Monetary incentives and the contagion of unethical behavior

Abstract

We examine how monetary incentives and information about others’ dishonesty affect lying decisions and whether these two dimensions interact with each other. Our experiment consists of a repeated cheating game where we vary the monetary incentives (Low, High, and Very High) and information about others’ dishonesty (With or Without information). We find that dishonesty decreases when payoffs are Very High. Information has only a weak positive effect on average. Conditioning on beliefs, we find that those who overestimate (underestimate) cheating reduce (increase) dishonesty. Information and payoffs do not interact with each other.

1 Introduction

It is traditionally assumed that dishonesty results from the comparison of the expected associated pecuniary benefits and costs (Becker, 1968). However, recent work has shown that there is less dishonesty than predicted by the standard model, suggesting the existence of intrinsic costs associated with cheating (e.g. Abeler et al., 2019; Gerlach et al., 2019). Furthermore, while the standard economic model of crime predicts that dishonesty should rise with stakes, the empirical evidence is not clear-cut. Some studies report that dishonesty increases with material payoffs (e.g. Dreber & Johannesson, 2008; Erat & Gneezy, 2012; Gneezy, 2005; Kajackaite & Gneezy, 2017; Sutter, 2008)¹ while other studies have found no significant effect (Abeler et al., 2019; Andersen et al., 2018; Fischbacher & Föllmi-Heusi, 2013; Gerlach et al., 2019; Gino et al., 2013; Hugh-Jones, 2016; Kajackaite & Gneezy, 2017; Mazar et al., 2008; Wiltermuth, 2011)² or even that dishonesty is reduced with monetary incentives (e.g. Balasubramanian et al., 2017; Cohn et al., 2019; Mazar et al., 2008).³

Empirical evidence also provides mixed findings regarding the role of information on others’ dishonesty. Some studies find a weak effect of information on average dishonesty (Abeler et al., 2019; Akin, 2019; Diekmann et al., 2015; Kroher & Wolbring, 2015) while other studies find no effect (e.g., Fortin et al., 2007). These unclear findings may result from the existence of several forces that go in opposite directions. On the one hand, observing others being (dis)honest may incite individuals to cheat less (more) depending on the social norm established in the reference group, i.e., a “social conformity” effect (e.g., Alm et al., 1999; Andreoni et al., 1998; Gordon, 1989; Myles & Naylor, 1996). On the other hand, being observed by others may induce reputational and shame costs, which may in turn reduce dishonesty, i.e., a “shame” effect (e.g., Erard & Feinstein, 1994; Fortin et al., 2007). Whether the conformity effect outweighs the shame effect remains an empirical question.

We attempt to contribute to the existing literature by investigating experimentally the role of stakes and information on dishonesty. We also investigate the interaction between payoff and information. Our experiment consists of a modified version of the non-strategic “observed game” developed by Gneezy et al. (2018), which allows us to identify individual cheating decisions (see also, e.g., Gibson et al., 2013; Faravelli et al., 2015). Participants observe a number from one to six where each number is associated with a payoff. They can either truthfully report the number they see or lie about it to increase their payoff. Participants play a repeated game of 20 periods, which allows us to measure the dynamics of cheating and assess how beliefs about others’ decision affect this evolution over time with and without information (Rauhut, 2013). We vary the size of the payoff (Low, High, and Very High) and information about others’ dishonesty (With and Without Information).

To our knowledge, no previous work has investigated simultaneously the role of payoffs and information and their interaction in a non-strategic setting. Mitra and Shahriar (2020) examine the interaction between information and stakes in a cheap-talk sender-receiver game. They find that when benefits for lying are low, senders lie more when they are induced to believe in a higher norm of lying compared to a control situation where the norm is not intervened. Yet, when the benefits to lie are raised, shifting the perceived norm downward does not significantly reduce the level of lying.While their game is a one-shot deception game, ours is a repeated cheating game. While their study only considers the effect of observing others, we also consider the effect of being observed by other participants.

