Machines in the Middle: Using Artificial Intelligence (AI) While Offering Help Affects Warmth, Felt Obligations, and Reciprocity

A central proposition of social exchange theory is that interdependent and contingent interpersonal interactions generate obligations toward others (Emerson, 1976). In work settings, employees establish unique exchange relationships with coworkers (such as subordinates, colleagues, and superiors), and individuals feel obligated to return the benefits they receive from others (Cropanzano & Mitchell, 2005). This compulsion is created by the norm of reciprocity, which encompasses the expected or “rule-based” behavior of reciprocating toward other social participants (Blau, 1964; Gouldner, 1960). The compulsion to reciprocate (i.e., “repay”) emerges when individuals, having internalized the norm of reciprocity as a moral standard, respond to the obligation created by previously received benefits (Korsgaard et al., 2010; Perugini et al., 2003). The concept of felt obligations has been frequently cited as a primary factor connecting received help to enacted or reciprocated help (Ehrhart, 2018; Lester et al., 2008; Organ, 1990). Eisenberger et al. (2001) applied social exchange theory in explaining employee–employer dynamics, suggesting that an employee’s felt obligation to their organization mediates the positive link between perceived organizational support and employee outcomes (like affective commitment and in-role performance). Following Eisenberger’s research, numerous studies employing social exchange theory concentrated on felt obligation as a key variable in the social exchange process between leaders and followers (Duan et al., 2022; Garba et al., 2018; Pan et al., 2012). More recently, studies have examined felt obligations between employees and their coworkers (Dierdorff & Rubin, 2022). As stated in Dierdorff and Rubin (2022), “the theoretical supposition that the norm of reciprocity is a causal mechanism underlying OCB is nearly axiomatic in extant scholarship” (p. 263).

How might using AI affect the reciprocity that typically develops between humans? To reason about its likely impact on recipients’ reactions, we integrate social exchange with theories of technology-mediated interactions in relational building (Duck, 1985; Taylor et al., 2022; Tong & Walther, 2011). These theories indicate that people monitor and evaluate discretionary investments from others in technology-mediated interpersonal contexts to discern whether someone’s behavior is directed at providing a fair and equitable relationship, even if the individuals have little experience with one another (Sigman, 1991). Moreover, investments in technology-mediated interaction contexts include one’s choice of medium as well as communication amount and frequency (Sosik & Bazarova, 2014). A common finding is that individuals using “lightweight” tools that minimize the resource costs of engaging with others are perceived less positively compared to those using more intensive tools and investing more effort (Walther, 2011). This perception stems partly from the fact that technology-mediated interaction is not merely conveyed by the technology itself but can also be modified, augmented, or even created by computational agents to achieve one’s personal or relational goals (Hancock et al., 2020). For example, it takes less time to click the “like” button than to compose and post a comment on a friend’s photo. As such, “likes” are more lightweight as they require less effort to express a sentiment compared to messages and photo comments (Mansson & Myers, 2011). In this way, selection of one type of medium over another can be a “signal of the resources one is willing to commit” to the relationship or interaction (Donath, 2007, p. 238), which can impact perceptions of equity within the interpersonal setting (Tong & Walther, 2011). As another example, the act of writing a letter, which takes time and dedication, is seen as more meaningful than routine text messages (King & Forlizzi, 2007) and considered more personal and valuable than emails (Riche et al., 2010). The explanation of these and similar findings is that the effort and perceived sincerity of the message and selected medium shapes perceptions of “the engagement of people in the conversation and hence in the relationship” (Riche et al., 2010, p. 2709). Consequently, technologies that lower one’s cost in engaging in interpersonal activities may also dilute relational assessments important for generating felt obligations toward others (Liu et al., 2023). Although much of the extant research has focused on friendships or romantic partners, similar findings in workplace-relevant interactions indicate they likely apply to workplace relationships as well. For instance, Brodsky (2021) finds that negotiators’ choice of medium impacts their counterparts’ perceptions of their emotional authenticity, where choosing more lightweight communication channels led them to seem less authentic, demonstrating that coworkers may use others’ choice of medium to gauge investment and perceptions of social exchange.

We build on and extend this work on technology-mediated relational building and the distinction between lightweight and heavyweight technology in interpersonal interactions to more deeply explore the resulting interpersonal perceptions and implications for social exchange. With the growth in AI capabilities, such as the rapid growth of large language models like ChatGPT, there are increasing opportunities for individuals to delegate tasks to AI-based tools that were, until recently, solely within the provenance of human ability (Vanneste & Puranam, 2024). Now, employees can delegate many tasks to AI-based tools, requiring less time and cognitive load from the human user (Hancock et al., 2020). AI capabilities are rapidly expanding in many domains: Physicians can delegate some of their clinical decision making to AI systems (Sun et al., 2023); software developers can delegate their coding tasks to ChatGPT (Ahmad et al., 2023); employees can use AI to modify or even generate messages for them (Hancock et al., 2020); and executives can use LLMs to generate and edit apology letters to stakeholders and fans when products underperform (Bevan, 2023). In the context of help giving and receiving, a coworker offering help may off-load some of his/her contribution to AI tools, such as ChatGPT (Liu et al., 2023; Wei et al., 2023). As a result, help recipients may value such help differently, in part due to the altered perception of social exchange toward the helper.

