Communication increases collaborative corruption ☆

Despite being a pivotal aspect of human cooperation, only a few studies within the field of collaborative dishonesty have included communication between participants, and none have yet experimentally compared this to non-communicative contexts. As a result, the impact of communication on unethical collaborations remains unclear. To address this gap, we conducted two well-powered studies ( N total = 1187), closely replicating and extending seminal research by Weisel and Shalvi (2015), introducing communication as a manipulated variable within a dyadic cheating task. Across both studies, we found evidence that communication increases the magnitude of cheating — even when coordination on the task is not allowed. Importantly, the effect of communication was linked to a stronger experienced collaboration among the communicating dyads, highlighting that communication is not only key to everyday ethically sound collaborations, but also to corrupt collaborations.


Introduction
Humans are a cooperative species, whose amicable ability to collaborate is regarded as one of the most fundamental factors in their evolution (Bowles & Gintis, 2011) and morality (Tomasello, 2016), and still constitute an important pillar of modern societal life.However, despite the clear benefits of collaboration, research in behavioral ethics and collaborative corruption has shown that collaborative settings increase the magnitude of economic dishonesty compared to instances where individuals behave independently (e.g.Beck et al., 2020;Conrads et al., 2013;Gerlach et al., 2019;Gross et al., 2018;Kocher et al., 2018;Weisel & Shalvi, 2015, see also; Leib et al., 2021;Zickfeld et al., 2022 for recent review and meta-analyses).Thus, human collaborative tendencies might, paradoxically, also play a key role in shaping corrupt behaviors.
The field of collaborative corruption has provided valuable insights into the interplay between cooperation and (un)ethical behaviors.Weisel and Shalvi (2015) was among the first to show that participants collaborating in a dyadic die rolling game, cheated more than individuals playing the game alone.Similar results have been observed in different decision contexts; for instance, dishonesty increases (1) when participants make decisions individually but report simultaneously (e.g., Kocher et al., 2018), (2) when the group makes one joint decision (e.g., Beck et al., 2020;Korbel, 2017) or (3) when each group member makes a decision and reports one after each other (e.g., Gross et al., 2018;Wouda et al., 2017).
Despite having greatly advanced our understanding of the role of collaboration in corrupt and fraudulent behaviors, many studies within the field are based on sequential decision-making with limited interaction between participants.However, true joint decision-making often involves some interaction, wherein group members engage in communication about possible courses of action (Beck et al., 2020).Consequently, a central element integral to collaborations outside the laboratory is often absent in these studies: verbal communication.
Verbal communication refers to communication through the use of language (words) in written and oral form and differs from nonverbal communication (e.g., gestures and facial expressions (Buck & VanLear, 2002)).Verbal communication is perhaps the most pervasive form of human communication (De Saussure & Rocci, 2016) and is essential for coordinating behaviors to achieve common goals (Tomasello, 2008) and for fostering commitment among individuals (Kerr & Kaufman-Gilliland, 1994).Furthermore, both oral and written communication have been found to increase collaboration in various games, such as ultimatum games (Roth, 1995;Zultan, 2012), trust games (Charness & Dufwenberg, 2006), and dictator games (Andreoni & Rao, 2011).Yet, despite its evident importance and although collaborative dishonesty has been defined as "joint unethical acts" (Shalvi et al., 2016, p. 134), experimental paradigms employed to study collaborative dishonesty frequently only allow minimal opportunities for joint decisions. 1 Theoretically, verbal communication can both foster and hinder dishonesty.It provides a platform for coordination and justifications that may ease dishonest acts (Kocher et al., 2018).However, as reputational concerns increase in more social settings (De Cremer & Barker, 2003), it also amplifies reputational risks, potentially deterring dishonesty.Moreover, directly discussing the opportunity of behaving dishonestly may enhance ethical considerations by making the unethical aspect of dishonesty more salient and less ambiguous, in turn potentially mitigating dishonesty.Indeed, research have indicated that increasing the salience of unethical aspects of behavior or moral awareness reduces dishonesty (e.g., Mazar et al., 2008;Sturm, 2017).This disparity raises a central question: How does the introduction of communication influence collaborative dishonesty?More broadly, does the inclusion of communication lead to an increase in dishonest behavior, and if so, what are the underlying mechanisms?
In response to these inquires, we present two well-powered experimental studies that directly investigate the role of communication in fostering collaborative dishonesty (Fig. 1).First, we examined a scenario where individuals had the opportunity to engage in dishonesty for monetary gains.In this context, we compared a non-communicative setting to a setting where participants were allowed to communicate and further related this to levels of collaboration by directly measuring participants experience of collaborating with their partner (Study 1, Fig. 1a).
Then, because communication plays a key role in fostering commitment to others (Kerr & Kaufman-Gilliland, 1994), we reasoned that being able to communicate in itself (without coordination and justifications) would increase dishonesty through feelings of commitment to one's partner.Indeed, feeling committed to others may constitute a crucial motivation for engaging in dishonesty (Zickfeld et al., 2022).It has been found that this tendency increases as people become better acquainted with each other (Beck et al., 2020) and when the dishonesty benefits the group or other individuals (Cadsby et al., 2016;Erat & Gneezy, 2012).Therefore, in a second study (Study 2, Fig. 1b) we manipulated the content of participants conversation varying whether participants were allowed to coordinate and discuss the task.In doing so, we examined how different communicative settings would impact collaborative dishonest behavior.
By including and manipulating communication, this paper advances the current literature on behavioral ethics in several ways.Firstly, whereas the majority of past work on collaborative dishonesty have restricted participant's ability to communicate and engage in joint decisions with their partner, this paper introduces this important aspect.This significantly broadens the applicability of findings to real-world scenarios, such as large corporate scandals, where communication plays a central role in coordinating such actions.Secondly, while a few studies have allowed communication between participants (e.g., Beck et al., 2020;Kocher et al., 2018), the present paper is the first to directly test the impact of communication in collaborative corrupt behaviors.Thirdly, by varying the kind of social communicative exchanges, our findings illuminate the conditions that exacerbate challenges related to collaborative dishonesty, shedding light on why and how communication might facilitate such behaviors.Lastly, by relating dishonesty to experienced collaboration, we contribute knowledge into the experiential nature of behaviors that arise in the settings used to study corrupt collaborations, thereby providing new and important insights into the degree and nature of collaboration captured in collaborative dishonesty studies.