To anticipate our findings, we observe that dishonesty is significantly lower in the Very High payoff treatment than in the High and Low treatments. Information has, on average, a positive but relatively weak effect on cheating. We find no evidence of an interaction between information and payoffs. Last, dishonesty increases significantly over time in all treatments.

The remainder of the paper is structured as follows. Section 2 describes the experimental design, and Sect. 3 sets out the theoretical predictions. The empirical results are shown in Sect. 4. Finally, Sect. 5 discusses our main findings and concludes.

2 Experimental design

2.1 Cheating game

Our game is a modified version of the observed game developed by Gneezy et al. (2018). The game is played for 20 periods. In each period, subjects are privately shown six boxes labeled a, b, c, d, e, or f on their screen. The numbers from one to six are randomly assigned to these boxes. Participants have to click on the boxes and report anonymously the number associated with the first box that they opened (e.g.,Fischbacher & Föllmi-Heusi, 2013; Shalvi et al., 2011). A higher reported number implies a higher payoff. Participants can either truthfully report or lie. In line with an expanding body of literature (see the review by Abeler et al., 2019) our controlled laboratory experiment has no penalties, and participants are told that their decisions are anonymous and private (see the instructions in Appendix A.1). The instructions make clear that mistakes are not punished and could be unintentional.

3 Treatments and procedures

We use a $2\times 3$ between-subjects design, as shown in Table 1. The first dimension is the payoff (see Table 2). In the treatments with low payoffs (Low), participants can earn between 10 cents (if they report a one) and 60 cents (if they report a six). In the high-payoff treatment (High), these payoffs are multiplied by 10, so that participants can earn between 1 and 6 Euros. In the very-high payoff treatment (Very High), the payoffs are multiplied by 40.

Table 1 Treatments

Information	Reward			N
	Low	High	Very high
Without	Without/low N = 70	Without/high N = 95	Without/very high N = 95	260
With	With/low N = 90	With/high N = 120	With/very high N = 90	300
N	160	215	185	560

Table 2 Payoffs (in Euros)

Treatment	Reported number
	1	2	3	4	5	6
Low	0.10	0.20	0.30	0.40	0.50	0.60
High	1	2	3	4	5	6
Very high	4	8	12	16	20	24

The second dimension relates to the information about others. At the beginning of the experiment, participants are randomly assigned to a group of five. The composition of the groups remains unchanged over the 20 periods. Without information, subjects do not receive any information about the behavior of the other four group members. With information, participants are informed at the end of each period about the percentage of others in the group who correctly reported their first click. In all six treatments, we elicited subjects’ beliefs about others’ over-reporting after the subject reports the number that they saw. Following Rauhut (2013), we are thus able to establish if beliefs are important in explaining the reaction to information. This elicitation of beliefs was incentivized. With Information, the belief was measured before the receipt of information about others’ behavior.

The laboratory experiment took place at the LABoratory of EXperiments in Economics and Management LABEX-EM (CREM-University of Rennes 1, France) in September 2017, March to May 2018, and September 2018. Ztree was used to program the experiment, and participant recruitment was organized via Orsee (Greiner, 2015). Our sample consisted of 560 participants. All participants are students. Table A.1 in the Appendix shows descriptive statistics. At the end of the experiment, we elicited risk aversion (Holt & Laury, 2002) and let participants play a memory game designed to measure difficulties in memorizing numbers.⁴ Last, participants completed a post-experimental questionnaire on their socio-demographic characteristics, attitudes, and perception and understanding of the experiment.⁵

4 Theoretical predictions

To illustrate the decision that agents face in our cheating game and derive testable hypotheses, we set out a theoretical model of the decision to be dishonest. In each period of our cheating game, participants observe a random number. They can either truthfully report the number seen and be paid m or report a higher number and be paid n (where n depends on the size of the lie). The payment scheme in the Low, High, and Very High treatments is designed such that m and n are multiplied by the same factor x as payoffs rise. Let $I$ be a binary variable indicating whether the agents are informed ( $I=1$ ) or not ( $I=0$ ) about the true fraction of cheaters. In line with the recent literature on dishonesty, the decision to cheat is assumed to be the result of a tradeoff between pecuniary benefits $B$ and non-monetary costs $C$ . The monetary benefits associated with cheating are assumed to rise with factor $x$ such that $B=B\left(x\right)$ . The non-monetary cost function $C$ consists of four elements described below. Agent $i$ will cheat in period $t$ if her material benefits $B$ of cheating outweigh the psychological costs $C$ of cheating such that:

(1) $B\left(x\right)>C\left(k_i,,,G,\left(x\right),,,R,\left(x\right),\times ,\pi ,\left(I\right),,,\widehat{{\beta }_t^j},\left(I\right)\right),j=L,H,$

where $k_i$ is a fixed Kantian cost. It is assumed that individuals are heterogeneous and make “unconditional” commitments to behave honestly (Abeler et al., 2019; Harsanyi, 1980; Kajackaite & Gneezy, 2017). This unconditional commitment, however, is weak in the sense that cheating costs are also sensitive to external factors and social influences (e.g., Abeler et al., 2019; Figuieres et al., 2013; Le Maux et al., 2019; Masclet & Dickinson, 2019).

The “guilt component” $G\left(x\right)$ is a variable lying cost that rises with stakes. It relies on the idea that people want to maintain a positive self-image, which makes lying costly even in the absence of observability (Kandel & Lazear, 1992). This also relates to the theory of self-concept maintenance suggesting that individuals strive to maintain a positive and consistent self-concept by using categorization or rationalization.⁶ However, while categorization and rationalization may work quite well when one lies for small amounts it may be less effective when stakes are higher (see Mazar et al., 2008). Costs $G\left(x\right)$ are thus expected to rise with payoffs $x$ since higher stakes are likely to reduce rationalization and categorization.

The “shame effect”, denoted $R\left(x\right)\times \pi \left(I\right)$ , relies on the idea that being observed as a cheater may induce a cost in terms of shame and potential loss in reputation (e.g. Garbarino et al., 2019). Shame is expected to rise with payoffs $x$ since higher stakes are likely to intensify external pressure (Vrij, 2008) and/or may cause larger costs for experimenters in ‘observed’ cheating games (Fischbacher & Föllmi-Heusi, 2013; Gneezy, 2005). Shame increases with the subjective probability of being observed denoted $\pi \left(I\right)$ . With information ( $I=1)$ , the agents know that their behavior will be displayed (yet anonymously) to others. Hence, we could expect $\pi$ to be larger with information.⁷

The “social conformity” effect, denoted $\widehat{{\beta }_t^j}\left(I\right)$ reflects that people comply with social norms and conform to others’ behavior. The cost of cheating is assumed to decrease with the fraction $\widehat{{\beta }_t^j}$ of other agents who are thought to be dishonest in period $t$ (Rauhut, 2013). This conformity effect is dynamic in the sense that an exogenous variation in $x$ that encourages (reduces) cheating at $t$ also reduces (increases) the conditional commitment to honesty at $t+1$ , which will make cheating even more (less) attractive in t + 2, and so on (see Appendix A.2 for detailed discussion of the model).

To sum up, our expectations about the effect of stakes $x$ and information $I$ are described as follows:

Hypothesis H1 (variable lying cost) Lying costs (either G, R, or both) increase with stakes $\text{x}$ such that an increase in stakes does not necessarily induce more agents to cheat.

Hypothesis H2 (shame effect) Information $\text{I}$ increases shame $\text{R}\left(\text{x}\right)\times \pi \left(\text{I}\right)$ which leads to a decrease in the share of liars.

Hypothesis H3 (conformity effect at the individual level) Lying costs are responsive to beliefs about the true fraction of liars and those beliefs are responsive to the information condition $\text{I}$ .

Hypothesis H4 (interaction related to shaming) Under information, people might feel more observed, which affects how they react to stakes.

Hypothesis H5 (interaction related to conformity) Under information, a change in stakes does not only affect behavior in t but also the social norm in subsequent periods.