Although extant work on technology-mediated relational building offers some explanation as to why some media choices are perceived differently than others, we do not know about the specific perceptions that recipients develop and their implications for the felt obligations that are normally experienced in response to helping behavior. To further develop theory in this area, we draw from Fiske’s theory of social perception, or the stereotype content model, to theorize about the responses that AI-enabled helping behavior may elicit in recipients. Specifically, we propose that whether a coworker uses an AI tool when providing help may affect the recipient’s perceptions of his/her warmth.

Warmth perceptions reflect evaluations of an actor as nice, friendly, and pleasant, and they are a foundational component of person perception across cultures and situations (Eisenbruch & Krasnow, 2022; Fiske et al., 2002; Inagaki & Eisenberger, 2013; Wojciszke et al., 1998). Theories of person perception suggest that the level of warmth observers attribute to a particular actor is based on inferences about motive: someone thought to be benevolent is likely viewed as warm, particularly compared to someone perceived as hostile or competitive (Cuddy et al., 2008). Cuddy et al. (2011) posit that warmth is especially relevant in organizations because of the common forms of interdependence connecting coworkers. They argue that employees witness the actions of their coworkers and subsequently make judgments about whether the actor wishes them well or harm (e.g., Huang et al., 2017; Wellman et al., 2016).

Building on this theoretical foundation and integrating it with existing theories of technology-mediated relationships, we suggest that applying an AI tool in the context of a helping event is likely to undermine the foundations of social exchange. Specifically, we predict that individuals who delegate their actions to an AI tool when providing help should be viewed as less warm by recipients, because they will be seen as deploying a lightweight technology signaling lower investment and effort relative to individuals who do not use an AI tool. As previously discussed, costs are important in fair and equitable social exchanges. As the AI-aided help sender spends few costs during his/her actions, the recipient may become less inclined to view the sender as “going out of their way” when providing help. For example, in a study by Liu et al. (2023), participants were asked to imagine texting with another individual about a variety of interpersonal contexts (e.g., experiencing burnout and needing support, having a conflict with a colleague and needing advice). When participants learned that their partner used AI to augment his/her text responses back to participants, they reported significantly lower relationship satisfaction. We predict that, in a helping context, using an AI tool transparently will affect recipient perceptions of warmth toward the helper.

We further expect warmth to predict felt obligations. Research on Fiske’s (2018) person perception model demonstrates a causal connection between warmth perceptions and subsequent emotions and intentions. Accumulated evidence supports the theory that in social situations, individuals assess whether an actor is high or low in warmth and then experience emotions, intentions, plans, and cognitive evaluations in light of that information (Cuddy et al., 2007; Wojciszke, 1994). Moreover, when people experience warmth from others, it often engenders positive emotions like gratitude and empathy (Lee et al., 2019; Sawyer et al., 2022). These emotions reinforce felt obligations as individuals are naturally inclined to act in a way that preserves and amplifies the positive emotional experiences.

We connect those findings to research demonstrating that felt obligations arise from a complex set of emotional and cognitive responses that individuals experience in response to perceived warmth and friendliness in social interactions, leading to an internalized sense of duty or responsibility to reciprocate positive behavior and maintain a positive social contract with others (Eisenberger et al., 2001). In connecting these two bodies of work, we theorize that help recipients perceiving high warmth in a helper are likely to experience greater felt obligations stemming from their natural inclination to uphold the positive social contract. Empirical evidence lends support to this prediction. For instance, warmth often leads to helping intentions (Simon et al., 2020), liking (Huang et al., 2017), desires to assist others (Levine & Schweitzer, 2015), obligations to purchase from sales targets (Wang et al., 2017), and when the target is a company, brand loyalty (Kervyn et al., 2022). We therefore hypothesize that perceptions of warmth (a) positively relate to felt obligations and (b) mediate the relationship between AI-assisted helping behavior and felt obligations.

Felt obligations (within the recipient), in turn, are expected to predict enacted help. Understanding why individuals reciprocate help is tied to the concept of reciprocal obligation. Interpersonal OCBs are fundamental in contributing to the workplace’s social environment and in fulfilling the social duties arising from interpersonal exchanges (Cropanzano & Mitchell, 2005; Organ, 1988). For example, research generally supports the notion that individuals offer more helping behaviors when they work in contexts of high-quality social exchanges—social exchanges either with their leaders (Eisenberger et al., 2001), the organization in general (Masterson et al., 2000), or after receiving assistance from coworkers (Deckop et al., 2003). Helping behaviors are an important way that individuals can fulfill obligations to reciprocate benefits received from others, especially if those benefits are in the form of assistance or support (Bowling et al., 2005). The felt pressure to reciprocate reflects a norm of reciprocity and stems from the dissonance that is created by the inequity of the social exchange (Adams, 1965). Individuals typically strive to resolve this dissonance by allocating resources to the relationship (Perugini et al., 2003) so as to avoid debt in social exchanges (Blau, 1964). Along these lines, research has found that when individuals feel that they are receiving more than reciprocating, they tend to increase cooperation (Clark & Sefton, 2001) and enact more helping behaviors (Halbesleben & Wheeler, 2011). The above theory and related research suggest that felt obligations should prompt individuals to enact help.

