Under pressure: how widespread vs severe competitor unethical practices shape responsible artificial intelligence deployment Open Access

The integration of artificial intelligence (AI) into digital products and services has created unprecedented opportunities and emerging challenges around ethics, privacy, security, human rights and fairness that managers must navigate to avoid unintended consequences of AI-integrated systems (Berente et al., 2021; Marjanovic et al., 2022). Responsible AI (RAI) practices can deliver value by strengthening fairness perceptions and lowering the incidence of discriminatory recommendations (Ebrahimi et al., 2025). A recent global McKinsey survey of over 750 leaders found that most plan to invest more than 1 million USD in RAI in 2025, with larger firms investing between 25 and 50 million (Luget et al., 2025). Yet, social injustices and harmful outcomes can occur when organizations implement AI systems without adequate governance and processes or prioritize economic gains over user welfare (Berente et al., 2021; Mikalef et al., 2022).

Although demand-side dynamics driven by consumer expectations push organizations to adopt AI systems, supply-side competitive pressures intensify the need for rapid innovation, often sacrificing ethical considerations in the process (Alsheibani et al., 2020; Enholm et al., 2022). These pressures create launch decisions for AI-based digital services and products, where organizations must navigate trade-offs between achieving competitive advantage and ensuring ethical responsibility (Benbya et al., 2021; Kirshner and Lawson, 2025; Lee et al., 2018; Papagiannidis et al., 2025; Wu et al., 2023). Competitive pressures force organizations into rapid design-to-market strategies that bypass thorough testing, creating AI systems that pass internal validation but fail under real-world conditions (Marabelli et al., 2021; Rhue and Washington, 2020). This testing gap significantly increases risks of privacy violations, bias and manipulative behaviors, while AI’s inherent opacity obscures these failures from detection (Wu et al., 2023; Kordzadeh and Ghasemaghaei, 2021; Hinds et al., 2020; Asatiani et al., 2021).

While competitive pressures can increase unethical pro-organizational behavior (UPB), actions that violate ethical norms but benefit the organization (Moore, 2008; Welsh et al., 2015), limited research explores how external competitive environments influence decisions to launch unethical AI. Ethical guidelines provide high-level principles like transparency and fairness (Jobin et al., 2019) but fail to account for competitive trade-offs (Birkstedt et al., 2023), such as prioritizing speed-to-market over thorough testing or accuracy over explainability (Liang et al., 2021; Langer et al., 2021). We seek to understand how competitive pressure influences decisions to launch unethical AI by examining two distinct forms of competition. First, firms may encounter horizontal competitive pressure, where many competitors deploy unethical AI systems, creating industry norms of lower ethical standards. Second, organizations may face vertical competitive pressure, where a key competitor deploys an extremely unethical AI system, establishing a new unethical floor. We posit that these competitive conditions can activate moral disengagement (Bandura, 1991). Horizontal unethical AI pressure from multiple competitors can dilute individual accountability, a process known as diffusion of responsibility (Bandura, 1991; Moore, 2008). Vertical unethical AI pressure can trigger advantageous comparison, where actions are justified by contrasting them with more egregious violations (Bandura, 1991).

As organizational goals can shape responses to external competitive forces, we employ regulatory focus theory (Higgins, 1998). Promotion-focused organizations, driven by growth and achievement, are more prone to trade-offs that compromise RAI (Kirshner and Lawson, 2025) and may use advantageous comparison with competitors’ more severe violations to justify launching unethical AI. In contrast, prevention-focused firms, motivated by security and avoiding losses, may succumb to horizontal unethical pressure through diffusion of responsibility when industry practices threaten their competitive position. This suggests that firm goals may moderate how different competitive conditions impact AI launch decisions. Thus, we investigate the following two research questions:

We conducted three experiments to examine how competitive pressures influence decisions to launch unethical AI. In all three studies, we manipulated competitive conditions to test the effect of horizontal and vertical pressure. Our results reveal that horizontal competitive pressure consistently increases unethical AI launches. Surprisingly, vertical competitive pressure interacted with prevention-focused goals, decreasing the likelihood of launching unethical AI compared to promotion-focused goals. We discuss how these findings respond to calls for exploring the dynamic nature of ethical AI implementation (Birkstedt et al., 2023; Li and Wang, 2023) and bridge the gap between high-level ethical principles and practical governance mechanisms (Morley et al., 2020), demonstrating how competitive environments and organizational orientations can shape RAI deployment.

