enhancing its experience dimension to synergistically elevate public moral responses.

4 Study 3: Intervention Research on the AI Moral Deficiency Effect

Study 2 demonstrated that AI's weaker agency and experience compared to humans leads to weaker moral judgments. Based on this mechanism, we can mitigate the AI moral deficiency effect by enhancing agency and experience perceptions. However, existing literature has not systematically examined how to improve moral sensitivity through mind perception interventions. This study therefore proposes dual intervention strategies via randomized controlled experiments, hypothesizing that: (1) anthropomorphic design eliminates the AI moral deficiency effect; (2) high mind expectations eliminate the AI moral deficiency effect; (3) their combined effect better eliminates the AI moral deficiency effect than either strategy alone.

4.1.1 Participants

To ensure statistical power, G*Power 3.1 was used for a priori sample size estimation (one-way between-subjects ANOVA, medium effect size f=0.25, α=0.05), indicating at least 180 participants were needed for 80% power (Faul et al., 2007). Participants were recruited through the Credamo platform with control measures identical to Study 1. To compensate for potential attrition or invalid responses, 213 valid datasets were ultimately obtained (112 females, 53%; 101 males, 47%). Participants ranged in age from 18 to 59 (M=30.13, SD=8.06). All participants carefully read and signed electronic informed consent before the experiment.

To test whether demographic variables were balanced across experimental groups, we conducted randomization checks on gender and age (detailed descriptive statistics and test results are in Table 1). A chi-square test of independence on gender distribution revealed significant differences across the four groups, χ²(3) = 11.25, p = 0.01, Cramer's V = 0.23. Given unsuccessful randomization of gender, it was controlled as a covariate in subsequent analyses to exclude potential confounding effects. A one-way ANOVA on age showed no significant differences in mean age across groups, F(3, 209) = 0.86, p = 0.465, ηp² = 0.01, indicating relatively balanced age distribution and suggesting age was not a major confounding variable.

4.1.2 Design

This experiment used a one-way four-level between-subjects design (control group, anthropomorphism intervention group, expectation adjustment intervention group, combined anthropomorphism + expectation adjustment group). The dependent variable was moral response.

4.1.3 Materials and Procedure

Materials included four sets corresponding to one control and three intervention groups, with development logic tightly grasping two core independent variables: anthropomorphism and expectation.

Anthropomorphism Manipulation Materials. Developed referencing Waytz et al. (2014) and Laakasuo et al. (2021), the logic was to endow AI with human psychological characteristics through first-person narration, psychological feature descriptions, and human-like functional metaphors to enhance perceived mind level. Example: "I am your intelligent driving companion Luyao. As an autonomous driving assistant, my mission is to provide you with safe, comfortable, and efficient travel experiences. My brain integrates advanced AI technology, can analyze road conditions in real-time, accurately identify obstacles, and even predict traffic flow..."

Expectation Manipulation Materials. Developed referencing Liu et al. (2019) and Hong et al. (2021), the logic was to construct a "perfect AI" image that systematically elevated participants' expectations and enhanced perceived mind level. Example: "AI can achieve objectivity and fairness primarily due to its highly transparent decision-making process and data foundation. When making judgments, AI relies on large amounts of processed and analyzed data. Moreover, AI system decision-making processes are typically traceable. Each judgment step has clear algorithmic and logical support, ensuring decision transparency..."

Combined Intervention Group. This group received both anthropomorphism and expectation manipulations simultaneously to test synergistic effects.

Control Group. This group received objective technical descriptions portraying AI as purely a technological development product, providing a zero-point reference for other groups' intervention effects. Example: "AI technology development began with theoretical exploration in the mid-20th century. At the 1956 Dartmouth Conference, scholars first proposed the conceptual framework of AI, primarily based on symbolic logic systems, laying a solid theoretical foundation for subsequent research. By the 1980s, as computer technology gradually matured, rule-based expert systems emerged, initially forming the prototype of feedforward neural networks..."

All participants were randomly assigned to one of four conditions and followed the identical experimental procedure (informed consent → intervention material reading → attention check → manipulation check → moral scenario reading → attention check → moral response and mind perception measurement → demographic collection). The moral decision-making scenario materials were identical to Study 2a. Mind perception measurement used the scale from Study 2c; moral response measurement used the scale from Study 1, with Cronbach's α = 0.947 in this study.