Hypothesis H6 (conformity effect at the average level) The share of dishonest agents is impacted by information if and only if the average belief about the true fraction of liars is itself impacted.

5 Experimental results

5.1 The effects of stakes and information

Figure 1a and b shows the percentage of cheaters over time across all treatments.⁸

Fig. 1 The percentage of dishonest subjects over time by payoff condition. Without information covers 5200 observations, and with information 6000 observations. Shaded areas are 95% confidence intervals

Without information, the average percentage of dishonest people over the 20 periods is 20% in the Very High treatment, 31.6% in the High payoff condition and 32.5% in the Low payoff condition. A two-tailed Mann–Whitney test (also used in subsequent results) indicates that the difference between Very High and Low is statistically significant (p = 0.000). No significant difference is observed between High and Low (p = 0.597). As shown in Fig. 1b, a similar pattern occurs with information. The average percentage of liars is 22.9% in Very High, 35.1% in High and 33.6% in Low. The difference between Very High and Low and Very High and High is statistically significant (p = 0.000). No significant difference is found between High and Low (p = 0.325). These results are partly consistent with H1.⁹ Figure 1a and b shows a similar rising trend in dishonesty in all treatments.

Comparing the percentage of cheaters over time across information conditions indicates that information has either no effect on dishonesty (Low) or a small positive effect (High and Very High) (see Figures A1a, A1b and A1c in the Appendix). In the Low treatments, the average percentage of liars is 32% without information and 34% with information (p = 0.487). The analogous figures in High and Very High are 32% vs. 35% (p = 0.016), and 20% vs. 23% (p = 0.029). These findings indicate that providing information does not induce a strong shame effect as hypothesized in H2. They do not indicate that shame effect does not exist at all but that it should be relatively small and that the conformity effect (H3) outweighs such shame effect (H2).¹⁰ Our results are summarized as follows:

Result 1 (a) The percentage of cheaters is significantly lower in the Very High compared to the Low and High treatments. (b) Information has at most a weak positive effect on dishonesty. (c) In all treatments, cheating increases over time.

To provide more formal evidence of our results, we run estimates on the probability of cheating using random-effect probit models. The results are shown in Table 3. Column (1) controls only for treatment variables and the trend effect. Column (2) adds socio-demographic characteristics and attitudes towards dishonesty. Column (3) adds controls for the first number seen. Finally, column (4) controls for both the belief about how many others cheated (Believes X/4 dishonest) and the actual fraction dishonest in the previous period (Actual X/4 dishonest).

Table 3 The marginal effects from dishonesty regressions (RE probit models)