To investigate the role of competitive environments on ethical AI decision making, we leverage moral disengagement theory. Moral disengagement is a context-dependent, malleable phenomenon shaped by social influences and organizational environments (Moore, 2008). Moral disengagement theory identifies eight mechanisms for rationalizing immoral behavior (Bandura et al., 1996), including behavioral activities (moral justification, euphemistic labelling and advantageous comparison), agency factors (diffusion of responsibility and displacement of responsibility), outcomes (distortion of consequences) and victim characteristics (attribution of blame and dehumanization). As downplaying harm is a function of moral disengagement, firms can take advantage of moral disengagement to implement unethical processes, pre-empting employee’s cognitive dissonance where immoral behaviors may contradict their values (Moore, 2008; Bandura, 1991).

Individual decision-making is shaped by motivations and context, influencing goal setting and actions, where both failure and success drive motivation (Van-Dijk and Kluger, 2004). Regulatory Focus Theory reconciles these views, distinguishing between promotion-focused individuals, driven by success and prevention-focused individuals, motivated by avoiding failure (Higgins, 1998). Promotion focus is associated with risk-taking, creativity and innovation, often incentivized by rewards such as bonuses (Johnson et al., 2017). In contrast, prevention focus prioritizes stability and security, aiming to minimize losses and maintain the status quo.

Regulatory focus can affect decision-making through its relationship with risk (Crowe and Higgins, 1997; Gino and Margolis, 2011). Promotion-focused individuals pursue ambitious goals, fostering exploration, growth and innovation, but also by engaging in riskier behaviors (Higgins, 2002). Prevention-focused individuals adopt cautious strategies, reducing risk and prioritizing ethical behavior through vigilance (Higgins, 2002; Gino and Margolis, 2011). At the firm level, regulatory focus affects employee behaviors and ethical outcomes, as individual tendencies, such as extraversion and neuroticism, can translate into professional orientations (Lanaj et al., 2012). A match between employee focus and organizational environment enhances engagement and moral reactions (Achar and Lee, 2022). Industry competition, leadership and organizational culture also shape focus, with promotion focus driving innovation and citizenship behaviors, while prevention focus supports safety and compliance (Lanaj et al., 2012). Recently, Regulatory Focus has found extensive application in IS, offering valuable insights into user behavior and digital platform engagement (e.g. Lwin et al., 2016; Lee et al., 2019; Wu et al., 2019; Zhou et al., 2024). Most recently, Kirshner and Lawson (2025) found support that promotion-focused behaviors can lead to increased AI-based UPB, by influencing decision-maker’s RAI values. The current work extends their findings by integrating regulatory focus with moral disengagement to examine how competitive pressures interact with organizational goals to impact AI-based UPB decisions.

Organizations deploying AI systems face competitive pressures that can affect ethical decision-making. We examine how two ethical competitive configurations activate distinct moral disengagement mechanisms to influence decisions to launch potentially unethical AI systems. First, when many competitors adopt similar unethical AI practices, we posit that this widespread adoption activates diffusion of responsibility. Decision-makers may experience reduced personal obligation to maintain ethical standards as the behavior becomes normalized (Bandura, 1991). Second, when a focal competitor implements extremely unethical AI practices, we posit that this scenario will activate advantageous comparison (Bandura, 1991), where individuals justify deploying moderately unethical AI by contrasting it against more severe misconduct.

We propose that competition activating these moral disengagement mechanisms predicts how external environments influence AI deployment decisions. Under horizontal pressure, decision-makers may rationalize deployment, since “everyone else is doing it” and delaying the launch while competitors gain market advantages would be strategically detrimental. Similarly, under vertical unethical pressure, decision-makers may launch a moderately unethical AI as a “lesser evil” while maintaining competitive positioning and appearing more ethical than their focal competitor. Based on these mechanisms, we hypothesize that different competitive environments will systematically affect AI deployment decisions as follows:

Kirshner and Lawson (2025) explored the relevance of promotion and prevention-based organizational goals for AI-based UPB, and found that promotion-focused goals lead to lower levels of AI-based UPB. We expand on Kirshner and Lawson (2025) by positing a relationship between the competitive environments and regulatory focus on AI deployment decisions. Prevention-focused goals and competitive threats: Prevention-focused individuals tend to be risk-avoiders, focusing on minimizing losses through security and stability, and are conservative in their efforts to avoid adverse outcomes (Higgins, 2002). When many competitors deploy unethical AI systems, this creates a direct competitive threat to the organization’s market position. Prevention-focused decision-makers may perceive this widespread adoption as creating a competitive disadvantage that threatens organizational survival and their own job security, given their motivation from resilience and the fear of failure, prioritizing the security of employment and maintaining the status quo (Johnson et al., 2017; Van-Dijk and Kluger, 2004). This fear of falling behind competitively may activate diffusion of responsibility, allowing them to justify deploying similar unethical AI by reasoning that “everyone else is doing it” and that maintaining competitive parity is necessary to avoid organizational losses. The industry-wide practice becomes normalized, reducing individual accountability for the ethical implications.