We based our main predictions on warmth, following Eisenbruch and Krasnow’s (2022) arguments that people react more strongly to warmth than competence in social and interpersonal contexts. However, we also consider competence as an exploratory part of the study. Competence perceptions reflect evaluations of an actor as capable, intelligent, and effective, and they represent the second foundational dimension of person perception across cultures and domains (Fiske et al., 2002; Wojciszke et al., 1998). Theories of person perception suggest that competence judgments are rooted in inferences about ability: observers assess whether the actor is skilled enough to achieve their goals or perform effectively in a given context (Cuddy et al., 2008). While warmth tends to dominate early impressions and emotional reactions, competence plays a critical role in shaping perceptions of status, influence, and task performance (Fiske et al., 2007; Judd et al., 2005). As employees observe the outcomes of others’ actions—such as whether a task is completed successfully and with sufficient quality—they infer levels of competence that shape how they evaluate coworkers and decide who deserves credit, deference, or future opportunities (Anderson & Kilduff, 2009). Relying on an AI tool to provide help may influence perceptions of competence. For example, some recipients may interpret AI use as a sign of efficiency and technological savvy—an indicator that the helper is skilled at using modern tools to solve problems. At other times, however, AI use may be a sign of limited ability, raising doubts about the helper’s underlying competence. Therefore, we also examine the following research question:

Participants were 371 individuals from the online research platform Prolific. They were required to work full or part time, have an approval rating of at least 98%, speak English as their primary language, and be citizens of the United States. The average age of participants was 37.7 years (SD = 11.9), and 48% were female. Most participants were White (71%) and held a college degree or higher (66%). We recruited 400 total participants, but 7 failed attention checks and 22 did not complete the study beyond the training sequence, yielding 371 as our final sample.

We conducted a preregistered, online behavioral experiment to examine whether providing assistance using an AI tool Changes perceptions, felt obligations, and subsequent enacted help. Participants were randomly assigned to one of four conditions that resulted in a 2× 2 between-subjects design, with AI tool use (coworker did or did not use ChatGPT when offering help) and help type (person versus task) as factors. Our preregistration can be found at https://​osf.​io/​ayudn?​view_​only=​f00a76754d3c495b​9f9d359e279fa76b​.

Participants were introduced to an experimental task involving a scenario in which they were asked to work as a member of a caregiving team, offering assistance, support, and problem solving when necessary to a care recipient. When designing this task, we adhered to principles described in Colquitt (2008). A thorough description of the task, which is a work simulation in which participants offer care support to an online recipient, is provided in Appendix 1.

Participants were informed that they would be working as a member of a caregiving team watching over a care recipient named Turto. Members of the team worked in shifts so that one individual was with the care recipient 24 hours a day. Participants were told that they would be picking up where the last caregiver finished. After being introduced to their role and team context, they completed a training sequence to learn how to support the care recipient online. Specifically, they needed to manage Turto’s symptoms by feeding and washing him and providing medicine at appropriate intervals. They were provided an opportunity to practice all necessary actions.

After completing the training, participants were told that one of their experienced teammates, Caregiver J, may send them messages during their shift. They were instructed that they could choose to respond or not, with no implications for their compensation for participation. Participants viewed the first message by Caregiver J before beginning their shift, which included the manipulation of AI tool use and help type. Person-oriented help took the form of an encouraging welcome message, trying to make the participant, as a new member of the care team, feel comfortable, whereas task-oriented help involved Caregiver J reporting on a task they took care of—writing a detailed activity report—on the participants’ behalf. To manipulate whether Caregiver J used an AI tool during the help event, participants saw a message that included a side comment such as “I used ChatGPT!” in describing their help or else made no mention of technological tool use. Specifically, participants received one of the following messages from Caregiver J prior to starting their shift:

No AI Person Help: “Hello – just wanted to say welcome to the team! It’s fantastic to have you with us, and we’re excited to get to work with you. We’re here to support you as we’ve all been in your shoes. Welcome aboard and stoked to have you with us!”

AI-Assisted Person Help: “Hello – just wanted to say welcome to the team (I’m using ChatGPT to write this message)! It’s fantastic to have you with us, and we’re excited to get to work with you. We’re here to support you as we’ve all been in your shoes. Welcome aboard and stoked to have you with us! (p.s., ChatGPT is an artificial intelligence tool on the internet that helps write and edit things).”

AI-Assisted Task Help: “Turto’s family wanted an update/report, including details about his health and recent activity. Normally this report is something you would write during your shift, but I went ahead and took care of it by using ChatGPT. All set – you don’t need to worry about this task. (p.s., ChatGPT is an artificial intelligence tool on the internet that helps write and edit things).”