	1	2	3	4
	b/se	b/se	b/se	b/se
Info	0.062* (0.033)	0.049 (0.032)	0.066 (0.050)	0.064 (0.048)
Payoff: low	(ref.)	(ref.)	(ref.)	(ref.)
Payoff: high	0.020 (0.048)	0.003 (0.048)	− 0.009 (0.084)	− 0.012 (0.078)
Payoff: very high	− 0.105** (0.042)	− 0.131*** (0.041)	− 0.180** (0.077)	− 0.178** (0.070)
Period	0.007*** (0.001)	0.008*** (0.001)	0.010*** (0.001)	0.006*** (0.001)
Age		− 0.003 (0.003)	− 0.003 (0.003)	− 0.003 (0.003)
Female (binary)		− 0.060** (0.030)	− 0.077 (0.049)	− 0.074 (0.050)
Economics (binary)		− 0.007 (0.031)	− 0.002 (0.048)	0.007 (0.049)
Self-reported risk aversion (1 = low to 10 = high)		0.009 (0.009)	0.012 (0.014)	0.011 (0.014)
Self-reported morality (1 = high to 10 = low)		0.039*** (0.011)	0.052** (0.020)	0.054*** (0.020)
Self-reported religiosity (1 = no to 10 = high)		− 0.020*** (0.005)	− 0.025*** (0.008)	− 0.025*** (0.007)
Self-reported political orientation (1 = left to 10 = right)		− 0.012* (0.007)	− 0.014 (0.013)	− 0.015 (0.013)
Self-reported financial knowledge (1 = none to 3 = high)		0.025 (0.029)	0.029 (0.045)	0.031 (0.047)
Memory game (2 = low to 9 = high)		0.024** (0.011)	0.036** (0.017)	0.034** (0.017)
First box: 1		(ref.)	(ref.)	(ref.)
First box: 2			− 0.021 (0.017)	− 0.029 (0.018)
First box: 3			− 0.144*** (0.027)	− 0.154*** (0.026)
First box: 4			− 0.273*** (0.046)	− 0.305*** (0.043)
First box: 5			− 0.360*** (0.070)	− 0.404*** (0.065)
First box: 6			− 0.376*** (0.077)	− 0.419*** (0.070)
Believes 0/4 dishonest				(ref.)
Believes 1/4 dishonest				0.053*** (0.014)
Believes 1/2 dishonest				0.079*** (0.021)
Believes 3/4 dishonest				0.088*** (0.025)
Believes all dishonest				0.135*** (0.042)
Actual 0/4 dishonest				(ref.)
Actual 1/4 dishonest				0.015 (0.012)
Actual 1/2 dishonest				0.017 (0.013)
Actual 3/4 dishonest				0.031* (0.017)
Actual all dishonest				0.042 (0.030)
N	11,200	11,200	11,200	10,640

The dependent variable is binary, indicating whether someone lied about the income. Reported are marginal effects from random-effect probit models, calculated as the probability of a positive outcome, assuming that the random effect for that observation's panel is zero. Standard errors clustered at the group level appear in parentheses. Significance levels: *p ≤ 0.1, **p ≤ 0.05,***p ≤ 0.01

The estimated coefficients in column (1) confirm that the percentage of liars is significantly lower with Very High payoffs and that information has a positive but weak effect. The trend variable is positive and significant, indicating that the percentage of liars rises over time. Controlling for demographics in column (2) leaves these effects unchanged except that information is no more significant.¹¹ In line with previous work, we find that women cheat less (e.g., Abeler et al., 2019; Alm & Malézieux, 2021; Grosch & Rau, 2017). A comparison of descriptive statistics shows that men overreport on average in 32% of the decisions, while women overreport 27% of decisions.¹² Religious people and those who consider cheating less justifiable are also less likely to cheat. Risk aversion is not significant. A greater number of correct answers in our memory game is negatively related to dishonesty, suggesting that over-reporting cannot be attributed to memory issues.

Column (3) indicates that dishonesty decreases with a higher true state, which is consistent with previous studies (e.g., Andersen et al., 2018; Gibson et al., 2013; Gneezy et al., 2013). Hence, we find that a higher first number produces less dishonesty (see also Figures A2a and A2b in the Appendix). This trend is similar over the payoff treatments. Column (4) shows that a higher expected fraction of dishonest others is strongly positively correlated to the probability of lying. We acknowledge that causality cannot be established since, for example, liars could also adjust their beliefs according to their own behavior to justify their actions. The actual previous-period behavior of others has no effect on average. Finally, to test whether information and payoffs interact, we include interaction terms. The marginal effects are shown in Table 4). Dishonesty is lower in the Very High treatments, both with and without information. However, these effects are similar across the Information treatments, suggesting no significant interaction, which rules out hypotheses H4 and H5.¹³

Table 4 The marginal effects from dishonesty regressions including interaction terms

Variable	Conditional on	ME	SE
Payoff low	Without info	Ref.	Ref.
	With info	Ref.	Ref.
Payoff high	Without info	− 0.057	0.098
	With info	0.035	0.07
Payoff very high	Without info	− 0.157*	0.091
	With info	− 0.139**	0.065

The dependent variable is binary, indicating whether someone lied about the income. We estimate column 3 from Table 3 including interactions of the information treatment dummy with dummies on the payoff treatment. Reported are marginal effects from random-effect probit models, calculated as the probability of a positive outcome, assuming that the random effect for that observation's panel is zero. Standard errors clustered at the group level appear in parentheses. Significance levels: *p ≤ 0.1, **p ≤ 0.05,***p ≤ 0.01

Our findings are summarized in result 2.