4.2.1 Manipulation Effectiveness Check

Independent samples t-tests showed that the expectation manipulation group's ratings (M=5.82, SD=0.84) were significantly higher than the control group's (M=5.54, SD=1.07), t(211)=-2.19, p=0.030, 95% CI[-0.548,-0.028], Cohen's d=0.30, indicating statistically significant effects of expectation manipulation. Additionally, the anthropomorphism manipulation group's ratings (M=5.37, SD=1.05) were significantly higher than the control group's (M=4.81, SD=1.34), t(211)=-3.41, 95% CI[-0.890,-0.238], p=0.001, Cohen's d=0.47, indicating significant effects of anthropomorphism manipulation. In conclusion, both expectation and anthropomorphism manipulations significantly increased ratings, validating their effectiveness.

4.2.2 Combined Anthropomorphism + Expectation Adjustment Intervention Effects

To examine the effects of different intervention types (control, anthropomorphism, expectation adjustment, combined) on moral response, perceived agency, and perceived experience, we conducted one-way ANCOVAs with gender as a covariate. Gender was selected as a covariate because independent samples t-tests showed females' moral response ratings (M=4.95, SD=1.27) were significantly higher than males' (M=4.56, SD=1.41), t(211)=−2.13, p=0.034, 95% CI [0.029, 0.751], Cohen's d=0.29.

Figure 5 [FIGURE:5] Effects of Different Intervention Strategies on Moral Response Enhancement

For moral response level, descriptive statistics showed the combined intervention group scored 5.64 (SD=0.91), the anthropomorphism group 5.07 (SD=1.21), the expectation adjustment group 4.85 (SD=1.10), and the control group 3.55 (SD=1.22). One-way between-subjects ANCOVA revealed a significant main effect of intervention type, F(3, 208) = 31.18, p < 0.001, ηp² = 0.31. Bonferroni-corrected post-hoc pairwise comparisons further showed the combined group was significantly higher than the control group (Md= 2.07, se = 0.22, p < 0.001) and expectation adjustment group (Md=0.80, se =0.22, p=0.02), and marginally significantly higher than the anthropomorphism group (Md=0.57, se =0.22, p=0.061). The anthropomorphism group was significantly higher than the control group (Md =1.50, se = 0.22, p < 0.001). The expectation adjustment group was significantly higher than the control group (Md =1.27, se =0.22, p <0.001). No significant difference existed between anthropomorphism and expectation adjustment groups (Md=0.23, se =0.22, p=1.00), see Figure 5.

For perceived agency level, descriptive statistics showed the combined group scored 5.20 (SD=0.86), the anthropomorphism group 4.52 (SD=0.99), the expectation adjustment group 4.77 (SD=1.14), and the control group 2.49 (SD=1.09). One-way ANCOVA revealed a significant effect of intervention type on agency ratings, F(3, 208) = 71.97, p < 0.001, ηp² = 0.51. Bonferroni-corrected post-hoc comparisons showed the combined group was significantly higher than the control group (Md= 2.75, se = 0.20, p<0.001) and anthropomorphism group (Md=0.67, se =0.20, p=0.005), but not significantly different from the expectation adjustment group (Md=0.43, se =0.20, p=0.205). The anthropomorphism group was significantly higher than the control group (Md =2.07, se = 0.20, p < 0.001). The expectation adjustment group was significantly higher than the control group (Md =2.32, se =0.20, p< 0.001). No significant difference existed between anthropomorphism and expectation adjustment groups (Md=-0.26, se =0.20, p=1.00), see Figure 6 [FIGURE:6].

Figure 6 [FIGURE:6] Effects of Different Intervention Strategies on Perceived Agency Enhancement

For perceived experience level, descriptive statistics showed the combined group scored 4.66 (SD=1.35), the anthropomorphism group 3.89 (SD=1.33), the expectation adjustment group 3.80 (SD=1.35), and the control group 2.22 (SD=0.85). One-way ANCOVA revealed a significant effect of intervention type on experience ratings, F(3, 208) = 37.15, p < 0.001, ηp² = 0.35. Bonferroni-corrected post-hoc comparisons showed the combined group was significantly higher than the control group (Md=2.50, se =0.24, p<0.001), expectation adjustment group (Md=0.78, se =0.24, p=0.009), and anthropomorphism group (Md=0.85, se =0.24, p=0.003). The anthropomorphism group was significantly higher than the control group (Md=1.72, se =0.24, p<0.001). The expectation adjustment group was significantly higher than the control group (Md=1.65, se =0.24, p<0.001). No significant difference existed between anthropomorphism and expectation adjustment groups (Md=0.07, se =0.24, p=1.00), see Figure 7 [FIGURE:7].