Promotion-focused individuals approach tasks eagerly, centering on achievements and positive outcomes (Higgins et al., 2003), and have greater risk tendencies, focusing on maximizing success by innovative and creative means, leading to unethical behavior (Gino and Margolis, 2011). Accordingly, we propose that promotion-focused goals provide an opportunity for individuals to take advantage of a primary competitor implementing an AI system with worse consequences to engage in advantageous comparison and justify an unethical AI action. Put differently, individuals may increase the likelihood of implementing moderately unethical AI when competitors implement even more unethical AI. Therefore, a scenario where an organization can implement a new AI and competitors are implementing worse AI may create a natural regulatory fit for a promotion-focused individual to leverage advantageous comparison, leading to AI-based UPB.

The complete research model and hypotheses are presented in Figure 1.

To explore how external competitive pressures impact AI-based UPB, Study 1 examined the relationship between individuals’ regulatory focus orientation and behaviors in a competitive scenario characterized either by a high number of unethical competitors (Horizontal) or by a severely unethical competitor (Vertical). Utilizing a between-subjects design with three treatments (Control vs Horizontal vs Vertical), we manipulated the potential for moral disengagement to test H1a and H1b and measured participants underlying regulatory focus to explore potential moderating effects of regulatory focus, testing H2a and H2b.

We recruited American MTurkers, compensating them with 2 USD for completing the study. We screened ^[1] participants and used attention checks to ensure quality engagement from participants. The sample size was 240 participants, which is above GPower’s recommendation (d = 0.5, alpha = 0.05, and 1-beta = 0.8) due to measuring regulatory focus. In total, 249 participants (females: 111; other: 2; age: M = 44.04, SD = 11.03) completed the experiment.

Participants indicated how strongly they agreed with eight promotion and prevention-focused statements adapted from Lockwood et al. (2002) to measure regulatory focus. We discussed that firms use AI to undertake targeted marketing and personalized pricing by tracking consumer activity and spending habits from previous interactions and third-party data. Participants were asked to imagine they worked for a firm that sells tourism services and reported directly to the Chief Marketing Officer (CMO), who planned to introduce AI into the organization. This scenario and attention check questions were taken from Kirshner and Lawson (2025) ^[2]. Participants answered two comprehension-check questions, and if they passed, they were asked to decide on a proposed AI implementation, addressing the question: “Should the firm use AI to target marketing and alter pricing of services based on data collected from their consumers based on previous interactions and data about their consumers purchased from third parties?”.

Participants were then allocated to one of the three conditions: Control, Horizontal or Vertical. All participants read the Control condition, reiterating that AI can track consumer activity through cookies and that the AI analyzes the data from previous interactions and data purchased from third parties for target marketing to improve profits, but it may have potential ethical issues. Participants in the Horizontal treatment were told (in addition to the Control group information) that 95% of their firm’s competitors use AI to target marketing and alter pricing using data from previous interactions and purchases from third parties. We noted that this was the same strategy as the CMO’s proposal (see Table A1). Participants in the Vertical treatment were told that the firm’s main competitor used AI for target marketing and altering pricing using data from previous interactions and data purchased from third parties, which is the same strategy as the CMO’s proposal. (This is the same as in the Horizontal treatment, but with one firm.) In addition, the competitor accessed microphones on consumer devices to collect data and sold this data without consent from consumers, which we emphasized is highly unethical.

After exposure to the treatment statements, participants indicated whether to implement the AI on a 7-point Likert scale (1 = “Strongly disagree” to 7 = “Strongly agree”). This was in accordance with the scale as used by Lee et al. (2018) to avoid participant fatigue. Participants in the moral disengagement treatments were asked a manipulation check question. We asked what percentage of the firm’s competitors disregard consumer privacy to undertake targeted marketing and personalized pricing in the Horizontal group, which was measured on a 5-point scale from “Much lower than average” to “Much higher than average.” The Vertical group was asked how they would rate the ethicalness of the main competitor compared to the CMO’s proposed action on a 5-point scale from “Much lower” to “Much higher.” Participants then reported age and gender.