Result 2 (a)In all treatments, the percentage of dishonest subjects falls with the true number. (b) Payoffs and information about others’ behavior do not interact.

6 Information, behaviors, and beliefs

The fact that the net effect of information is positive suggests that conformity effect totally offsets shame effect. To better understand the underlying mechanisms behind this result, we investigate the role played by beliefs. We categorize at the participant-period level to consider that beliefs may change over time. We acknowledge that over- and underestimation may be related to own behavior. However, we are mostly interested in the effect of information on correcting misperceptions.

Figure 2 shows the level of dishonesty by information condition and belief type. We create two groups based on the observed gap between beliefs and truth (see Figure A3 in the Appendix). Underestimators believe that the percentage of dishonest subjects is lower than it actually is, and overestimators believe that it is higher. Over- and underestimators represent 39% and 35% of the sample without information, respectively and 25% and 30% with information. We analyze the effect of information on the behavior of over- and underestimators in the subsequent period.

Fig. 2 The percentage of liars by information and over-/underestimation. Overestimators include 1498 observations and underestimators 704 observations. Shaded areas are 95% confidence intervals

Figure 2 indicates that, on average, overestimators are more dishonest (48%) than underestimators without information (13%) (Δ = 35 ppts, p = 0.000). This gap narrows substantially with information (40% for overestimators and 28% for underestimators, (Δ = 12 ppts, p = 0.000). This gives support to hypothesis H3, i.e., lying costs are not only responsive to beliefs, but those beliefs are also responsive to information about others’ lying.

In the first five periods, the difference between underestimators and overestimators is 23% without information (p = 0.000) and 22% with information (p = 0.000); the corresponding figures in the last five periods are 39% (p = 0.000) and 6% (p = 0.090). Without information, we thus see divergence in the dishonesty of over- and underestimators, while on the contrary there is convergence with information. Overall, there is no fundamental change in average beliefs, as hypothesized in H6. Looking at the size of the changes, underestimators react more strongly to information than overestimators (overestimators: ∆ = − 8 ppts, p = 0.000, and underestimators: ∆ = 16 ppts, p = 0.000).¹⁴ This may explain why we find on average a weakly positive effect of information on dishonesty. Our findings are summarized as follows:

Result 3 (a) Those who overestimated (underestimated) dishonesty in the previous period are more (less) likely to be dishonest themselves. (b) Underestimators react more strongly to information than do overestimators.

7 Discussion and conclusion

A first main finding from this study is that people tend to lie significantly less in the treatments with very high payoffs. Why do participants cheat significantly less when payoffs are very high? A possible explanation, in line with our theory, is that the marginal cost of lying rises with the magnitude of a lie. This explanation relates to the theory of self-concept maintenance suggesting that individuals attempt to maintain a positive self-concept and that the ability to categorize and rationalize behaviors has its own limits. In other words, while categorization and rationalization may work quite well when one lies for a small amount it may be no longer effective when high stakes are involved.

A second possible explanation is that participants may fear the possible consequences of being exposed and that such feeling of being observed may be amplified when incentives are raised (Gneezy et al., 2018). However, we found no evidence of such feeling of being observed from our post experiment questionnaire (see Appendix A4). Only one participant out of 560 reported feeling of being observed. Another possible explanation is that some participants may be willing to cheat more for smaller stakes in order to ensure a minimum payoff.¹⁵ Consequently, they may not need to lie in the very high treatment, as a relatively low number already gives them a fair payoff but would do so when stakes are low. Interestingly, our post-experimental questionnaire tends to confirm this interpretation showing that 9.15% of participants mentioned that they lied to ensure a minimum payoff (see Appendix A4). This may potentially explain differences across treatments but also within the same treatment, where lying is less frequent for higher values of the true number. Attempting to disentangle these possible alternative explanations is beyond the scope of this current study and may constitute an interesting extension of this work.