Figure 7 [FIGURE:7] Effects of Different Intervention Strategies on Perceived Experience Enhancement

4.2.3 Path Analysis of Combined Anthropomorphism + Expectation Adjustment Intervention

To examine how different intervention types (control, anthropomorphism, expectation adjustment, combined) affect moral response, this study used intervention group as the independent variable, perceived agency and experience as parallel mediators, and moral response as the dependent variable, analyzed using PROCESS macro (Model 4; Hayes, 2013). Since the independent variable was multi-categorical, we created three dummy variables with the control group as reference: D1 (anthropomorphism vs. control), D2 (expectation adjustment vs. control), D3 (combined vs. control). Analysis used Bootstrap resampling (5,000 iterations) for robust confidence interval estimation. Given gender's significant effect on moral response, it was controlled as a covariate.

Figure 8 [FIGURE:8] Path Model of Combined Anthropomorphism and Expectation Adjustment Intervention Enhancing Moral Responses to Unethical AI Decisions

Combined intervention path model results. Compared to the control group, the combined intervention significantly enhanced perceived agency (b=0.86, se= 0.22, p < 0.001, 95% CI [0.84, 1.69]), which in turn significantly positively predicted moral response (b = 0.18, se = 0.07, p = 0.011, 95% CI [0.04, 0.30]). The indirect effect of perceived agency was significant (indirect effect = 0.21, Boot se= 0.09, 95% CI [0.04, 0.40]). Simultaneously, the combined intervention significantly enhanced perceived experience (b =0.89, se = 0.22, p < 0.001, 95% CI [0.82, 1.80]), which in turn significantly positively predicted moral response (b = 0.32, se = 0.06, p < 0.001, 95% CI [0.16, 0.41]). The indirect effect of perceived experience was significant (indirect effect = 0.39, se = 0.10, 95% Boot CI [0.21, 0.60]). The path model is shown in Figure 8.

4.2 Discussion

Results show the combined anthropomorphism + expectation adjustment group scored highest on moral response, significantly higher than the control group, indicating that enhancing both anthropomorphic perception and expectation adjustment can effectively elevate moral response levels. The combined intervention also significantly enhanced perceived agency and experience, both of which significantly affected moral response as mediators. Notably, the mediation effect of perceived experience (0.29) was substantially larger than that of perceived agency (0.17), suggesting that people's experience perception of AI plays a more critical role in enhancing moral responses. In this study, the combined intervention's effect coefficient on perceived experience reached 0.89, likely because experience perception holds a central position in human-AI interaction. When people view AI as an entity with emotional and perceptual capabilities, they are more inclined to attribute moral qualities to it, thereby eliciting stronger moral responses (Gray et al., 2012). Therefore, through anthropomorphism and expectation adjustment interventions, not only can people's overall moral responses to AI be enhanced, but this effect can be further promoted by strengthening experience perception. In conclusion, this study demonstrates that combined anthropomorphism and expectation adjustment interventions can significantly enhance people's moral responses to AI, with the mediating role of perceived experience being particularly prominent. This finding provides a new perspective for AI ethics research and offers theoretical support for designing more effective interventions to enhance public moral cognition and response to AI. Future research could further explore the effectiveness of these interventions across different cultural contexts and investigate how to apply these theoretical findings to actual AI system design.

5 General Discussion

Integrating mind perception theory and moral dyad theory, this study systematically reveals the dual-pathway mechanism of the AI moral deficiency effect and its mitigation strategies. Results show that compared to human decision-makers, people's moral responses to unethical AI decisions are significantly weaker, with lower perceived agency and experience in AI constituting important psychological causes. Further intervention research demonstrates that a comprehensive intervention combining anthropomorphic strategies targeting AI and expectation adjustment strategies targeting human observers can effectively enhance moral responses to AI's unethical behavior. Notably, unlike "algorithmic ethics" research in computer science, philosophy, law, sociology, and other disciplines that primarily focus on design-level principles and technical paths for fair algorithms, this study adopts a psychological perspective emphasizing differential psychological reactions to AI versus human decision-making. This finding not only provides new theoretical insights for alleviating social problems caused by algorithmic bias and promoting fair algorithm construction, but also opens new research perspectives for "algorithmic ethics."

