Remembering the bad ones: Does the source memory advantage for cheaters influence our later actions positively?

Abstract

Previous research has demonstrated that people remember information that is (emotionally) incongruent to their expectations, but it has left open the question if this memory enhancement has also an influence on our later actions. We investigated this question in one pilot study and two experiments. In all studies, participants first interacted with trustworthy and untrustworthy looking partners in an investment game. Facial trustworthiness was manipulated to stimulate social expectations about the behaviour of the partners. In a later second investment game, participants played against old opponents from the first game and new opponents. Overall, willingness to cooperate in the second game was influenced by the formerly behaviour of the opponent. However, facial trustworthiness affected economic decisions, too. Furthermore, we analysed source memory data that indicated no differences in memory between cheaters and cooperators. Instead, source guessing was related to cooperation: The more participants guessed that an untrustworthy looking face belonged to a cheater, the less they cooperated with untrustworthy looking opponents. Interestingly, in Experiment 2, we found a positive correlation between old-new recognition and later cooperation. In sum, the results demonstrate that memory and guessing processes can influence later decisions. However, economic decisions are also heavily affected by other social expectations like facial trustworthiness.

Keywords

Cheater recognition cooperation source memory trustworthiness

From an evolutionary point of view, social cooperation among unrelated individuals is still puzzling (Bell et al., 2010). Mutual cooperation is a common phenomenon among humans (Fehr & Fischbacher, 2004), so it seems obvious that individuals benefit from it. However, cooperation usually implies some kind of fitness costs for the acting individual and is therefore risky (Trivers, 1971). When explaining this puzzle of human cooperative sentiment, researchers have long relied on reciprocity (e.g., Axelrod, 1984; Trivers, 1971): Individuals should not provide benefits to others unconditionally, but should only cooperate when there is the possibility of a favour being returned later on (Axelrod, 1984; Axelrod & Hamilton, 1981; Tooby & Cosmides, 1992; Trivers, 1971). This, however, also implies that cooperators must avoid exploitation from cheating individuals. According to Barclay (2008), different cognitive abilities are required to ensure that cooperation is a stable behaviour. More precisely, humans should have “cognitive abilities that can solve two tasks: (1) detecting instances of non-cooperation (cheating detection) and (2) remembering who has been cooperative and who has not (cheater recognition) and interacting preferentially with other cooperators” (Barclay, 2008, p. 818).

Whereas there is a lot of research concerned with memory advantage for recognising cheaters and research interested in economic decisions, only few experimental studies investigate the postulated effect of supposedly enhanced memory for cheaters on actual decisions in social cooperation. In the following experiments, we thus investigated the influence of (source) memory and appearance-based expectations on economic decisions in social cooperation games.

(Source) memory for cheaters

According to social contract theory (Cosmides, 1989; Tooby & Cosmides, 1992, 2005), a specialised cognitive module has evolved that comprises a cheater detection algorithm. This idea is supported, for instance, by the finding that individuals are better able to solve the Wason Selection Task (Wason, 1966) when it is presented as a cheater detection situation instead of an abstract logical problem (Cosmides, 1989; Gigerenzer & Hug, 1992). Moreover, previous research on source memory for faces of cheaters, that is memory for the cheating context in which these faces were previously encountered, showed that source memory for faces of cheaters is enhanced as compared with faces of people described as trustworthy or cooperative (e.g., Bell et al., 2010; Buchner et al., 2009). In these studies, the moral status of these faces (i.e., source of faces) was manipulated either by using short descriptions in which the stimulus faces were associated with cheating behaviour, trustworthy/cooperative behaviour, or neutral behaviour, or by using a social-cooperation game in which the interactants cheated, cooperated, or showed neutral behaviour. The test phase included old photographs (targets) and new photographs (distractors) and the participants were asked if they had already seen the face (i.e., “old”) or not (i.e., “new”). Given the answer was old, participants were additionally required to indicate whether each face belonged to a cheater, to a trustworthy person, or to a neutral person (not every study included this last category).