No AI Task Help: “Turto’s family wanted an update/report, including details about his health and recent activity. Normally this report is something you would write during your shift, but I went ahead and took care of it. All set – you don’t need to worry about this task.”

Then, participants were asked to rate their perceptions of Caregiver J’s warmth (Fiske et al., 2002) as well as their felt obligations toward their caregiving team (Eisenberger et al., 2001; Lee et al., 2023; Lorinkova & Perry, 2019). Subsequently, participants started their caregiving shift. After 15 s into the shift, all participants received a message from Caregiver J asking for help—their response to this message was our behavioral measure of enacted help (see Christian et al., 2012 for a similar behavioral measure). In total, the shift was 7 min long, and participants were paid $12.00 per hour.

We used validated surveys from established literature to measure key constructs. Unless otherwise noted, response scales ranged from 1 = strongly disagree to 5 = strongly agree.

We controlled for age and gender (0 = female, 1 = male) due to possible differences in attitudes toward technology (Cai et al., 2017). We also controlled for task performance, or participants’ efficacy in carrying out their caregiving tasks. This score was operationalized as the total time between problematic symptoms arising in Turto, and the moment ailments were cured via actions such as feeding or providing medicine. High scores indicated low performance—when participants failed to address Turto’s symptoms or were delayed in responding appropriately, scores on this variable increased. For robustness, the results reported below include controls. The results do not change significantly when controls are removed.

As an attention check, participants were provided with a color at the end of their shift (e.g., “blue”) and then asked to insert this color into a subsequent question on their post survey. Seven participants failed this attention check. As a manipulation check, participants were asked whether Caregiver J used ChatGPT when first messaging them (Yes/No). One hundred percent of the retained 371 participants answered this question correctly.

To assess the factor structures of warmth, competence, and felt obligations, we conducted a series of confirmatory factor analyses. A 3-factor solution with warmth, competence, and felt obligations as 3 distinct factors demonstrated good fit (χ² = 128.05, df = 32, comparative fit index (CFI) = 0.97, Tucker–Lewis index (TLI) = 0.96, root mean square error of approximation (RMSEA) = 0.09, standardized root mean square residual (SRMR) = 0.04), which was significantly better than 2-factor structures combining warmth and competence (χ² = 516.11, df = 34, CFI = 0.87, TLI = 0.83, RMSEA = 0.20, SRMR = 0.07; (Δχ² = 388.06, p = 0.00) or warmth and felt obligations (χ² = 1245.54, df = 34, CFI = 0.67, TLI = 0.56, RMSEA = 0.31, SRMR = 0.19; Δχ² = 1117.50, p = 0.00) into a single dimension.

We present means and standard deviations of study variables in Table 1. Consistent with our theorizing, warmth was significantly associated with felt obligation (r = 0.42, p = 0.00) and felt obligation was significantly associated with enacted help (r = 0.12, p = 0.02). In addition, performance was unrelated to felt obligation (r = 0.02, p = 0.78) or enacted help (r =  − 0.06, p = 0.22). There were no significant differences in performance across the study conditions (F(3, 367) = 0.94, p = 0.42). We present conditional means and analysis of variance (ANOVA) contrasts in Table 2. Consistent with our theorizing, the effect of AI use on warmth was significant (F(1, 367) = 53.37, p = 0.00, \(\eta\) ² = 0.13), such that participants who received help from a coworker using ChatGPT saw their coworker as significantly less warm (M = 3.75, SD = 0.78) than did those receiving help from a coworker who did not use ChatGPT (M = 4.26, SD = 0.52; t(312.68) =  − 7.25, p < 0.01). In addition, there was no moderating effect of help type on the relationship between AI tool use and warmth. Regarding competence, there was a significant effect of AI use on competence (F(1, 367) = 21.66, p = 0.00, \(\eta\) ² = 0.06), such that participants who received help from a coworker using ChatGPT saw their coworker as significantly less competent (M = 4.00, SD = 0.70) than did those receiving help from a coworker who did not use ChatGPT (M = 4.31, SD = 0.57; t(347.46) =  − 4.54, p < 0.01). There was also significant effect of help type on competence (F(1, 367) = 15.93, p = 0.00, \(\eta\) ² = 0.04), such that participants who received person help saw their coworker as significantly less competent (M = 4.03, SD = 0.61) than did those receiving task help (M = 4.29, SD = 0.55; t(369.00) =  − 3.85, p < 0.01).