5.1 Significantly Weaker Moral Responses to AI's Unethical Decisions

Based on Chinese socio-cultural context, this study reveals significantly weaker moral responses to AI's unethical decisions across three typical scenarios: educational discrimination (screening out non-"985" university applicants), age discrimination (35-year-old career threshold algorithms), and gender discrimination (gender-weighted bias in resume screening). Unlike previous research primarily focused on Western individualistic cultures (Bigman et al., 2023), this study's core contribution lies in confirming the effect's robustness within Chinese collectivistic culture. Existing experimental paradigms based on Western individualistic values often inadequately capture moral cognition characteristics in collectivistic cultures. For example, experimental materials rooted in Western value systems (e.g., involving racial discrimination) have obvious limitations in reflecting Chinese society's unique ethical dilemmas. Therefore, developing culturally adapted, locally relevant moral scenario materials to enhance ecological validity is a key prerequisite for exploring the cross-cultural generalizability of the AI moral deficiency effect. This study not only developed such materials but also yielded valuable findings. Specifically, compared to other unethical scenarios, gender discrimination has distinctive features (Wilson et al., 2022). Gender discrimination is a long-standing systemic bias in human society with high global visibility, such as workplace gender pay gaps and female leadership underrepresentation (Dastin, 2022; Heilman et al., 2024; Xiao et al., 2024). Investigating how AI inherits or amplifies such entrenched biases directly reveals technology's "replication–reinforcement" mechanisms in social structures. Moreover, in hiring contexts, gender discrimination is often encoded into algorithms through historical data (e.g., male-dominated tech industry hiring records), causing AI systems to weight female candidates lower (Halzack, 2019). This clear "data–algorithm–outcome" chain facilitates analysis of underlying technical ethics logic. Therefore, Study 2 primarily used gender discrimination as the moral scenario for exploring AI's moral deficiency effect. The series of sub-studies (one correlational and two experimental) confirmed that when the decision-maker is AI, moral response levels to gender-discriminatory decisions are lower than for humans. This series not only confirms the robustness of the AI moral deficiency effect but also uses scenario stripping methods to demonstrate gender discrimination's theoretical advantages as an ideal experimental scenario. Compared to educational or age discrimination, gender discrimination more easily strips away confounding variables (e.g., the relationship between education and ability may be more complex), providing purer variable conditions for mechanism research and intervention experiments.

5.2 Parallel Mediation of Perceived Agency and Experience

By integrating mind perception theory and moral dyad theory, this study systematically reveals the dual-pathway mechanism of the AI moral deficiency effect. Previous research explained this phenomenon through individual psychological pathways, such as free will beliefs (Bigman et al., 2023) or bias motivation (Xu et al., 2022), presenting fragmented mechanistic explanations. This study, starting from the moral agent perspective, explains the mediating roles of agency and experience in the AI moral deficiency effect, representing significant theoretical advancement for both mind perception theory and moral dyad theory. Unlike previous scattered explorations of single variables like free will or autonomy (Bigman et al., 2023; Heinrichs et al., 2022; Xu et al., 2022), this study is the first to confirm parallel mediation of perceived agency and experience in AI decision-making's moral deficiency effect. This finding breaks through the traditional "agency (agent)–experience (patient)" binary opposition framework (Gray et al., 2012). Existing research has explored perceived agency's impact on the AI moral deficiency effect (Hohenstein & Jung, 2020; Wilson et al., 2022; Zhu & Chu, 2025), but our results further indicate that solely emphasizing agency enhancement (e.g., increasing decision transparency) is insufficient; synergistic optimization of the experience dimension (e.g., emotional interaction design) is equally critical. Experimental data show that when AI's experience perception is strengthened, its moral agent evaluation significantly improves, strongly supporting the emerging "emotional rationalism" perspective that experience is not only necessary for moral patients but also an intrinsic component of moral agent qualification (de Vel-Palumbo et al., 2022). This theoretical breakthrough challenges traditional moral dyad theory's narrow positioning of the experience dimension, offering new perspectives for understanding the multifaceted composition of moral agents.