Although there is evidence in favour of some kind of cheater detection module, more recent research suggests that more general mechanisms are responsible for the effects found: better performance in the Wason Selection Task when framed as a cheater detection situation can be attributed to, for example, greater relevance as compared with its abstract version (Carlisle & Shafir, 2005), or to the influence of world knowledge and syntactic cues (Ayal & Klar, 2014). Regarding enhanced source memory for cheaters, Bell et al. (2012), for example, demonstrated that source memory is not generally better for the context of cheating behaviour, but that source memory is enhanced for a certain behaviour if this behaviour is somehow unexpected. A more general memory mechanism which prioritises the processing of unanticipated information seems more sensible in light of further corroborative evidence (see, for example, Barclay, 2008; Kroneisen & Bell, 2013; Kroneisen et al., 2015).

To summarise, these memory studies showed that even after a very short contact with another person, individuals are able to memorise some of the past behaviour of this person. Cheating or cooperation in earlier encounters seems to have important informational value, especially when it is unexpected. Yet, a detection module specifically for cheaters would fail to take into account the importance of new information that violates expectancies, which people are able to memorise especially well (Bell et al., 2012). In particular, there seems to be a sensitivity for violations of positive expectations (Kroneisen et al., 2015).

Trustworthiness as cue in economic decisions

Besides memory that is assumed to affect decisions, various factors can influence economic decisions; for instance, when considering social interactions, information on others’ trustworthiness may play an important role. Delgado et al. (2005), for example, showed that additional information about three fictional partners indicating good, bad, or neutral moral behaviour led participants to make more risky choices with partners introduced as morally good. Participants were also more likely to share money with them in comparison to the other partners formerly introduced as bad or neutral even after the expectation of good behaviour was violated. Another study demonstrated that after the possibility to speak with the two other gaming partners for 30 min, participants were more likely to cooperate in one-shot prisoner`s dilemmas given they perceived the partners as cooperative (Frank et al., 1993).

Facial appearance as indicator of trustworthiness

Besides direct experiences in a game, Chang et al. (2010) showed that facial trustworthiness as an implicit social signal for trust can also influence the investment behaviour in a trust game. In the first round of a cooperation game, participants invested more, when the opponent looked trustworthy. In social-dilemma games, too, participants rather defected against untrustworthy looking partners and cooperated with trustworthy looking partners (Rezlescu et al., 2012; van’t Wout & Sanfey, 2008). Facial trustworthiness thus seems to be used as a reliable cue for later behaviour by triggering expectations and stereotypes. When only seeing a face we already draw conclusions about the trustworthiness of this person (Winston et al., 2002). Impressions based on appearance are formed quickly (Todorov et al., 2009; Willis & Todorov, 2006) and automatically (Engell et al., 2007). Apparently, there is also a high consent about the question who looks trustworthy and who does not (Todorov, 2008). Moreover, Chang et al. (2010) found that trustworthiness and experience interact: trustworthy looking opponents that also showed trustworthy behaviour were entrusted with the most money in a trust game.

Overall, these studies suggest that impressions, reputation, and experience may influence people’s economic decision making. Trust in a partner is determined by two things: (1) an initial trustworthiness rating influenced by social signals (e.g., facial trustworthiness) and (2) subsequent experience or reputational information about this partner.

Interaction effects of memory and (facial) trustworthiness on economic decisions

According to the adaptive memory framework (e.g., Buchner et al., 2009; Kroneisen & Bell, 2018; Nairne & Pandeirada, 2016), memory did not only evolve for the purpose of remembering, but to inform judgements and decisions based on the remembered information. Bell et al. (2012), for instance, demonstrated that facial trustworthiness in combination with reputational information influences memory: unexpected information was remembered especially well. Participants showed very good memory for trustworthy looking cheaters and for untrustworthy looking cooperators. Yet, it is still an open question how exactly memory interacts with beliefs, for example, about the moral behaviour of a partner and their influence on economic decisions. As outlined above, people are better able to memorise cheaters in comparison to trustworthy or neutral opponents even after very short contact and also among distractor faces in a classical source-memory design. In the economic decision studies discussed, however, participants only encountered a small number of players. Therefore, memorising their faces is not a difficult task. Initial findings by Suzuki et al. (2013) demonstrated that formerly learned bad lenders persist in memory even after verbal extinction in comparison to good borrowers. In their experiments, however, participants only played against old opponents. Murty et al. (2016) also investigated if episodic memory can influence decision-making. They found that participants had better memory for opponents making less fair offers in a Dictator game. Furthermore, they were more likely to reengage with old opponents in a later decision task if these opponents made more favourable offers during the preceding Dictator game. Schaper et al. (2019), too, found evidence that the willingness to cooperate in a trust game may be determined by the memory for the previous behaviour of opponents. Both, the results from Murty et al. (2016) and Schaper et al. (2019), underline the importance of source memory in adaptive decision making.