Study 1: Does communication corrupt?
The importance of verbal communication in human social behavior is clear.First, the origin of verbal communication is one of the most essential cornerstones in human evolution (Nowak, 2000).Second, it serves as a primary channel through which we transmit socially shared knowledge (Mesoudi, 2011) and establish joint intentions and goals to effectively coordinate our behaviors to achieve a common goal (Tomasello, 2008), which is crucial for cooperation (Mitkidis et al., 2013).Therefore, we reasoned that verbal communication would also be instrumental when people collude on deceiving others for collective or monetary gain.
Few studies on collaborative corruption have included communication between participants.In one study, participants in groups of three were given the opportunity to chat with each other via a chat box during a die-roll game.This finding suggests that communication could facilitate dishonesty in group settings by allowing participants to coordinate their actions and justify deceitful behavior (Kocher et al., 2018).Similarly, another study, in which participants were able to engage in joint decision-making in face-to-face interactions, found further evidence that groups act more dishonest, because they are more readily exposed to justifications for dishonesty (Beck et al., 2020).While such studies provide important insights into how communication can facilitate corrupt collaborations, these studies do not provide direct evidence that communication increases dishonesty.In fact, although a recent metaanalysis indicates that communication is linked to increased dishonesty (Zickfeld et al., 2022), no studies have systematically isolated the effect of communication by comparing it to non-communicative contexts.Consequently, the extent to which communication influences collaborative dishonesty remains largely understudied.
Considering this, we closely replicated and extended the work by Weisel and Shalvi (2015), which has played a central role in establishing the corruptive role of collaboration.In the original study, participants engaged in a private die roll game either individually or in pairs, in which they could cheat for monetary gains by misreporting the result of their die rolls.Specifically, in the authors' aligned incentives treatment, participants playing in dyads would receive the same payoff based on the number of doubles they reported.Participants playing alone instead rolled the die twice and reported the results of their die rolls.Notably, compared to individual players, pairs exhibited significantly greater levels of cheating (indicated by higher numbers of reported doubles above chance level), underscoring the facilitative role of collaboration in dishonest behavior.Nevertheless, the interaction in the pairs was confined to receiving feedback on the partner's die-roll on a computer, minimizing the salience of collaboration.While participants might still be able to communicate their intention and willingness to engage in dishonesty through their reporting behavior, participants could not directly discuss the task and explicitly coordinate on how to proceed during the game.We therefore extended Weisel and Shalvi (2015) original aligned paradigm to enhance the conditions for collaboration by manipulating whether participants in the dyadic conditions could communicate with each other or not.In doing so, we aimed to investigate the interplay of communication and collaborative dishonesty.
1 While communication within the joint action literature has been described as a joint action in and of itself (e.g., Moore, 2018;Tomasello, 2008), we focus on communication as a means of collaboration, serving as an important channel through which people can collaborate more effectively.Building on Weisel and Shalvi (2015), we hypothesized that dyads (both those that could communicate and those that could not) would report more doubles compared to individuals (H1).Additionally, we expected that permitting communication would allow participants to align their goals and more easily coordinate on which numbers to report.We therefore further hypothesized that dyads who were able to communicate would report a higher number of doubles compared to dyads that could not communicate (H2).
Furthermore, because communication is important to collaboration and joint decision-making (Beck et al., 2020;Zultan, 2012), constituting a channel to effectively coordinate behavior and exchange justifications (Kocher et al., 2018), we reasoned that the effect of communication on number of reported doubles would depend on the experience of collaborating with the other player.Thus, by allowing participants to communicate, we increase the conditions for collaboration and goal coordination, which may also increase coordinated dishonesty.To test this, we therefore measured the subjectively perceived collaboration within both dyadic conditions, hypothesizing that the effect of communication on dishonest reporting would be mediated by an increased experience of collaborating with the other player (H3).2