To test the full, hypothesized model including main effects, indirect effects, and controls, we specified a simultaneous path analysis using lavaan in R. Table 3 presents the unstandardized coefficient estimates for the proposed model (path model fit: χ² = 155.57, df = 18, CFI = 1.00, TLI = 1.00, RMSEA = 0.01, SRMR = 0.01). Snijders and Bosker’s (1994) formulas were used to calculate pseudo-R² for the effect sizes in predicting outcomes. Hypothesis 1 predicted that coworker AI tool use during a helping event would be negatively related to perceptions of coworker warmth. As Table 3 shows, coworker AI tool use was negatively related to coworker warmth (estimate =  − 0.51, p = 0.00), supporting Hypothesis 1. Hypothesis 2 predicted that warmth would be positively related to felt obligations. As Table 3 shows, warmth was positively related to felt obligations (estimate = 0.21, p = 0.00), supporting Hypothesis 2. Hypothesis 3 predicted that coworker AI-assisted help would be indirectly related to felt obligations via perceptions of coworker warmth. The indirect effect was tested via a Monte Carlo simulation procedure using R. This procedure was used to accurately reflect the asymmetric nature of the sampling distribution of an indirect effect (MacKinnon et al., 2004; Preacher et al., 2010). As shown in Table 3, with 10,000 Monte Carlo replications, we found that the indirect effect of coworker AI tool use on felt obligations via warmth was significant (indirect effect =  − 0.23; 95% confidence interval (CI) of [− 0.32, − 0.15]). This finding indicates that coworker AI use was negatively related to felt obligations via warmth perceptions, providing support for Hypothesis 3. Hypothesis 4 predicted that felt obligation would be positively related to enacted help. As Table 3 shows, felt obligation was positively related to enacted help²  (estimate = 0.23, p = 0.01), supporting Hypothesis 4. As shown in Table 3, we also found that the indirect effect of perceived coworker warmth on enacted help via felt obligation was significant (indirect effect = 0.10; 95% CI of [0.03, 0.18]). Finally, the serial mediation (with 10,000 Monte Carlo replications) effect of AI tool use on enacted help via our serial mediators was significant (− 0.023; 95% CI of [− 0.055, − 0.002]). We replicated these results in a sample of undergraduate students while controlling for alternative mediators, reported in Appendix 2 (study 3). We also replicated these results using different approaches to interpreting our dependent variable, described in detail in footnote 1 (Tables 4 and 5).³ Regarding our research question and competence, coworker AI tool use was negatively related to coworker competence (estimate =  − 0.31, p = 0.00), the indirect effect of coworker AI tool use on felt obligations via competence was significant (indirect effect =  − 0.16, 95% CI of [− 0.23, − 0.09]), and the indirect effect of perceived coworker competence on enacted help via felt obligation was significant (indirect effect = 0.11; 95% CI of [0.03, 0.20]).

We used Prolific to recruit 241 participants in the United States working full or part time with an approval rating of at least 98% and whose primary language was English (as well as no overlap with Study 1). Across participants, 46% were female, 81% White, 6% Asian, 5% Black, 4% Mixed, and 4% Other. The average age of participants was 41.47 years (SD = 12.40), and they were employed in a mixture of jobs including healthcare (14%), information technology (15%), legal services (15%), transportation (11%), education (9%), finance (8%) as well as media, government, security, and cosmetics. Finally, 57% held a bachelor’s degree or higher. We recruited 241 total participants, but 3 failed an attention check, yielding 238 as our final sample.

Study 2 was a preregistered, 2 × 2 between-subjects vignette experiment, with AI tool use as the first condition (coworker did or did not use ChatGPT when providing help) and help type as the second condition (coworker provided task versus person-oriented citizenship; preregistration: https://​osf.​io/​x9j6c/​?​view_​only=​28314603be3c465a​976b7771a9ab8472​). Participants were asked to imagine themselves as a member of a caregiving team providing 24-h care to a care recipient named Turto, with caregiving teammates working in shifts throughout the day. All participants were told, “You are arriving at your first shift and taking over for a teammate who just finished their shift – we will refer to this person as ‘your teammate’ when describing more details below.” After reading this background information, participants were randomly assigned a scenario in which we manipulated AI-assisted help and help type. All of the vignettes used language directly from scale items used in prior research (Dierdorff & Rubin, 2022; Settoon & Mossholder, 2002). Participants read one of the following vignettes.

Similar to study 1, we controlled for age and gender (0 = female, 1 = male) due to possible differences in attitudes toward technology (Cai et al., 2017). For robustness, the results reported below include these additional controls.

To assess the factor structures of warmth, competence, felt obligations, technological self-efficacy, technology readiness, and perceived usefulness of ChatGPT, we conducted a series of confirmatory factor analyses. A 9-factor model demonstrated better fit (χ² = 883.72, df = 524, CFI = 0.94, TLI = 0.93, RMSEA = 0.05, SRMR = 0.06) than a 7-factor model which combined warmth and competence into a single factor (χ² = 1110.49, df = 532, CFI = 0.90, TLI = 0.89, RMSEA = 0.07, SRMR = 0.07; Δχ² = 226.77, p = 0.00).