Combining the different strands of research, in one pilot study and two main experiments, we examined the interaction of memory and appearance–based expectations on economic decisions. Mainly, we were interested in two things: (1) how much does the behaviour in an investment game depend on memory for the previous experience with the partner and (2) what influence do moral expectations like facial trustworthiness have on both, source memory and economic decisions? Similar to Bell et al. (2012), we combined the two factors facial appearance (trustworthy vs. untrustworthy) and behaviour in a trust game in the encoding phase. According to the incongruity effect (Bell et al., 2012; Kroneisen et al., 2015), unexpected combinations of facial trustworthiness and actual behaviour (e.g., a trustworthy looking cheater) should be remembered especially well. With regard to economic decision, we expected source memory to influence participants’ investments: participants should defect more when confronted with formerly introduced cheaters as compared with cooperators. Based on the incongruity effect, participants should especially defect more when confronted with trustworthy looking cheaters in comparison to untrustworthy looking cheaters and invest more in formerly learned untrustworthy looking cooperators as compared with trustworthy looking cooperators. Moreover, based on the results discussed above, facial trustworthiness seems to be an important cue for trust. Therefore, participants might invest more when confronted with trustworthy looking opponents as compared with untrustworthy looking opponents.

Pilot study

The pilot study’s objective was to investigate the effects of facial trustworthiness and prior behaviour of fictitious opponents on economic decisions involving these opponents. We set the stage for the two subsequent experiments that tested the role of source memory.

Participants

Thirty-nine students (27 females) of the University of Mannheim participated. They received a participation credit and 1% of the money they earned during the two games. Their age ranged from 18 to 35 years (M = 21.92 years; SD = 3.19 years).

Apparatus and materials

Thirty-six coloured portrait photographs (512 × 768 pixels) of male White adults were taken from the FERET database (Phillips et al., 1998). The faces were selected based on a norming study conducted at the Heinrich-Heine University in Düsseldorf.¹ The same norming study was already used to select pictures in other experiments (see, for example, Bell et al., 2012; Mieth et al., 2016). All faces were unfamiliar to the participants. Only forward-facing pictures with a neutral facial expression were selected. Mean a priori trustworthiness ratings for the selected 36 faces were M = 4.18 (SD = 0.18) for faces with high a priori trustworthiness and M = 2.20 (SD = 0.14) for faces with low a priori trustworthiness.

The faces were randomly assigned to two sets. One consisted of 24 photographs and was presented in the debt game (see the following for details). Here, participants saw 12 trustworthy looking pictures and 12 untrustworthy looking pictures. Out of these, 6 of the trustworthy and 6 of the untrustworthy looking pictures were combined with the behaviour of a good lender or bad lender in the debt game, each. The second set consisted of 12 photographs and served as new pictures in the two-player Public Goods game (again, half of them received high facial trustworthiness ratings and the other half received low ratings in the norming study).

Procedure

Debt game—learning phase

Our debt game used in the encoding phase was similar to the debt game used by Suzuki and Suga (2010). Participants were first informed that they had to invest in a friend’s enterprise. To do so, they had to borrow money from a lender. On each trial, participants were first asked to rate the likeability of the lenders face. Each trial started with a headline (“How likeable do you find this person?”) and a photograph shown for 2 s. Subsequently, the likeability rating scale, ranging from 1 (not likeable) to 6 (extremely likeable), was shown. Participants made their choice by using the computer mouse. After that, participants had to decide whether to borrow €15 or €30 from a lender whose face was again presented on the monitor. The participants were aware that after investing the borrowed money, they would temporarily earn a 20% profit. Then, the lender demanded repayment. The participants were told that there were different lenders who had been classified as good or bad beforehand. Good lenders demanded only the money they lent in repayment (i.e., no interest), yielding the 20% profit to the participants. Bad lenders demanded the money they lent plus 40% (i.e., high interest), yielding a 20% net loss to the participants. After participants played 24 trials of the game and had a short pause of 20 s, they were informed that they now had to complete the second part of the experiment, a two-player Public Goods game.