Participants
A total of 430 participants from Prolific completed the study.Seven participants failed the attention check placed in the end of the study in a measure of the personality trait Honesty-Humility (see Supplementary Material; SM section 2.4.2.), and two participants were missing a partner after we removed the participants who failed the attention check.Following our pre-registration, these participants were excluded from the herein presented analyses, leaving a final sample size of N = 421 (161 dyads and 99 individual players; 47.5% women, 50.6% men, M age = 40.32,SD age = 12.57).Results including all participants are reported in SM (see section 2.5.).
Due to the magnitude of the effect size in the original study (d = 0.94), which yielded a rather small sample size estimate (34 participants [12 dyads and 22 individual players]) in an a priori power analysis (d = 0.94, 1-β = 0.80, α = 0.05, allocation ratio N1/N2 = 1.8, two-tailed Mann-Whitney U test), we determined our sample size based on the smallest effect size of interest.
We follow Simonsohn (2015), and define a small effect as the effect that would give the original study a power of 33% (d 33% ).Based on this, we estimated the minimum sample size of 53 dyads (n = 106) in each dyadic condition and 95 participants in the individual condition (d 33% = 0.44, α = 0.05, 1-β = 0.80, allocation ratio N1/N2 = 1.8, one-tailed In Study 1 participants were randomly assigned to either an individual condition, a no-communication condition, or a communication condition.After a practice round, participants would begin phase 2, consisting of 10 trials of a die-roll task adapted from previous work by Weisel and Shalvi (2015).After completing the trials, participants indicated their subjective level of experienced collaboration in the dyadic conditions.Study 2 followed a similar procedure.Participants were randomly assigned to a no-communication condition, an unrestricted communication condition (reflecting the communication condition in Study 1) or a restricted communication condition, wherein participants were not allowed to talk about the task and instead were given specific topics to discuss.

Design and procedure
In this study, we report all measures, manipulations and exclusions.A thorough description of measures and manipulations can be found in Supplemental Material.Study 1 was conducted online using Qualtrics and the add-on application SMARTRIQS (Molnar, 2020), which allows for real-time interaction between participants on the Qualtrics platform.
To study the impact of communication on cheating behavior and experienced collaboration, the study used a between-subjects design, in which participants were randomly assigned to either of three conditions; (1) Communication (n = 164; 82 dyads), (2) No-communication (n = 158; 79 dyads) or ( 3) Individual (n = 99). 5 Following the protocol of Weisel and Shalvi (2015) participants in the No-communication condition engaged in 10 trials of the dyadic die roll game, on a computer.We used 10 trials, instead of 20 as in Weisel and Shalvi (2015).We did this to reduce fatigue among participants since the study took place in an online setup, and the finding that behavior in collaborative cheating paradigms typically stabilizes around the 10th round in sequential cheating paradigms (Zickfeld et al., 2022).Participants were given a base payment for taking part in the study and were informed that they would be able to earn an additional bonus, which was determined by random selection of one of the 10 die-rolls.
After having provided informed consent, participants began Phase 1 of the experiment, which consisted of thorough instructions of the rules of the die-roll game, including the payment structure and procedure.Participants were made aware that they would be rolling a die on an external website (www.random.org)and report their outcome of this die-roll in an allotted text box.We used a reputable external website to ensure that the die-rolls remained truly private to the participants in line with the original study.Because the die rolls were truly private, participants could inflate their profit by misreporting the actual outcome of their die rolls.Following Weisel and Shalvi (2015) aligned incentives condition, participants were informed that their payment was based on the reported outcomes and that they would only receive money, if both Player A and B reported the same number.In this case, participants would earn their reported number in Experimental Currency Units (ECU's), in which 1 ECU = £0.30.That is, if both participants reported a 1, they would both earn £0.30 in that round, while they would earn £1.8 if they both reported a six.
After reading the instructions and filling out a payment scheme to ensure that participants understood the payment structure, participants were paired with another anonymous player, and then randomly assigned either the role of Player A or Player B. Dyads then engaged in a practice round of the dyadic die-roll task.First, Player A would be asked to go to the external website, click on "roll die" and then report their outcome.This information was sent to Player B, who would then go to the external website, roll the die, and report their respective outcome.After this, participants were shown a table explaining the outcome of the round.After the practice round (Phase 1), participants began Phase 2, which followed the same procedure (see SM section 2.1.).
Dyads in the Communication condition underwent the same procedure as in the no-communication condition.However, in this condition Player A and B additionally had the possibility to chat with each other using an online chat.Participants were instructed that they had min to discuss and exchange information of their choice, except for sensitive personal information.They could end the chat at any point, but the chat would automatically close after 2 min.Participants could chat after the practice round, before beginning the 10 trials, as well as after every round of the 10 trials.In the Individual condition, participants engaged in the die roll task by playing both the role of A and B as in Weisel and Shalvi (2015).
To account for the possible extra time in between trials introduced by the chats in the Communication condition, participants in the individual and No-communication conditions answered a set of filler questions after each trial.These filler items were in the form of mundane episodic recall questions, such as "What did you have for dinner yesterday?"or "What is your favorite hobby?".We chose these writing tasks to simulate the interactive chats as much as possible and to minimize the risk of inducing unintended confounds in the experimental setup.After the trials, one round was randomly selected as payoff for Phase 2 and showed to the participants.