We present means and standard deviations of study variables in Table 6. Consistent with our theorizing, warmth was significantly associated with felt obligation (r = 0.50, p = 0.00). In addition, warmth was associated with a number of technology controls, including technological self-efficacy (r = 0.31, p = 0.00), tech readiness optimism (r = 0.29, p = 0.00), and tech readiness innovative (r = 0.21, p = 0.00). Competence was also positively related to felt obligation (r = 0.44, p = 0.00). We present conditional means and ANOVA contrasts in Table 7. Consistent with our theorizing, the effect of AI tool use on warmth was significant (F(1, 234) = 17.84, p = 0.00, \(\eta\) ² = 0.07), such that participants who read about received help from a coworker using ChatGPT saw their coworker as significantly less warm (M = 4.20, SD = 0.61) than did those receiving help from a coworker who did not use ChatGPT (M = 4.52, SD = 0.60; t(235.99) =  − 4.13, p < 0.01). Regarding competence, there was a significant effect of AI use on competence (F(1, 234) = 4.11, p = 0.04, \(\eta\) ² = 0.02), such that participants who read about receiving help from a coworker using ChatGPT saw their coworker as significantly less competent (M = 4.4.31, SD = 0.57) than did those receiving help from a coworker who did not use ChatGPT (M = 4.46, SD = 0.62; t(233.37) =  − 1.98, p = 0.049). There was also significant effect of help type on competence (F(1, 234) = 12.00, p = 0.00, \(\eta\) ² = 0.05), such that participants who read about receiving person help saw their coworker as significantly less competent (M = 4.24, SD = 0.61) than did those receiving task help (M = 4.51, SD = 0.55; t(230.8) =  − 3.44, p < 0.01).

To test the full, hypothesized model including main effects, indirect effects, and controls, we specified a simultaneous path analysis using lavaan in R. Table 8 presents the unstandardized coefficient estimates for the proposed model (path model fit: χ² = 159.25, df = 21, CFI = 0.92, TLI = 0.71, RMSEA = 0.09, SRMR = 0.03). Snijders and Bosker’s (1994) formulas were used to calculate pseudo-R² for the effect sizes in predicting outcomes. Hypothesis 1 predicted that coworker AI tool use during a helping event would be negatively related to perceptions of coworker warmth. As Table 8 shows, coworker AI-assisted help was negatively related to coworker warmth (estimate =  − 0.39, p = 0.00), supporting Hypothesis 1. Hypothesis 2 predicted that warmth would be positively related to felt obligations. As Table 8 shows, warmth was positively related to felt obligations (estimate = 0.38, p = 0.00), supporting Hypothesis 2. Hypothesis 3 predicted that coworker AI tool use would be indirectly related to felt obligations via perceptions of coworker warmth. The indirect effect was tested via a Monte Carlo simulation procedure using R. This procedure was used to accurately reflect the asymmetric nature of the sampling distribution of an indirect effect (MacKinnon et al., 2004; Preacher et al., 2010). As shown in Table 8, with 10,000 Monte Carlo replications, we found that the indirect effect of coworker AI tool use on felt obligations via warmth was significant (indirect effect =  − 0.20; 95% CI of [− 0.30, − 0.11]). Regarding our research question and competence, coworker AI tool use was negatively related to coworker competence (estimate =  − 0.20, p = 0.01), competence was positively related to felt obligations (estimate = 0.21, p = 0.00), and the indirect effect of coworker AI use on felt obligations via competence was significant (indirect effect =  − 0.10, 95% CI of [− 0.18, − 0.03]).

Our results contribute to the current research in several ways. First, our findings extend theories of technology-mediated relationship building and maintenance (Rusbult et al., 2012; Walther, 2011). In the context of technology-mediated relationships, digital tools can be used to enhance interpersonal communication efficiency, enabling individuals to participate in innovative and lightweight relationship activities (Maruping & Agarwal, 2004). In their review, Tong and Walther (2011) acknowledged the positive impact of technology but also raised concerns about potential drawbacks. Citing equity theory (Hatfield et al., 1978), they suggested that recipients might perceive low-cost actions as less genuine in a relational context and called for future research to examine the effects of technology “from a relational transaction costs perspective” (Tong & Walther, 2011, p. 112). Our study explored the consequences of receiving help from an AI-supported coworker. In our experiment, we varied whether coworkers used ChatGPT to complete their helping behaviors while keeping the message properties (near) constant. Despite recipients observing the same actions from Coworker J, the variation in the processes behind those actions—whether involving ChatGPT or not—influenced how recipients responded. This resulted in participants perceiving less warmth when coworkers used ChatGPT during their actions. Furthermore, diminished warmth perceptions prompted participants to feel a reduced sense of obligation and were less likely to reciprocate help. Therefore, while AI tools may enhance individual efficiency, they can also incur relational costs and undermine the social glue that managers and practitioners desire from citizenship behaviors.

Second, we extend citizenship theory, much of which assumes that when employees go out of their way to help, they are making a discretionary investment of time, energy, or expertise—and that this investment triggers social and psychological mechanisms like gratitude, obligation, or reciprocity. However, as new technologies enable people to offload parts of their work, this assumption no longer holds. Our findings suggest that the relational effects of helping depend not just on what someone does, but how they do it. Perceiving a helping behavior as supported by ChatGPT is viewed as less warm and reduces felt obligations to reciprocate. We therefore expand citizenship theory to account for the growing role of automation and AI in how helping behaviors are interpreted and responded to.