Participants in the Horizontal group perceived the number of the firm’s competitors using the same AI as higher than average (M = 4.60, SD = 0.92, t(80) = 15.74, 95% CI: [4.40, 4.81], p < 0.001), and participants in the Vertical group perceived the ethicalness of the main competitor compared to the CMO’s proposal as being lower (M = 1.94, SD = 1.05, t(83) = −9.29, 95% CI: [1.71, 2.17], p < 0.001). Therefore, both competition manipulations were successful in creating moral disengagement scenarios. The responses were internally consistent across the four promotion-focused measures (Cronbach’s alpha = 0.84) and the four prevention-focused measures (Cronbach alpha = 0.85). Overall, participants were slightly more promotion-focused (M = 5.68, SD = 0.99, t(248) = 26.75, 95% CI: [5.56, 5.81], p < 0.001) than prevention-focused (M = 3.79, SD = 1.39, t(248) = −2.44, 95% CI: [3.61, 3.96], p = 0.016). We deducted participants’ prevention-focus scores from their promotion-focus scores to evaluate regulatory focus (Lockwood et al., 2002).

To examine how participants’ competitive context (Vertical and Horizontal) and regulatory focus influenced RAI decision-making, we used Hayes’ PROCESS Model 1 with indicator coding for the multi-categorical treatment variable (Hayes and Montoya, 2017). The dependent variable was participants’ decision on whether the firm should implement the AI system, reverse-coded so that higher values indicated the more ethical choice. The moderator regulatory focus was mean-centered and conditional effects were estimated at low (−1 SD; 16th percentile), medium (0 SD, 50th percentile) and high (+1 SD; 84th percentile) values of regulatory focus.

Table 1 presents the estimates of the overall model, which was significant (F(5, 243) = 3.97, p = 0.002). At the mean regulatory focus level, participants in the Horizontal group were significantly less likely to make ethical decisions than those in the Control group (b = −0.70, SE = 0.31, p = 0.023), supporting H1a. The difference between the Vertical and Control conditions at the mean level of regulatory focus was not significant (b = 0.42, SE = 0.30, p = 0.165), providing no support for H1b. The analysis above addresses RQ1.

We next tested whether regulatory focus moderated the effects of treatment condition. The interaction between regulatory focus and Horizontal was not significant (b = 0.04, SE = 0.15, p = 0.772), providing no support for H2a. However, the interaction between regulatory focus and Vertical was marginally significant (b = −0.30, SE = 0.17, p = 0.070). Probing this interaction revealed that participants low in regulatory focus (−1 SD) were significantly more likely to make ethical decisions in the Vertical condition compared to the Control group (b = 1.00, SE = 0.44, p = 0.024). At high levels of regulatory focus (+1 SD), there was no significant difference between Vertical and Control (b = −0.22, SE = 0.46, p = 0.640). These results support H2b and suggest that participants lower in promotion focus were more sensitive to a competitor’s severely unethical behavior and responded by making more responsible AI recommendations. The analysis above addresses RQ2.

We then examined if regulatory focus moderated the effects of competitive context on launching the AI. The interaction between regulatory focus and Horizontal was not significant (b = 0.04, SE = 0.15, p = 0.772), providing no support for H2a. However, the interaction between regulatory focus and Vertical was marginally significant (b = −0.30, SE = 0.17, p = 0.070). Contrary to the direction predicted by H2b, probing this interaction revealed that participants low in regulatory focus were significantly more likely to make ethical decisions in the Vertical condition compared to the Control group (b = 1.00, SE = 0.44, p = 0.024, 95% CI [0.13, 1.86]; see Table 2). This effect was not significant at the mean level of regulatory focus (b = 0.32, SE = 0.31, p = 0.307), and reversed but remained nonsignificant at high levels of regulatory focus (b = −0.22, SE = 0.46, p = 0.640). These findings suggest that participants lower in promotion focus (i.e. more prevention-focused) responded more strongly to the presence of a severely unethical competitor by making more responsible AI decisions.

These observations are supported by Figure 2, which displays the predicted ethical decision scores across the range of regulatory focus levels for each treatment group. Participants in the Vertical condition made the most ethical decisions when their regulatory focus was low, but this effect diminished as regulatory focus increased. In contrast, participants in the Horizontal condition consistently made less ethical decisions regardless of regulatory orientation. Overall, these results provide support for H1a, showing that widespread unethical competition undermines ethical decision-making, while the pattern of interaction observed for H2b runs counter to expectations, indicating that prevention-focused individuals were more responsibly responsive to the ethical severity of a focal competitor.

Study 2 tested the robustness of our findings in Study 1. Moreover, we explicitly tested the interaction effects by manipulating the regulatory focus goal of an imagined employee and organization. Thus, the study is a 2 (regulatory focus: promotion vs prevention-focused goals) x 3 (competitive environment: Control vs Horizontal vs Vertical) between-subjects design to test the effects of competition on RAI launch decisions.