Measures
After having completed both experimental phases, participants were asked a set of questions related to their experience throughout the experiment.Because shared goals and intentions are central concepts within the joint actions literature (Bratman, 1992;Butterfill, 2012;Tuomela, 2020) we measured participants' experienced collaboration and having a shared goal, by including three items measuring the degree to which participants experienced (1) that they were collaborating with the other player, (2) that they had a shared goal to obtain a certain result together during the game, and (3) that they had a shared intention with the other player to report a certain outcome during the game.To examine personality differences related to dishonesty, we included the Honesty-Humility subscale of 10 items from the HEXACO personality inventory (Lee & Ashton, 2004).This scale has, repeatedly, been found to correlate with higher honesty and trustworthiness in previous work (Ashton & Lee, 2007;Heck et al., 2018;Ścigała et al., 2020;Thielmann & Hilbig, 2015).At the end of the survey, participants were asked a range of socio-demographic questions, including gender, age, education, household income, and number of people in the household.

Analysis
In line with the original study (Weisel & Shalvi, 2015), we calculated (1) the percentage of reported doubles (i.e., identical die-rolls by Player A and B) in each condition, (2) the mean reported number of doubles in each condition, and (3) the percentage of (totally) brazen B and A players (i.e., dyads who always report a double) across conditions.The number of reported doubles was calculated based on the number of times participants were shown feedback for an equal outcome across the 10 trials in Phase 2.
We also followed the same analytical strategy as in Weisel and Shalvi (2015).First, we used a Wilcoxon-signed rank test to test for differences in percentage of reported doubles with the expected percentage assuming honesty (16.7%)6 in each condition.However, as the present study included three conditions (vs. two in Weisel & Shalvi, 2015), we deviated from the original study's Mann-Whitney U test and instead used a Kruskal-Wallis test to test for differences in mean reported 3 We calculated the power analysis on a one-tailed test in the pre-registration for Study 1.A two-tailed power analysis revealed a sample of 67 dyads and 121 individual players. 4Our funding was specifically earmarked for the respective study by the funder. 5To get a similar allocation ratio as the original study and thus avoid oversampling in the Individual condition compared to the dyadic conditions, the allocation to the Individual condition was set to stop at 100 participants.After this point, participants were randomly allocated to one of the two dyadic conditions.Additional correlation analyses showed no significant association between time-of-participation and individual characteristics, such as age (r = − 0.05, p = .757),gender (r = 0.06, p = .757)and Honesty-Humility (r = − 0.01, p = .768).
number of doubles between the conditions.Multiple comparisons were tested using Dunn's test.Effect sizes (Cohen's d) was calculated using the R package effsize (Torchiano, 2016).We further used a Fisher's exact test to analyze whether the proportion of (totally) brazen B players differed between conditions.This was similarly done for brazen A players. 7  To test whether participants in the dyads affected each other's behavior, we modelled a linear-mixed effects model for each player (Player A and Player B).For Player A we predicted the die-roll reported starting from round 2 (due to the sequential nature of the study design) with fixed effects for 1) Player B's report in the prior round (lagged), 2) round number, 3) gender, and 4) age. 8As participants interacted in the practice round prior to the main trials, we also conducted an exploratory linear mixed effects model using Player B's report as an unlagged variable.We deviated from the pre-registration and specified a random intercept for group ID instead of participant ID with a random slope for round number.Because the analyses involved dyads, we used group ID to avoid singular fits for participant ID.Similarly for Player B, we predicted Player B's reported die-roll with fixed effects for: 1) reported dieroll of Player A, 2) round number, 3) gender, and 4) age, and the same random effects structure as for the model on Player A's behavior.Exploratory linear mixed effects models were used to investigate the differences in mean reported doubles predicted by condition and round number (time), with an expected interaction between condition and round number (see SM section 2.5.3.).The interaction between condition and round number is based on previous research finding that verifying behavior among leaders in collaborative dishonest settings increased over time, when the leader and subordinate had conflicting incentives, but not when incentives were disentangled (Karg, 2021).
Mean experienced collaboration and shared goals were calculated for each condition.The three items were highly correlated (Pearson's r > 0.50), and were therefore aggregated to create a mean score for each participant (cf.Mitkidis et al., 2022).We used an independent sample ttest to test for differences in experienced collaboration between the two dyadic conditions.We formulated a mediation model using the Rpackage mediation (Tingley et al., 2014) to further test whether the effect of condition on mean reported number of doubles was mediated by experienced collaboration. 9We hypothesized that communication would increase participants sense of collaboration by allowing them to coordinate and make joint decisions, thereby increasing the number of doubles reported.However, we acknowledge that an alternative direction for the relationship may be equally viable.Specifically, it is possible that participants experience increased collaboration due to the shared success of reporting doubles during the task.
Correlations between number of doubles, experienced collaboration (overall aggregate and individual items) and Honesty-Humility were analyzed using Pearson's correlations (see SM section 2.5.1.for a correlation matrix).Chatlogs in the Communication condition were cleaned10 and analyzed based on word frequencies in terms of top 30 words written in the chats to get insights into the most commonly used words in the social exchanges (See SM section 2.5.8. for results).11All analyses were conducted in the statistical environment R (version 4.2.2.R Core Team, 2022).
An exploratory mixed effects model predicting number of reported doubles based on condition and number of rounds did not find cheating increased over the course of the 10 rounds of the task (OR = 1.02, 95% CI [0.95, 1.07], p = .416;see also Supplemental Material section 3.2.7 for full model results).Furthermore, a correlation analysis indicated that the number of reported doubles was negatively correlated with the personality measure of Honesty-Humility (r = − 0.14, 95% CI [− 0.23, − 0.02], p = .027).
Overall, Wilcoxon-signed rank tests revealed that dishonesty was not only related to an inflated number of doubles but also to the magnitude of reports (i.e., higher numbers yielding greater monetary outcomes).Both A and B players reported significantly higher numbers than one would expect when assuming honesty (3.5). 13 However, analyses related to brazen A players were pre-registered under "analysis" in the pre-registration.See SM section 2.5.5 for results.Totally brazen A players are conceptualized as players who try to maximize their earnings by consequently reporting a 6, while totally brazen B players are defined as players who consistently report a double (see Weisel & Shalvi, 2015). 8We included age and gender due to previous research suggesting that dishonesty differs for gender and age, with males and younger participants often cheating more than females and older participants (Conrads et al., 2013, see Gerlach et al., 2019). 9We deviate from the pre-registration.Instead of using the shared goal item for the mediation analysis, we changed this to the single item for experienced collaboration.Because the experimental task in itself provides participants with a shared goal of getting doubles, the shared goal item was deemed less suitable for capturing variance between the conditions.Furthermore, in hindsight, the questions about having shared goals and intentions might not be as intuitive for the participants as question about the degree they experienced they were collaborating with their partner.The experienced collaboration item thus may more accurately capture the participants' experience of the collaboration during the task.We therefore chose to focus on the experienced collaboration item.
.001), reported significantly higher numbers than expected.There was no difference between conditions for A players (χ2 (2) = 4.592, p = .10),while B players reported higher numbers in the communication condition (Individual: p < .01;No-communication: p = .010).Furthermore, the proportion of brazen A and B players, 14 were higher in the Communication condition (14% A players, p = .009)and 30.5% B players, p < .001,see SM section 2.5.5.).The full description and analyses for personality measures, differences in magnitude of dishonesty, brazen players and linear mixed effects models predicting reports based on partners' behavior are presented in SM (section 2.5.1.,2.5.4-2.5.6.).