Public goods game—testing phase

Here, each participant played a two-player Public Goods game with 36 opponents. Out of these, 24 had been presented during the encoding phase (12 bad lenders, 12 good lenders) and 12 were new persons. Participants were informed that they played for real money, and that their decisions influenced the amount of money they would receive at the end of the study. They were told that they would receive 1% of the money earned during the Public Goods game. They were also informed that the old opponents would behave in line with their behaviour in the debt game. The game played was very similar to the game used in several other studies (see, for example, Bell et al., 2010, 2012, 2015). On each trial, participants were first asked to rate the likeability of the presented face. Each trial started with a headline (“How likeable do you find this person?”) and a photograph shown for 2 s. Subsequently, the likeability rating scale (ranging from 1 [not likeable] to 6 [extremely likeable]) was shown. Participants made their choice by using the computer mouse. On the next screen, the participant saw a black and white contour, representing themselves, on the left side of the screen. On the right side of the screen, opposite to the participant’s contour, a facial photograph of the opponent was shown. The current account balance of the participant was shown under the participant’s contour. In every trial, participants had to decide whether to cooperate or not. More precisely, they had to decide whether to invest either €0 or €30. Once confirmed, the selected amount was shown on the screen, at the same time the opponent’s investment appeared. An opponent introduced as good lender during the encoding phase invested the same amount of money as the participant. An opponent introduced as bad lender during the encoding phase invested nothing. If the opponent was a new person, he randomly invested either €0 or €30. On the next screen, the sum of investments, a profit, and the total sum appeared. The profit was always 20% of the sum of both investments. The sum of the investments and the profit were added up and the resulting total sum was split up between the participant and the opponent. Both received half of the total sum regardless of their investments. For example, if a participant decides to invest €30, an opponent induced as bad lender during the encoding phase would invest €0. Accordingly, the sum of investments would be €30. The profit would be €6, and the total return would be €36. Each opponent would receive €18. Thus, the participant would incur a loss of €12, whereas the bad lender would gain €18. Thus, depending on the investment of the participant and the opponents’ behaviours, participants could win or lose money. Finally, the participant’s current account balance was updated, and the participant could initiate the next trial by pressing the space bar.

Design

A 3 × 2 design was used with behavioural history (bad lender vs. good lender vs. new) and trustworthiness (trustworthy vs. untrustworthy) as within-subject factors. Likeability ratings and investments during the two-player Public Goods game were the dependent variables.

Results

Likeability ratings

During the Debt Game (learning phase) trustworthy looking persons were rated more likable than untrustworthy looking persons, a 2 (behaviour: bad vs. good lender) × 2(trustworthiness) repeated measurement analysis of variance (ANOVA) revealed a significant main effect of trustworthiness, F(1, 38) = 522.51, p < .001, $η_{p}^{2} = . 93$ . There was no difference in likeability depending on later behaviour, F(1, 38) = 2.64, p = .11, $η_{p}^{2} = . 07$ . In addition, there was no significant interaction between facial trustworthiness and later behaviour, F(1, 38) = .25, p = .62, $η_{p}^{2} = . 006$ .

During the Public Goods game (test phase), trustworthy looking persons were rated more likable than untrustworthy looking persons, a 3 (behavioural history: good lender vs. bad lender vs. new) × 2 (trustworthiness) repeated measurement ANOVA revealed a significant main effect of trustworthiness, F(1, 38) = 220.92, p < .001, $η_{p}^{2} = . 85$ . The previous association with specific behaviour (bad vs. good lender vs. new) had an effect on likability, F (1.62, 61.35²) = 20.12, p < .001, $η_{p}^{2} = . 35$ . Descriptively, good lenders and new opponents were rated more likeable than bad lenders (M_{bad; trustworthy} = 3.37, SD = 0.68; M_{good; trustworthy} = 3.67, SD = 0.64; M_{new; trustworthy} = 3.75, SD = 0.66; M_{bad; untrustworthy} = 1.96, SD = 0.50; M_{good; untrustworthy} = 2.21, SD = 0.46; M_{new; untrustworthy} = 2.38, SD = 0.58). There was no significant interaction between the two factors, F(2, 76) = 0.23, p = .79, $η_{p}^{2} = . 006$ .