We recruited American MTurkers, incentivizing them with 1.85 USD for completing the study. As in the previous studies, we screened participants using Qualtrics’ fraud direction services and incorporated attention-check questions. We aimed to recruit 50 participants per treatment as the sample size based on GPower’s recommendation for an effect size of d = 0.5, alpha = 0.05, and 1-beta = 0.8. In total, 304 participants (females: 118; other: 3; age: M = 39.13, SD = 11.16) completed the study.

The procedure resembled Study 1 except that we introduced a character named “Alex”, who works for a company called MarkTech and reports to the CMO ^[3]. The survey began with an explanation of black box AI, like Study 1, with the same checks. Participants were randomly assigned to a promotion or prevention-focused manipulation that gave descriptions of Alex’s goal orientation as being either a promotion-focused or prevention-focused employee (see Table A2). We asked a manipulation check for whether the participant believed “Alex is someone who is primarily striving to become their ought or ideal self?” Participants were then given a description of MarkTech as either a promotion-focused or prevention-focused organization and were asked a similar manipulation check question (i.e. “Do you think MarkTech primarily strives on achieving gains or preventing losses?”).

We utilized the competitive environment manipulations from Study 1 for the Control, Horizontal, and Vertical groups. The same AI-based UPB question was asked to participants, but regarding what Alex should recommend: “Should Alex recommend that MarkTech use black box AI for targeted marketing and personalized pricing of products and services based on data collected from their consumers from previous interactions and data about their consumers purchased from third parties to increase profits and build technological capabilities?”. Finally, we asked participants for demographic information.

We first tested the success of the regulatory focus manipulations. Participants rated Alex as more focused on achieving ideal goals (M = 4.45, SD = 2.04) in the promotion-focused group, which was significantly higher than the scale midpoint of 4, t(158) = 2.80, p = 0.006, 95% CI [4.13, 4.77]. In contrast, participants in the prevention-focused condition rated Alex as more focused on avoiding undesired outcomes (M = 3.03, SD = 1.93), significantly below the midpoint, t(144) = −6.02, p < 0.001, 95% CI [2.72, 3.35]. Participants also rated MarkTech as more promotion-focused in the promotion condition (M = 6.02, SD = 0.88), t(158) = 28.86, p < 0.001, 95% CI [5.88, 6.16], and more prevention-focused in the prevention condition (M = 2.61, SD = 1.88), t(144) = −8.94, p < 0.001, 95% CI [2.30, 2.91]. These results confirm the success of the regulatory focus manipulations.

We first conducted a 2 × 3 factorial ANOVA to examine the main and interaction effects. The dependent variable was participants’ recommendation for whether Alex should advise MarkTech to implement the AI system, reverse-coded so that higher values indicated more ethical (i.e. more cautious and responsible) choices. The ANOVA revealed a significant main effect of regulatory focus, F(1, 298) = 33.47, p < 0.001, a significant main effect of competitive condition, F(2, 298) = 6.42, p = 0.002 and a significant interaction between regulatory focus and competitive condition, F(2, 298) = 4.05, p = 0.018.

To further examine the interaction and test our specific hypotheses, we used Hayes’ PROCESS Model 1 (see Table 3). The moderator regulatory focus was coded as promotion-focused (0) versus prevention-focused (1), with conditional effects estimated for both regulatory focus orientations. The overall PROCESS model was significant, F(5, 298) = 10.88, p < 0.001. At the promotion-focused level of regulatory focus, participants in the Vertical condition were directionally, but not significantly less likely to make ethical decisions than those in the Control group (b = −0.35, SE = 0.34, p = 0.305), providing little support for H1b. However, the difference between Horizontal and Control at the prevention-focused level was moderately significant (b = −0.61, SE = 0.32, p = 0.057), providing support ^[4] for H1a.

We next tested whether regulatory focus moderated the effects of the treatment. Table 3 shows that the interaction between regulatory focus and the Horizontal group was not significant (b = 0.06, SE = 0.47, p = 0.894), while the interaction between regulatory focus and Vertical was (b = 1.20, SE = 0.47, p = 0.012). Probing this interaction revealed that participants in the prevention-focused frame were significantly more likely to make ethical decisions in the Vertical condition compared to the Control group (b = 0.85, SE = 0.33, p = 0.011, 95% CI [0.20, 1.51]; see Table 4), but the effect was not significant for promotion-focused participants (b = −0.34, SE = 0.34, p = 0.305).