Experienced collaboration
Overall, independent-samples t-tests revealed that dyads in the Communication condition experienced a higher level of collaboration compared to the dyads who did not communicate (t(319.27),− 2.85, p = .005,d = 0.32, see Table 1).This was primarily driven by an increased level of experienced collaboration with the other player, captured by a single item (i.e., experienced collaboration).Namely, communication had a medium to large effect of d = 0.57 on experienced collaboration, while there was no significant differences between conditions for the single items measuring experience of joint goals and shared intention.
Lastly, the effect of communication on number of reported doubles was partly mediated by experienced collaboration, thereby supporting H3.The regression coefficient between Communication and number of reported doubles and the coefficient between experienced collaboration and number of reported doubles was significant (Fig. 3.).The bootstrapped unstandardized indirect effect was ACME = 0.704, and the 95% confidence interval ranged from 0.40 to 1.07 (p < .001;1000 simulations).The effect of communication remained significant (ADE = 0.889, 95% CI [0.187; 1.59], p < .008).

Study 2: Does mere communication without task-related exchanges promote collaborative dishonesty?
Overall, Study 1 provides evidence suggesting that communication is associated with increased experiences of collaboration, leading to greater magnitudes of dishonesty.However, the main underlying mechanism for this effect remains unclear.We therefore aimed at isolating different potential mechanisms, by investigating whether communication could corrupt irrespective of the possibility to coordinate and share justifications for dishonesty (e.g., Kocher et al., 2018) through an increased feeling of being committed with one's co- 14 Totally brazen A players are conceptualized as players who tries to maximize their earning by consequently reporting a 6, while totally brazen B players are defined as players who consistently report a double (Weisel & Shalvi, 2015).conspirator.We did this by varying whether participants in the communicative settings were allowed to discuss and coordinate on the task.
Overall, Study 2 aimed to replicate the finding that (unrestricted) communication increases dishonesty (H1, Study 1 H2).Since communication may facilitate a sense of commitment, we further expected that dyads that were restricted from task-relevant communication would report a higher number of doubles compared to pairs that were not able to communicate at all (H2).Similarly to Study 1, we expected that the effect of communication would be mediated by an increased experience of collaborating with the other player (H3).