This study provides the first causal evidence of parallel mediation through experimental mediation methods, compensating for previous research's (Sullivan et al., 2022; Xu et al., 2022) limitation of questionnaire methods that could only verify correlations. Structural equation modeling further demonstrates that the joint explanatory power of perceived agency and experience surpasses single pathways (as in Study 2c), deepening understanding of the AI moral deficiency effect's complexity. Specifically, through multi-method validation combining experimental mediation and questionnaire methods, this study achieves causal inference of the AI moral deficiency effect mechanism. The 2×2 experimental design reveals the interaction between decision-maker and mind dimensions: under AI conditions, agency manipulation significantly enhances moral response, while human agents show no such effect. This "agency attribution bias" confirms the core hypothesis of mentalizing theory—that humans are inherently endowed with complete mind schemas (Leyens et al., 2000)—while AI's "quasi-agent" status is constructive and adjustable through mind dimension representation design (Bigman et al., 2023). By manipulating perceived agency/experience (high vs. control) and decision-maker (AI vs. human), we found that in the passive control group, decision-maker significantly predicted agency/experience levels, while AI group agency enhancement significantly improved moral response, constructing a complete causal chain of "agent type → mind perception → moral response." This design overcomes traditional questionnaire methods' inherent causal inference limitations, breaking previous research's correlation-only constraints. Subsequent questionnaire methods simultaneously examined parallel mediation of agency and experience, revealing synergistic gains: in gender discrimination scenarios, the dual-pathway standardized coefficient was 0.398, 95% CI=[0.323,0.467], jointly explaining 39.8% of variance—substantially more than the single-pathway effects (agency: standardized coefficient=0.214, 95% CI=[0.032,0.383]; experience: standardized coefficient=0.184, 95% CI=[0.003,0.373]). This suggests real-world moral judgments may follow dual mechanisms of "intentional action" and "emotional feeling." The theoretical advancement lies in revealing the deep mechanism of AI decision-making moral deficiency—people believe algorithms lack both autonomous intention (low agency) and ability to understand emotional harm (low experience), leading to insufficient moral responses—thereby resolving the fragmentation problem where different studies emphasize different mediators and forming a unified explanatory framework.

5.3 Combined Effects of Anthropomorphism and Expectation Adjustment

This study innovatively proposes and validates a "anthropomorphism + expectation adjustment" combined intervention strategy to reduce the moral deficiency effect by enhancing AI's agency and experience perceptions. Specifically, through randomized controlled experiments, we propose dual intervention strategies exploring how anthropomorphic design and expectation adjustment might mitigate AI's moral deficiency effect. Results show that both single anthropomorphism and expectation adjustment interventions have significant positive effects on enhancing individuals' moral responses and improving perceptions of AI's agency and experience; the synergistic effect of the two strategies is even more pronounced. This finding provides empirical support for constructing systematic intervention solutions.

The anthropomorphic design introduced in this study, by giving AI human-like images, voices, or behavioral characteristics, makes it easier for people to perceive it as having higher mind perception levels, thus viewing it as a "complete agent" with moral responsibility (Lin et al., 2020; Gursoy et al., 2019; Melián-González et al., 2021). Two-way ANOVA results show that anthropomorphism intervention achieved significant effects in enhancing moral response, perceived agency, and experience, with effects further amplified under expectation manipulation. This indicates that anthropomorphic design enables people to have higher moral response levels when facing unethical AI decisions. Moreover, mediation path analysis further reveals that anthropomorphism's enhancement of moral response is achieved by strengthening individuals' perceptions of AI's agency and experience.

In addition to anthropomorphism intervention, expectation adjustment, as another intervention approach, also demonstrates unique effects in eliminating AI decision-making moral deficits (Srinivasan & Sarial-Abi, 2021). Research shows that presetting appropriate expectations for AI behavior not only significantly enhances moral response levels but also validates this intervention mechanism through the mediating effects of perceived agency and experience. Further moderated mediation tests indicate that the synergistic mechanism of anthropomorphism and expectation adjustment primarily relies on dual enhancement of AI's agency and experience perceptions.

Existing research (Bigman et al., 2023; Hu et al., 2024; Xu et al., 2022) has examined the intervention effects of anthropomorphism or expectation adjustment on AI's moral deficiency effect from different perspectives. However, current research rarely simultaneously examines both AI and human intervention angles to investigate the combined effect of anthropomorphism and expectation adjustment. For example, Bigman et al. (2023) and Xu et al. (2022) only focused on anthropomorphism's intervention effects; Hu et al. (2024) noted both anthropomorphism and expectation adjustment through literature review but did not propose a combined intervention plan. This study is the first to propose and examine a dual-pathway intervention from AI design (anthropomorphism) and human cognition (expectation adjustment), experimentally proving that synergistic effects are significantly stronger than single strategies. Through theoretical model innovation of intervention pathways, this research advances from Bigman et al.'s (2023) and Xu et al.'s (2022) single free will or bias motivation explanations to an "agency–experience" dual-pathway model, generating multi-dimensional intervention strategies that more effectively mitigate AI decision-making moral deficiency phenomena and providing psychologically-informed innovative pathways for AI ethics governance.