Our finding that increasing payoffs decreases lying is consistent with some previous studies (e.g., Balasubramanian et al., 2017; Cohn et al., 2019; Mazar et al., 2008). However, it sharply contrasts with other studies reporting either no (significant) relationship (Abeler et al., 2019; Andersen et al., 2018; Fischbacher & Föllmi-Heusi, 2013; Gerlach et al., 2019; Gino et al., 2013; Hugh-Jones, 2016; Kajackaite & Gneezy, 2017; Mazar et al., 2008; Wiltermuth, 2011) or positive effect of incentives on dishonesty (e.g., Dreber & Johannesson, 2008; Erat & Gneezy, 2012; Gneezy, 2005; Kajackaite & Gneezy, 2017; Sutter, 2008's mind game).

How can we explain these mixed findings? One possible explanation is that these studies are based on different games. Indeed, most studies that report a positive relationship between payoffs and dishonesty relying on interaction games, namely the two-players deception game, in which lying is an option suggested in the rules of the game. In contrast, several studies that use non-strategic games, in which lying is not explicitly mentioned as an option, report either little sensitivity to stakes or even a negative relationship. One may thus argue that participants might be more afraid of possible consequences of a lie in deception games (see the meta-analysis by Gerlach et al., 2019). Making cheating more explicit may make people aware of the harm of their fraud to the other player or experimenter causing them to be more honest when stakes increase. Another important difference across studies is whether the experimenter can observe decisions at the individual level. In most cheating games lying is detected only by comparing the reports with the statistical distribution (e.g. Fischbacher & Föllmi-Heusi, 2013; Kajackaite & Gneezy, 2017). In contrast, in other variants of this game, such as the Gneezy et al (2018)'s observed game or in our current study, the experimenter knows exactly participants’ individual decisions. In such context, one may reasonably argue that participants may fear the possible consequences of being exposed and that such feeling of being observed may be amplified when incentives are raised (Gneezy et al., 2018).¹⁶

A second important finding of our study, in line with previous work (Diekmann et al., 2015; Kroher & Wolbring, 2015; Rauhut, 2013), is that information on others’ dishonesty has, at most, a small positive average effect. This finding suggests that if a shame effect exists, it is offset by a stronger conformity effect. And this conformity effect seems to be asymmetrical, as underestimators react to a little extent more strongly to information than overestimators, implying small net positive effect of information. This finding is in line with Colzani et al. (2023) who find that while lying is contagious, truth-telling is weakly so.

Third, we find that dishonesty rises over time in all treatments. Most studies based on repeated cheating games find similar results (see for instance Gneezy et al., 2013 ¹⁷ and Abeler et al., 2019 for a meta analysis¹⁸). This finding cannot be explained by contagion in behavior, as it also appears in the without information treatment. A possible explanation is that participants may learn more about the rules of the game. However, we doubt that this may be the case since the game was quite simple and answers to our post-experimental questionnaire did not reveal any misunderstanding of the game.¹⁹ One may also reasonably argue that participants may cheat more over time because they realize that nothing happens when they lie. Note however that, as mentioned above, we did not find any evidence of feeling of being observed from our post experiment questionnaire (see Appendix A4). Alternatively, participants’ lying costs may fall over time, as they get used to being dishonest (e.g., Garrett et al., 2016). Thus, a further examination of the role of repetition on dishonesty may constitute an interesting direction for future research. Fourth, we find no evidence for an interaction effect between information and payoffs. While dishonesty is lower in the Very High treatment, there is no evidence of more honesty with information. A possible reason is that the effect of information is relatively weak on average, which as suggested by our theory, may make any interaction effect difficult to identify. Our result contrasts with Mitra and Shahriar (2020). An important difference is that in their study participants learn about the behavior from another session in which lower/higher payoffs were paid. It should be noted that our results align with their study in that learning about higher cheating has a stronger effect than learning about lower cheating.