Participants
We recruited a total of 796 participants from Prolific.Eleven participants failed an attention check, and eleven participants were missing a partner after we removed the participants who failed the attention check and were subsequently removed from the analyses.Four dyads in the Restricted communication condition did not comply with the rules of their conditions and discussed coordinating their rolls in the die-roll game, leaving a final sample size of N = 766 (n = 383 dyads; 49.35% women, 49.61% men, mean age = 41.04,SD = 12.85).Identification and exclusion plan as well as results including participants who failed attention checks are reported in the SM (section 3.1.2.).
Based on the effect size of communication (vs. no-communication) in Study 1, a Mann-Whitney U a priori power analysis yielded a required sample size of 56 dyads in each condition (d = 0.49, 1-β = 0.80, α = 0.05). 15Additionally, due to more financial resources than originally estimated, we continued recruiting until our resources were depleted.A one-way ANOVA sensitivity power analysis (α = 0.05, 1-β = 0.8) indicated that our sample size (N = 383 dyads), is sensitive to detect an effect of Cohen's d = 0.16 (G*Power 3.1).

Design and procedure
In this study, we report all measures, manipulations and exclusions.
A thorough description of measures and manipulations can be found in Supplemental Material.Participants were randomly assigned to one of three conditions; 1) No-communication (n = 260; 130 dyads); 2) Unrestricted communication (n = 256; 128 dyads); and 3) Restricted communication (n = 250; 125 dyads).As in Study 1, the experiment was programmed in Qualtrics using the SMARTRIQS application (Molnar, 2020).The experimental setup was identical to Study 1, except for exclusion of the Individual condition, and the new Restricted communication condition.In the Restricted communication condition, participants were given the opportunity to chat together using an online chat function after each round, but, importantly, they were instructed to not chat about the die-roll task.Instead, participants in this condition were given specific neutral topics to discuss.To reduce variance between conditions, we based the topics on the filler items in the Nocommunication condition (e.g., What is your favorite hobby?).

Measures
We used the same items as in Study 1 to capture participants' experienced collaboration with the other player and also included a measure of participants' sense of commitment (Michael et al., 2016b).Inspired by Michael et al. (2016a), we included three items in terms of 1) how likely the participants think it is that they would help the other participant if they were to collaborate again (Likert-type scale from 1 = very unlikely to 7 = very likely), 2) how likely they think it is that the other participant would help them if they were to collaborate again (1 = very unlikely to 7 = very likely), and 3) how obligated they feel to help the other participant (from 1 = not at all to 7 = very much).The items were used based on the prediction that if participants felt a sense of commitment with their partner, they would be more likely to help (and expect help from) their partner.Following Study 1, the Honesty-Humility subscale (Lee & Ashton, 2004) and socio-demographic measures were also included.

Analysis
We replicated all analyses from Study 1. Inspired by (Zultan, 2012), we further conducted an exploratory analysis testing differences in proportion of coordination on Nash equilibrium under moneymaximization (i.e., Player A reports a 6 and Player B matches Player A's report) 16 over the 10 trials between conditions using Chi-square tests.In addition to analyses for experienced collaboration included in Study 1, mean differences in commitment between conditions were analyzed using a one-way ANOVA.This was followed by separate mediation analyses for the unrestricted and restricted communication conditions testing whether the effect of the two communication conditions on number of reported doubles was mediated by experienced  collaboration.We further ran a similar exploratory mediation analysis for sense of commitment.Similar to Study 1, the chatlogs in both communication conditions were cleaned and analyzed based on word frequencies in terms of top 30 words written in the chats (see SM section 3.2.8.).
There was no significant difference between the two communication conditions (z = − 1.84, adj, p = .198,d = 0.25).However, an exploratory chi-square analysis revealed that there was a higher proportion of coordination on the Nash equilibrium under money maximization in the Unrestricted condition (20.47%) compared to both the Nocommunication (8.14%; χ 2 (1) = 71.362,p < .001)and the Restricted communication condition (11.62%; χ 2 (1) = 29.157,p < .001),indicating higher levels of coordination between participants on maximizing their payment, when task-related communication was allowed.
Similar to Study 1, an exploratory mixed effects model predicting number of reported doubles by condition and round number revealed no significant effect of round (odds ratio = 1.02, 95% CI [0.99, 1.05], p = .127,see SM section 3.2.3.).Results for magnitude of cheating, proportion of brazen players, and linear mixed effects models predicting  reports based on partners behavior, are also presented in SM (section 3.2.4.-3.2.6.).
In line with Study 1, we found a significant partial mediation effect of experienced collaboration on the effect of Unrestricted communication, supporting H3.The bootstrapped unstandardized indirect effect was ACME = 0.472 and the 95% confidence interval ranged from 0.28 to 0.70.The direct effect of Unrestricted communication was ADE = 1.287 (95% CI [0.80; 1.77]).The effect of Restricted communication was likewise partly mediated by experienced collaboration.The mediation analysis revealed a significant bootstrapped unstandardized indirect effect of ACME = 0.089 (95% CI [0.16; 0.19]).The direct effect of Restricted communication was ADE = 0.869 (95% CI [0.40; 1.37]).All mediation analyses used 1000 simulations to bootstrap.
We performed similar exploratory analyses for commitment.There was a significant indirect effect of Unrestricted communication through commitment (ACME = 0.351, 95% CI [0.17; 0.57]), with a significant direct effect of ADE = 1.416 (95% CI [0.97; 1.88]).Thus, the effect of Unrestricted communication was partly mediated by commitment.The effect of Restricted communication was, however, not mediated by commitment.The indirect effect was not statistically significant (ACME = 0.069, 95% CI [− 0.01; 0.16]) (see SM Fig. S4 section 3.2.7.for overview of mediations for experienced collaboration and commitment).