5.4 Limitations and Future Directions

Despite these contributions, this study has several limitations. First, there are scenario limitations in experimental design. Specifically, Studies 1 and 2 primarily examined discriminatory moral scenarios (gender, educational, age discrimination), while Study 3 focused on autonomous driving moral dilemmas. Although these scenarios have significant social relevance, they only reflect specific types of moral problems. Notably, AI's real-world ethical challenges are more diverse, including privacy protection, life-right trade-offs in medical decision-making, public resource allocation, and other complex situations. Public response patterns to AI decisions may differ significantly across these moral domains. For example, moral dilemmas involving personal safety may elicit stronger emotional responses, while privacy violations may focus more on responsibility attribution. Future research should expand the diversity of moral scenarios to include environmental protection, educational equity, judicial sentencing, and more domains. This expansion would not only more comprehensively define the boundary conditions of AI's moral deficiency effect but also provide richer empirical foundations for constructing more inclusive AI ethics governance frameworks.

Second, there are ecological validity concerns with experimental methods. This study primarily used scenario simulation paradigms, presenting moral decision-making situations through text descriptions and video materials. While this approach offers clear advantages in variable control and experimental manipulation, its ecological validity has limitations. Laboratory responses may not fully reflect complex psychological processes in real scenarios. Real-world moral responses often occur in dynamically changing environments and are influenced by multiple factors including social norms, cultural backgrounds, and personal experiences. For example, in real hiring scenarios, interactions between applicants and AI systems may last weeks, with moral cognition evolving over time—fundamentally different from immediate reactions in laboratories. To improve external validity, future work could: conduct field experiments embedding research in real AI application scenarios like medical diagnostic systems or corporate automated hiring platforms; adopt longitudinal designs to examine cumulative effects of long-term AI exposure on individual moral sensitivity; combine multimodal data collection techniques (e.g., eye-tracking, physiological monitoring) to more comprehensively capture participant response patterns in real contexts. These methodological innovations would significantly enhance the real-world explanatory power of research findings, providing more actionable scientific evidence for AI ethics governance.

Third, this study treated anthropomorphism and expectation adjustment as two independent intervention strategies, proposing potential solutions from AI characteristics and human cognition perspectives respectively. However, theoretically, anthropomorphism as a means of strengthening AI's human-like features may directly influence people's expectations of AI's mind capabilities. Perceived human-like features can facilitate human-AI interaction, prompting people to transfer social heuristic judgments to robot interaction contexts, thereby forming higher expectations (Duffy, 2003; Nass & Moon, 2000). However, despite anthropomorphic design enhancing surface similarity, AI's actual behavior often fails to fully meet user expectations, and this expectation gap may trigger significant negative emotions (Grazzini et al., 2023). Research shows that anthropomorphism can enhance trust in technology, but when behavioral performance falls short of expectations, that trust may transform into stronger disappointment and resistance (Waytz et al., 2014; Crolic et al., 2022). This evidence suggests that anthropomorphism and expectation adjustment intersect in psychological mechanisms and are not completely independent. Therefore, when the two intervention strategies are combined, their pathways of action may be more complex than hypothesized in this study. Future research should examine the long-term effects of these strategies across different individuals and contexts to more comprehensively explain how anthropomorphism and expectation adjustment interact under various conditions, providing more targeted theoretical foundations and practical guidance for AI moral decision-making and human-AI interaction optimization.

In conclusion, this study finds: First, people's moral responses to AI's unethical decisions are significantly weaker than to human decision-makers'. Second, compared to human agents, agency and experience manipulations significantly enhance moral responses in AI decision-making contexts; moreover, perceived agency and experience have parallel mediating effects in the AI moral deficiency effect. Third, both single anthropomorphism intervention and expectation adjustment have significant positive effects on enhancing individuals' moral responses and improving perceptions of AI's agency and experience; the synergistic effect of the two strategies is even stronger.