Thus, overall, the pattern of results in Study 2 largely replicates and extends the findings from Study 1. Prevention-focused participants demonstrated greater ethical sensitivity when faced with a severely unethical competitor, making significantly more responsible AI decisions in the Vertical condition compared to the Control. In contrast, promotion-focused participants showed no significant difference in ethical decision-making across conditions. These findings provide additional support that regulatory focus can moderate responses to competitive ethical contexts. These results are presented in Figure 3.

Study 3 tested whether the Horizontal and Vertical competition treatments used in Studies 1 and 2 activated moral disengagement. Specifically, we assessed whether exposure to a high number of unethical competitors and whether exposure to a severely unethical competitor increased perceptions of unethicalness. This study helps validate the theoretical interpretation that these treatments encourage unethical decisions by enabling moral disengagement. Thus, this study tests the following hypothesis:

We recruited 160 British participants through Prolific and paid each participant 1.50 GBP, based on a target sample size of approximately 50 participants per treatment ^[5]. Participants were 18 years or older and fluent in English. Attention checks were embedded throughout the survey to ensure data quality and one participant who completed the study was dropped because Qualtrics flagged them as a duplicate, leaving a sample of 159 (females: 69; other: 1; age: M = 43.29, SD = 13.58).

Participants read the same AI implementation scenario from Study 2, in which MarkTech was considering adopting a black box AI system for targeted marketing and personalized pricing. The system was described as potentially increasing profits, but raising ethical concerns due to its use of consumer data. As in Studies 1 and 2, participants were randomly assigned to one of three treatment conditions. In the Control condition, no information was provided about competitors. In the Horizontal condition, participants read that 90% of competing firms had already implemented similar AI systems. In the Vertical condition, participants read that the focal competitor had implemented the AI and additionally accessed consumer device microphones to collect sensitive data. After reading the materials, participants were told that Alex had recommended implementing the AI. They then evaluated this recommendation using three items on 7-point scales assessing ethicality, justice and moral acceptability (following Reidenbach and Robin, 1990).

Participants then completed nine accountability items designed to assess whether the treatment conditions influenced how responsible Alex was perceived to be for potential ethical consequences. Each participant responded to three items per condition (Control, Horizontal, Vertical), presented in a fixed order. The Control items assessed baseline moral accountability (e.g. “If no other companies were using this type of AI, then Alex should feel accountable for any ethical issues that arise”). The Horizontal items reflected diffusion of responsibility (e.g. “If many competitors were already using this AI, Alex should feel personally responsible for any unfair outcomes”), and the Vertical items reflected advantageous comparison (e.g. “If other firms are doing worse things than using this AI, then Alex should feel accountable for any ethical issues that arise”). Responses were made on 7-point Likert scales (1 = “Strongly disagree” to 7 = “Strongly agree”). Lower agreement with these accountability statements in the Horizontal and Vertical conditions indicated reduced perceived responsibility, consistent with disengagement reasoning. While this study included both ethical evaluation and accountability measures, its primary aim was to test whether the competition manipulations reduced perceived ethical responsibility through the mechanisms theorized to underpin moral disengagement.

We first examined the internal consistency of the three ethical evaluation items (ethicality, justice and moral acceptability). Responses were highly consistent (Cronbach alpha = 0.93), and we averaged them to form a composite ethical decision score. We then conducted an ANOVA analysis with the treatment conditions (Control, Horizontal, Vertical) as the independent variable. The effect of treatment was significant, F(2, 156) = 4.68, p = 0.011, n² = 0.06. Tukey’s HSD revealed that participants in the Horizontal condition evaluated Alex’s recommendation as significantly less ethical than participants in the Control condition (Difference = 0.86, 95% CI [0.19, 1.52], p = 0.007). The difference between the Vertical and Control condition was not significant (Difference = 0.45, 95% CI [−0.21, 1.11], p = 0.241), nor was the difference between the Vertical and Horizontal conditions (Difference = −0.40, 95% CI [−1.06, 0.26], p = 0.323).

We then assessed whether the Horizontal and Vertical manipulations activated moral disengagement mechanisms. Participants completed nine moral disengagement items (three per condition), where reliability was high for the Control (Cronbach alpha = 0.85), Horizontal (Cronbach alpha = 0.81) and Vertical items (Cronbach alpha = 0.79). Paired-samples t-tests revealed that participants expressed less moral accountability in the Horizontal condition than in the Control condition, t(158) = −9.55, p < 0.001, 95% CI [−1.00, −0.66], Difference = −0.83. They also expressed less accountability in the Vertical condition than in the Control, t(158) = −10.24, p < 0.001, 95% CI [−1.15, −0.78], Difference = −0.96. Figure 4 shows the results for both the between-subject comparisons on ethical evaluations and the within-subject responses to accountability statements reflecting the moral disengagement logic. Although participants in the Vertical condition did not judge Alex’s recommendation as significantly more or less ethical than in the Control condition, they nevertheless held Alex significantly less personally accountable for potential ethical consequences. This is consistent with the logic of advantageous comparison, where unethical actions are less problematic when compared to worse alternatives. Similarly, participants in the Horizontal condition both rated the decision as significantly less ethical and reported lower personal accountability, consistent with diffusion of responsibility.