General discussion
Instead of deterring dishonesty, our results suggest that including communication facilitates dishonesty in collaborative settings.Overall, Study 1 successfully replicated the core findings of Weisel and Shalvi (2015), finding that participants engaged in a dyadic die roll paradigm cheated more for monetary gains compared to when people played alone.Study 1 offered further compelling evidence indicating that communication between participants increases the magnitude of cheating beyond the original study's non-communicative settings.Importantly, this effect was linked to a stronger experienced collaboration among the communicating dyads, highlighting that communication is not only key to everyday ethically sound collaborations, but also to corrupt collaborations.Study 2 replicated this finding, showing that communication significantly increases collaborative cheating compared to non-communicative contexts, even when coordination is restricted.Both commitment and experienced collaboration mediated the effect of unrestricted communication on dishonesty.However, when dyads could chat but were not allowed to coordinate on the task, dishonesty was only mediated by the experience of collaborating and not a sense of commitment.
Study 2's findings illuminate the role of communication in collaborative dishonesty.Similar to the presumptions of prior research (Beck et al., 2020;Kocher et al., 2018), the ability to communicate about the task, potentially enabling coordination and justification exchange, resulted in the highest levels of dishonesty.Interestingly, this was not significantly different from dyads who were not allowed to discuss the task.This suggests that discussing everyday topics alone might increase collaborative dishonest behavior, regardless of coordination and justification exchange.This effect was partly linked to experienced collaboration, but not significantly connected to commitment.
However, despite the non-significant difference in number of reported doubles between the two communicative settings in Study 2, the results also indicate that restricting the communication to non-task 17 See also Table S15 in SM section 3.2.7.for an overview of results and pairwise comparisons.related topics reduces the magnitude of the cheating in terms of coordination on obtaining a maximum payoff.This is in line with previous research studying the differential effect of task-related and non-taskrelated communication in collaboration.For instance, Zultan (2012) found that removing strategic communication by excluding gamerelated discussion led to similar levels of offers in the ultimatum game, while the responders were more willing to accept lower offers when communication was restricted.However, there was a slightly higher proportion of equal splits between participants when communication was not restricted, indicating higher coordination on a mutually beneficial outcome for the participants (Zultan, 2012).Together with the study by Zultan (2012), our study suggests that behavior under different communicative settings may be related to different mechanisms.While we did not find a significant effect of commitment when communication was restricted, other mechanisms, such as prosocial concerns for the other person may be at play.Thus, rather than looking at a unified explanation for the effect of communication, more research into the differential mechanisms is important, both when it comes to ethical and unethical collaborations, respectively.