Investments in the debt game (learning phase)

In the learning phase, participants invested more in the game when the opponent looked trustworthy than when he looked untrustworthy, $F (1, 38) = 306.11, p < . 001, η_{p}^{2} = . 89$ . As the individual opponent has not been identified as good lender versus bad lender at the time of investment, there was no effect of behavioural history, $F (1, 38) = 1.04, p = . 32, η_{p}^{2} = . 03$ , and no interaction, $F (1, 38) = 1.46, p = . 23, η_{p}^{2} = . 04$ .

Public goods game investments³ (testing phase)

The average investments depending on behavioural history (good lender vs. bad lender vs. new) and facial trustworthiness (trustworthy vs. untrustworthy) in the testing phase can be seen in the upper panel of Figure 1.

A 3 (behavioural history) × 2 (trustworthiness) repeated measurement ANOVA revealed a significant main effect of behavioural history, $F (2, 76) = 21.50, p < . 001, η_{p}^{2} = . 36$ . Participants spent less money when playing against an opponent identified as bad lender in the learning phase in comparison to a good lender, $(t (38) = - 4.75, p < . 001, d = - 0.76$ , or a new opponent, $(t (38) = - 6.29, p < . 001, d = - 1.01$ . There was no difference in investment between good lenders and new opponents, t(38) = 1.54, p = .13, $d = - 0.25$ .⁴ In addition, the ANOVA revealed that participants invested more in trustworthy looking opponents than in untrustworthy looking opponents, $F (1, 38) = 68.51, p < . 001, η_{p}^{2} = . 64$ . The two main effects were accompanied by a significant interaction, $F (2, 76) = 10.71, p < . 001, η_{p}^{2} = . 22$ . Post hoc tests showed that given trustworthy looking opponents participants spent less money to bad lenders than to good lenders $(t (38) = - 4.26, p < . 001, d = - 0.68)$ or new opponents $(t (38) = - 7.64, p < . 001, d = - 1.22)$ . There was also a difference between good lenders and new opponents ( $t (38) = - 3.39, p = . 005, d = - 0.54$ . For untrustworthy looking opponents, there was an equivalent difference in investments between bad lenders and good lenders $(t (38) = - 2.84, p = . 02, d = - 0.45)$ , but no difference between bad lenders and new opponents $(t (38) = - 1.75, p = . 17, d = - 0.28)$ or between good lenders and new opponents $(t (38) = 1.09, p = . 28, d = 0.17)$ .

Discussion

Similar to other studies (Mieth et al., 2016; van’t Wout & Sanfey, 2008), our results suggest that individuals are less likely to invest money when their opponent looks untrustworthy than when he looks trustworthy. Moreover, the opponent’s previous behaviour seems to influence investments: participants invested more in the Public Goods game when the opponent was experienced as good lender or new as compared with an opponent experienced as bad lender in a previous debt game. Contrary to findings by Bell et al. (2012) and Kroneisen et al. (2015) on source memory, however, we did not find an effect of expectation violation. Instead, participants invested less money when a formerly learned bad lender looked untrustworthy as compared with trustworthy and cooperated more often when a formerly learned good lender looked trustworthy as compared with untrustworthy. Borrowing money represents a risky strategy that pays only when one expects the opponent to be nice. Our finding suggests that participants had more positive expectancies towards faces with high a priori trustworthiness.

Still, the effect of the factor facial trustworthiness was reduced by behavioural history in comparison to its effect when confronted with new opponents. For new opponents, only facial appearance could be used as cue. Interestingly, trustworthy looking new opponents got the highest cooperation rate, a result also found in other studies (Bell et al., 2013; Mieth et al., 2016). This might indicate bad source memory for old opponents resulting in a more cautious investment strategy when facing old opponents as well as a high reliance on facial trustworthiness.

Due to these partially unexpected results, in Experiment 1, we tested if these remained stable when using a larger sample size. In addition, we addressed limitations in the pilot study: First, in the Public Goods game of the pilot study’s test phase, participants received feedback on their opponents’ behaviour after each trial. This may have functioned as additional learning potentially influencing further decisions. Second, we did not measure whether the investment decisions participants made were actually connected to source memory (supposedly enhanced for cheaters). Tackling these aspects, we adjusted our procedure for Experiment 1: Our participants first played a game with cheating opponents and cooperating opponents. This was followed by a decision task (participants could decide if they wanted to cooperate or to cheat) including old and new opponents without feedback to ensure that this task did not function as an additional learning phase. Similar to Murty et al. (2016), a classical source-memory test followed after this.