To formally evaluate H3, we tested if the effect of treatment condition on ethicality was mediated by participants’ perceived responsibility. Using PROCESS Model 4 with 5,000 bootstrap samples (see Table 5), the treatment was a multi-categorical independent variable, with the mediator matched to the moral accountability associated with the treatment condition, and the ethicality of the decision being the dependent variable. Relative to the control condition, the Horizontal treatment significantly reduced perceived responsibility (b = −1.01, SE = 0.26, p < 0.001, 95% CI [−1.51, −0.50]), as did the Vertical treatment (b = −0.80, SE = 0.26, p = 0.002, 95% CI [−1.30, −0.30]). In turn, perceived responsibility negatively predicted ethicality (b = −0.32, SE = 0.08, p < 0.001, 95% CI [−0.48, −0.15]). The indirect effect of the Horizontal treatment on ethicality via responsibility was significant (b = 0.32, SE = 0.13, 95% CI [0.10, 0.59]), as was the indirect effect for the Vertical treatment (b = 0.25, SE = 0.12, 95% CI [0.06, 0.52]). Neither the direct effect of the Horizontal treatment on ethicality (controlling for the mediator) (b = 0.54, SE = 0.28, p = 0.059, 95% CI [−0.02, 1.09]) nor the direct effect of the Vertical treatment were significant (b = −0.20, SE = 0.28, p = 0.472, 95% CI [−0.35, 0.75]). This pattern provides strong support for H3, demonstrating that competitive pressure operates through the theorized moral disengagement mechanisms. These findings validate that the Horizontal and Vertical manipulations activate distinct moral disengagement mechanisms underlying the behavioral effects observed in Studies 1 and 2.

While RAI literature has made strides in examining internal organizational factors related to accountability, transparency, governance structures and leadership policies (e.g. Papagiannidis et al., 2025), it has overlooked the role of competition in shaping RAI decision-making. We address calls for more precise conceptualization of how external pressures influence ethical decision-making in AI contexts (Birkstedt et al., 2023) by offering the first systematic study of how competitive environments influence organizational choices related to RAI.

Drawing parallels to economic product differentiation ^[6], we distinguish between horizontal unethical competition, which arises from pressure to match direct competitors engaging in equally unethical AI practices, and vertical unethical competition, which stems from a competitor engaging in severely unethical AI practices. By theorizing these forms of unethical competition, we advance the understanding of how external market forces condition the implementation and sustainability of RAI initiatives. Thus, existing RAI decision-making models (e.g. Papagiannidis et al., 2025) should integrate external competitive factors to better reflect firm challenges. By incorporating the dynamics of horizontal and vertical competition with regulatory focus, our study shows how external pressures can undermine or reshape internal commitments to ethical AI.

Our empirical findings also reveal distinct and robust effects of competitive pressures on RAI deployment. Horizontal unethical competition undermines ethical behavior, fostering a race to the bottom that normalizes irresponsible AI deployment. Contrasting our initial theorization predicting a uniformly negative effect on the launch of unethical AI, the prevention-focused orientation can dampen or even reverse this effect. Our counterintuitive finding can be explained through the interaction of ethical ambiguity and slippery slope mechanisms. Mittelstadt (2019) argued that RAI principles are inherently abstract and vague, making it difficult to draw clear ethical lines in deployment decisions. This baseline ambiguity allows decision-makers to rationalize launching potentially unethical AI systems, as the ethical boundaries remain unclear. Welsh et al. (2015) showed that people become more morally disengaged when potential unethicalness develops gradually rather than abruptly, as small ethical transgressions pave the way for larger future violations. However, ethical risks presented abruptly, e.g. when exposed to severe competitor unethicalness like accessing consumer microphones without consent, can disrupt the moral disengagement process by making the potential endpoint of unethical behavior vividly apparent.