Of course, other workplace tools distinct from AI are available too. What we view as particularly unique about AI or ChatGPT is the degree to which people perceive it to take over the task entirely. Research shows that when people are told a product was created “with AI assistance,” they often assume the AI generated it almost entirely, with little human involvement (Altay & Gilardi, 2024). This perception of full delegation makes AI qualitatively different from tools like templates, calculators, spell-checks, or search engines, which are seen as supporting human investment/effort rather than replacing it. In helping contexts, this matters: recipients may infer that the helper exerted minimal investment, leading to reduced feelings of warmth, obligation, and willingness to reciprocate. Using ChatGPT and related generative AIs may reshape how help is socially evaluated.

Third, we extend social perception theory by considering warmth as one mediator between received help and felt obligations. This discovery illuminates one possible pathway explaining why AI-assisted help may lead to diminished enacted help and felt obligations in the realm of coworker interactions. Although prior research has acknowledged the potential impact of AI on social dynamics, our theory describes one specific mechanism by which this can occur. By pinpointing warmth as a mediator between AI-assisted help and subsequent behavioral outcomes, we provide an understanding of one psychological process at play. Although we focused on warmth, there may be other mechanisms that explain why received help, when augmented with AI, backfires. Additional constructs may include effort (Wei et al., 2023), time spent (Rusbult, 1980), or authenticity (Brodsky, 2021). For example, Liu et al. (2023) asked participants to imagine exchanging text messages with a hypothetical friend, Taylor, regarding recent interactions (e.g., speaking about an upcoming birthday, needing advice after having a conflict with a colleague). Compared to those who believed Taylor sent messages individually, those who learned that the messages were aided by an AI system perceived significantly lower effort and relationship satisfaction in the hypothetical interaction. Thus, there may be additional mediators in addition to warmth that could operate as psychological perceptions relating to future felt obligations. Or, it may be that these constructs are upstream mediators from received help to warmth. In other words, perhaps individuals form perceptions of presumed time and effort taken, and then these perceptions fuel an overall assessment of warmth, leading to additional downstream consequences. These ideas may be interesting domains for future studies.

There is an important and ongoing societal conversation about the fourth industrial revolution and the rapid incorporation of AI into the twenty-first century workplace. The conversation often centers on how this technology can enhance employee productivity by helping them to understand trends, improve communication, and improve data analytics. However, it is vital to acknowledge that the implementation of new technology can bring about real costs that are sometimes overshadowed by these promises. Our study introduces a significant cautionary note, highlighting that while technology and AI systems may reduce execution costs, they could potentially elevate relational costs. In other words, although these systems streamline tasks, they might impact interpersonal relationships, potentially influencing perceptions of warmth and, consequently, feelings of obligation and enacted help. Practically, organizations should be mindful of the social implications when incorporating AI into the workplace. The findings suggest that coworkers using AI for assistance may be perceived as less warm, which, in turn, could affect feelings of obligation and subsequent collaborative behaviors. This insight encourages a thoughtful approach to the integration of AI, considering not only efficiency gains but also the potential impact on workplace relationships and collaborative efforts.

Given the speed at which AI tools, LLMs, and algorithms are being integrated into organizations across the globe (Bamberger, 2018; Makridakis, 2017; Von Krogh, 2018), we urge managers to be aware of the potential downside of over-relying on AI tools when employees enact helping behaviors. This insight may be particularly useful for organizational design efforts, whereby managers and organizations may find it necessary to think deeply about the extent to which LLMs creep into typically human-centered social interactions. Just as new technologies bring with them digital undertows at the industry level (Scott & Orlikowski, 2022), this same phenomenon may operate at the individual level as well. The crux of why citizenship behaviors are important in applied contexts is because they, in the aggregate over time and people, raise the overall effectiveness of a collective. However, if one individual realizes efficiency gains by using an AI tool but then later incurs costs from failed reciprocity, the collective net outcome is potentially zero. In addition to the growing literature on when received help backfires, this study urges leaders to consider helping processes in their entirety across individuals.

An important caveat to our findings is that the broader relational costs and benefits of AI-assisted help at the collective level remain an open question for future research. One could argue that the sender may in fact gain from using AI because it frees up time to assist with other tasks or support additional coworkers. However, our studies point the fact that, although AI may increase helping availability initially, it can disrupt the give-and-take of interpersonal exchange relationships. For instance, in real organizational settings, imagine a team member who uses AI to summarize client feedback, saving a colleague from doing the task manually. The colleague may appreciate the help in the moment but—per our findings—may be less likely to reciprocate later (e.g., by helping with a different task under deadline). If that second task then goes unfinished, the team’s collective performance could suffer, even though the initial AI-assisted help was well intended and efficient. Moreover, the recipient may now see the sender as less warm (relative to someone who provided help without AI). So, while the sender may “gain” from using AI (by freeing up time for other tasks), the recipient perceives the sender as less warm and becomes less likely to reciprocate, and the subsequent task that required support may go unaddressed. From a dyadic perspective, this creates a potential relational cost: the efficiency gained in the initial helping act is offset by a breakdown in reciprocity. On the other hand, if we zoom out even further and consider the total number of people the sender might be able to assist due to AI-enabled efficiency, the broader organizational impact could look quite different. This tension between efficiency and reciprocity is an important area for continued investigation.