Limitations
Besides certain differences between the present studies and the original study of Weisel and Shalvi (2015),18 some general limitations deserve consideration.The first is related to the inherent nature of experimental studies prioritizing high internal validity at the expense of external validity.Transitioning from die roll tasks to real-life cases of collaborative dishonesty outside the lab is indeed a substantial leap.Some studies have found that dishonesty in the lab correlate positively with real-world dishonesty (e.g., Dai et al., 2018;Schild et al., 2021).However, because communication between individuals has often been a necessary prerequisite for corruption to evolve in many major fraud scandals, including communication into collaborative paradigms is a step towards increasing the generalizability towards real-world unethical collaborations.
Secondly, another issue pertains to the demand-effects associated with the die roll task.Cheating paradigms often provide participants with ambiguous instructions that may indicate an expectation for cheating, potentially influencing their behavior (for similar concerns, see Frollová et al., 2021;Heyman et al., 2020;Karg, 2021).Nevertheless, this does not diminish the actual dishonesty of misreporting in the present studies.Furthermore, people differ in how (dis)honest they are (Thielmann et al., 2023) and several studies have shown that behavior in cheating paradigms is also related to basic personality traits, such as the Honesty-Humility trait (e.g., Pfattheicher et al., 2019;Ścigała et al., 2020;Thielmann & Hilbig, 2015).If dishonesty in cheating paradigms were only due to demand effects, it would arguably be unlikely that interindividual differences would be linked to basic personality traits. 19ndeed, in line with previous research, dishonesty was negatively correlated with the trait of Honesty-Humility in the present study.
Thirdly, as previous research has shown that coordinating in a die roll task increases trust and social bonds (Karg et al., 2023), the direction of the mediations in the present studies may be reversed.More research that disentangles the relationship between communication, collaboration, and dishonesty is needed.Lastly, our commitment items were focused on future help, which may not fully capture the commitment fostered by communication.Instead, the effect could be related to relationship building.The restricted communication topics constituted different forms of self-disclosure (i.e., the sharing og personal information; Cozby, 1973), that allowed the participants to get to know each other.In fact, previous research has indicated that people become more generous towards others the more they know about them (Bohnet & Frey, 1999;Charness & Gneezy, 2008).Further research into the role of such social exchanges in collaborative dishonest behaviors is warranted.
In conclusion, this paper highlights two key directions for future research.First, it emphasizes the role of communication in shaping collaborative dishonest behaviors, urging future investigation into its various mechanisms.Although Study 2 offers initial insights into such mechanisms, future studies should differentiate between the exchange of justifications, coordination, commitment, and other factors such as relationship building to understand how communication influences collaborative dishonesty.Specifically, it is crucial to further develop strong measures and manipulations to help differentiate between relevant mechanisms.This also relates to measures of experienced collaboration and commitment in these settings.
Additionally, exploring situations where communication not only facilitates but rather hinders collaborative dishonesty is pivotal.How strategic (unrestricted) communication may alter the beliefs and  preferences for (dis)honesty of players in these settings is especially important.That is, under which circumstances may strategic communication be more effectful in either persuading others to act dishonestly, or more importantly, when may it serve as an important tool to talk someone out of dishonesty?Furthermore, the reduced coordination when communication was restricted to non-task related topics warrants further examination.
Another interesting avenue is to study the impact of communication in a more real-word context, where individuals can perceive both their own and other's reputations as being even more at stake.For instance, research has suggested that reducing distance and anonymity may decrease dishonesty (Conrads & Lotz, 2015).Conversely, face-to-face communication may increase the effect of communication.Indeed, face-to-face communication has previously been related to stronger effects than those observed when using written communication (Brosig, Weimann, & Ockenfels, 2003;Zultan, 2012).Thus, studying the impact of communication in face-to-face interactions in the lab would provide important knowledge on how communication impact dishonesty when anonymity is reduced.As collaborative endeavors rarely happen without prior interaction and communication, future research investigating how these behaviors unfold in already established social groups (e.g., employees or friends), would further help move the field of collaborative dishonesty closer to real-life corrupt collaborations.Although few studies have included friends (e.g., Okano & Goto, 2023), we are not aware of studies investigating these behaviors among colleagues in organizations.
Second, given that our results reveal different levels of experienced collaboration between conditions, future research should consider variations in collaboration.Investigating how different levels of collaboration are related to distinct forms of corrupt behaviors, will offer important insights for mitigation strategies aimed at hindering corruption and corporate fraud.Indeed, some collaborations may be initiated by private, self-directed goals, other may be based on the goal of the group (Tuomela, 2020), potentially giving rise to different forms of corrupt collaborations.Finally, if we want to study people colluding on deceiving others together, it is essential understand the circumstances for this type of collaboration and to develop experimental designs that can captures the nuances of collaboration.Our assertion is that the integration of communication serves as the initial stride in advancing our understanding of the collective dimension inherent in collaborative dishonesty.

Fig. 1 .
Fig. 1.Experimental design in Study 1 (a) and Study 2 (b).In Study 1 participants were randomly assigned to either an individual condition, a no-communication condition, or a communication condition.After a practice round, participants would begin phase 2, consisting of 10 trials of a die-roll task adapted from previous work byWeisel and Shalvi (2015).After completing the trials, participants indicated their subjective level of experienced collaboration in the dyadic conditions.Study 2 followed a similar procedure.Participants were randomly assigned to a no-communication condition, an unrestricted communication condition (reflecting the communication condition in Study 1) or a restricted communication condition, wherein participants were not allowed to talk about the task and instead were given specific topics to discuss.

Fig. 3 .
Fig. 3. Mediation model of experienced collaboration on the effect of communication on number of reported doubles.

F
statistic based on one-way ANOVA.Overall collaboration is based on an aggregated measure of the three items.Overall Commitment is likewise based on an aggregate measure of three items.NC = No communication, UC = Unrestricted communication, RC = Restricted communication, M = mean, SD = standard deviation, df = degrees of freedom.

Table 1
Mean differences in experienced collaboration in dyadic conditions.

Table 2
Differences in mean experienced collaboration and commitment across conditions.