Importantly, Welsh et al. (2015) showed that regulatory focus moderates this slippery slope effect, with prevention-focused individuals being particularly motivated to avoid negative ethical trajectories once they become salient. The severe competitor behavior thus serves as an abrupt ethical clarification that breaks through the baseline ambiguity, activating prevention-focused individuals’ core motivation to step back from the ethical edge entirely rather than risk sliding down the slippery slope. In contrast, promotion-focused individuals, even when seeing the severe competitor behavior, remain more willing to take calculated risks and may still engage in advantageous comparison reasoning. By demonstrating that competition can activate different moral disengagement pathways and lead to different decision-making depending on goals, we also respond to Achar and Lee’s (2022) call for studying disengagement mechanisms individually.

Our research highlights that firms should analyze competitors’ AI use and be cognizant of how moral disengagement tactics can influence decision-makers to justify implementing unethical AI. Comparisons to unethical competitors may provide prevention-focused firms with motivation for implementing RAI. However, as more organizations implement unethical AI, horizontal unethical competition scenarios may increase and RAI could become a competitive advantage as more competitors engage in AI-based UPB. Firms wishing to improve RAI decision-making by emphasizing prevention-focused objectives must also heed slippery slope effects (Welsh et al., 2015) when comparing against unethical competitors. Managers can also respond by establishing competitor intelligence functions that track both technical and ethical choices, running ethics stress-tests of potential launch decisions and integrating RAI criteria into product approval gates. Managers can also reinforce these measures by mandating regular reviews of how competitive pressures are influencing AI deployment decisions.

Technologically advanced firms implementing new and ethically questionable AI products and services may create vertical unethical competition scenarios in their respective industry. Whereas prevention-focused firms may resist engaging in similar behaviors, promotion-focused firms can leverage this as an opportunity to maximize their own success through irresponsible AI deployment. However, a closer analysis of our results indicates that the benefits of prevention-focused goals may only serve to slow the adoption of industry-wide unethical AI. Over time, promotion-focused firms will likely implement unethical practices in advantageous comparison situations. Eventually, the number of firms engaging in these practices may cross a critical threshold, transforming the market from vertical to horizontal unethical competition. As more promotion-focused competitors implement unethical AI systems, horizontal unethical competition may become the dominant method for unethical AI deployment, driving prevention-focused firms to adopt unethical AI practice to protect their competitive position.

To guard against this erosion, managers can take several steps. Firstly, managers can embed RAI checkpoints into launch processes, including independent review of privacy, fairness and explainability before release. Secondly, managers can invest in transparency mechanisms (e.g. public audits and model documentation) to build trust with regulators and customers. Thirdly, joining industry alliances or standard-setting bodies may prevent a race to the bottom by making responsible practices part of the competitive baseline. If firms neglect these steps, market dynamics risk eroding ethical standards across an entire industry. By contrast, early investment in RAI practices can protect reputation, attract ethically minded customers and employees, and reduce the likelihood of costly regulatory intervention. These effects highlight the need for swift organizational, policy and legislative actions. Delaying responses to unethical AI systems allows for more diffusion of responsibility, opportunities to arise, rendering interventions too late.

Our exploration centered on two types of competitive scenarios that can activate specific moral disengagement mechanisms. Our results indicate that these competitive scenarios impact RAI decision-making through distinct pathways, with severely unethical competitors unexpectedly motivating more responsible decisions under prevention-focused conditions. We hope future research formally examines the mechanisms underpinning this surprising result. Moreover, this result highlights the value of investigating moral disengagement mechanisms individually since it may expose novel effects across applications and generate interesting interactions with regulatory focus and other relevant behavioral-based theories.

We focused on diffusion of responsibility and advantageous comparison as they are the mechanisms most directly activated by competitive environments. While competitive dynamics are likely to manifest through these mechanisms, the other six moral disengagement mechanisms could also influence AI-based UPB through organizational, cultural or individual-level antecedents, representing avenues for future investigation. Future research should also explore how organizations may systematically leverage different moral disengagement mechanisms to justify AI-based UPB, potentially examining how organizational cultures, leadership practices or industry norms activate specific disengagement pathways beyond competitive dynamics. The studies used vignettes and online samples, creating limitations regarding external validity. While online samples reduce stakes and protect anonymity, therefore are appropriate when investigating ethical dilemmas, future research could enhance external validity by replicating these findings with professionals in relevant fields and contexts.

We thank the Special Issue editors and the anonymous review teams whose feedback substantially strengthened this manuscript. We are grateful to the UNSW Business School and BIZLAB for financial support for the experiments. We also thank seminar participants at University College Cork and the University of Houston for their valuable feedback, with special acknowledgements to Wynne Chin for insightful suggestions that informed the additional study designs conducted after the University of Houston seminar.