Our findings should be considered in light of several limitations. The distinction between task-focused and person-oriented helping behaviors was based on Settoon and Mossholder’s (2002) traditional classification, which has been a fundamental distinction applied to helping research since its inception (Ehrhart, 2018). However, in the last several years, a number of additional dimensions of help type have emerged in the literature, including anticipatory versus reactive helping (Harari et al., 2022), autonomy versus dependency helping (Chernyak-Hai et al., 2024), and empowering versus non-empowering helping (Lee et al., 2023). Anticipatory help involves providing assistance in advance of being asked. Autonomy helping occurs when the helper provides the recipient with information and resources that will allow him/her to resolve the issue in the future without help. Empowering help occurs when the recipient is an interactive participant in the resolution of an issue and the event is viewed as a mastery experience. It may be beneficial for future research to extend our work by considering the extent to which AI-assisted help across these various help types influences recipient perceptions and attributions.

We also held the qualities of Caregiver J constant across conditions. This individual, in our experimental scenario, not only provided help but also asked for assistance from participants. Prior research has shown that several individual differences, including status (Burke et al., 1976; Tai et al., 2023; Toegel et al., 2013), gender (Akinola et al., 2018; Heilman, 2012), personality (Hughes et al., 2022), and minority identity (Nguyen & Ryan, 2008) have important implications for help receiving and giving dynamics. For example, Tai et al. (2023) found that recipients felt more threatened when receiving help from individuals perceived to be of greater status. We recommend exploring in future research how these characteristics of the help giver (and asker) may influence the recipient’s cognitive, emotional, and behavioral reactions.

Moreover, we assessed warmth, competence, and felt obligations concurrently throughout our studies, which raises potential concerns for common method variance (CMV) bias (Podsakoff et al., 2012). This simultaneous measurement makes it difficult to discern the true separation between warmth and obligations to reciprocate. Similarly, our research provides a snapshot of the immediate reactions to AI-assisted help. However, it is crucial to understand how these perceptions and behaviors evolve over time. For instance, repeated exposure to AI-assisted help may lead to changes in perceptions of warmth and felt obligations as individuals become more accustomed to this type of assistance. Longitudinal studies would provide valuable insights into how AI-assisted help influences social dynamics in organizations over the long term. Last, in our studies, AI-assisted help was explicitly disclosed to the recipients. However, the dynamics of revealing or concealing AI involvement in helping behaviors is a complex issue that warrants further investigation. The identity management literature suggests that how individuals manage information about their identity and actions—for example, choosing to reveal, conceal, or signal characteristics about oneself (Clair et al., 2005)—influences interpersonal cognitions and behaviors (Button, 2001; Jones et al., 2016). Future research could explore the strategic choices behind these disclosures and their implications. Questions, such as how third parties discover AI delegation, whether individuals are transparent about their reliance on AI, and the extent to which organizational culture supports open use of AI tools, will be critical to understand moving forward.

Our studies focused on situations in which AI use is explicitly acknowledged by the helper—where the recipient clearly knows that a tool like ChatGPT was used to generate the assistance. We modeled this after instances demonstrated in professional and public communication, where individuals voluntarily disclose AI involvement. For example, emails may include statements such as “Drafted with ChatGPT,” academic presenters sometimes begin talks with “Slides generated using AI,” and social media platforms like TikTok may automatically display “AI-generated” tags on videos. In workplace settings, employees might preface a message with “I used ChatGPT to help write this,” or an executive’s letter to stakeholders may be revealed as AI generated. These kinds of explicit disclosures create a social context in which recipients are aware that AI was involved in the process. While our findings apply directly to such clear and transparent use cases, we recognize that many real-world situations involve more implicit or ambiguous AI use. Future research should explore how undisclosed or passively signaled AI assistance affects social perceptions differently, particularly regarding issues of intentionality and transparency.

In our studies, participants did not directly use the output that Caregiver J generated. As such, our design held the perceived utility of the help relatively constant across conditions, focusing instead on the source of the help (AI-assisted vs. unassisted) and the type of help (task vs. person-oriented). We attempted to capture utility indirectly in the task condition, where Caregiver J removed a responsibility from the participant’s agenda. However, this differs meaningfully from situations in which recipients evaluate or depend on the actual quality of an AI-generated product. It is possible that when help includes a deliverable—like an initial financial model or source code—the effectiveness of that product could influence perceptions of obligation, potentially increasing felt obligation if the AI-generated result is especially useful. Future research should explore this alternative pathway by examining how actual use and perceived utility of AI-assisted outputs interact with perceptions of effort and warmth in shaping social exchange outcomes.