Experiment 1

Method

Participants

Sixty-nine students of the University of Koblenz-Landau participated. One participant had to be excluded because of low task performance. The final sample comprised 68 students (54 females). Participants’ age ranged from 18 to 45 years (M = 21.67, SD = 4.29). They received a participation credit and €2 at the end of the experiment. Given N = 68, α = .05, and 36 responses in the source-memory test, effects of size w = 0.06 (Buchner et al., 2009) of the behavioural history variable on memory could be detected with a probability of 1−β = .84. All power calculations were conducted using G*Power (Faul et al., 2007).

Apparatus and materials

Forty-eight coloured portrait photographs (640 × 480 pixels) of male White adults were taken from the lifespan database of adult facial stimuli (Minear & Park, 2004) which has been extensively used in face research (e.g., Stahl et al., 2008; Wiese, Schweinberger, & Hansen, 2008; Wiese, Schweinberger, & Neumann, 2008). Only forward-facing pictures with a neutral facial expression were selected. All faces were unfamiliar to the participants. The faces were selected on the basis of a norming study conducted at the Heinrich-Heine University in Düsseldorf. The same norming study was already used to select pictures in the pilot study. Mean a priori trustworthiness ratings for the selected 48 faces were 3.72 (SD = 0.40) for faces with high a priori trustworthiness and 2.50 (SD = 0.35) for faces with low a priori trustworthiness.⁵ The faces were randomly assigned to a learning set (24 photographs; 12 faces low in facial trustworthiness and 12 faces high in facial trustworthiness) and two distractor sets (12 photographs; 6 faces low in facial trustworthiness and 6 faces high in facial trustworthiness, each).

Procedure

Public goods game—learning phase

During the first phase of the experiment, 12 trustworthy looking and 12 untrustworthy looking facial photographs were presented in the context of a Public Goods game (learning set). During the game, subjects were supposed to learn from their own experience whether the respective target person showed either cheating behaviour or cooperative behaviour. Thus, the experimental design comprised two independent variables: a priori expectation (trustworthy looking vs. untrustworthy looking) and behaviour in the game (cheating vs. cooperative). Half of the trustworthy looking opponents and half of the untrustworthy looking opponents were randomly assigned to either the cheater condition or the cooperator condition. Consequently, the learning phase comprised 24 trials, with six trials for trustworthy cheaters, trustworthy cooperators, untrustworthy cheaters, and untrustworthy cooperators, each.

The game was similar to that used in previous studies on memory for reputations (Bell et al., 2010; Giang et al., 2012). Each participant started with a deposit of €450. The participant was represented by a black and white shape that was presented on the left side of the screen. On the right side of the screen, opposite to the participant’s contour, a facial photograph of the opponent was shown. The current account balance of the participant was shown under the participant’s contour. First, the participants had to decide whether to invest €10, €20, or €30. On the next screen, the investment of the participant was shown on the left side and the opponent’s investment appeared on the right side of the screen. A cooperator invested the same amount of money as the participant. A cheater invested €0. Then, a bonus was added to the sum of both investments, consisting of 20% of the total sum. At the end of each round, the total money was split by two, with each player receiving half of the total money. Thus, depending on the investment of the participant and the opponents’ behaviours, participants could win or lose money.

After participants played 24 trials of the game and had a short pause of 20 s, they were informed that they now had to complete the second part of the experiment, a decision task.

Decision task—testing phase

Each participant played a game with 36 opponents. Out of these, 24 had been presented before (i.e., target items from the first Public Goods game) and 12 were new (distractor set 1, no previous encounters, 6 trustworthy and 6 untrustworthy looking faces).

First, participants were instructed that a new game would start. They would play against old and new opponents. Old opponents would behave in the exact same way as in the Public Goods game before. In each round, participants could decide if they wanted to invest €0 or €30. In contrast to the test phase of the pilot study, participants could only select the amount of money they would invest, feedback about the behaviour of the opponent or their current account was never given. Participants were told that they would have the opportunity to be rewarded at the end of the experiment if they showed good performance during this game. They were also explicitly told, that, therefore, a good memory for the behaviour of the old opponents would be beneficial.

Memory test

After a short pause of 20 s, the third part of the experiment started, a surprise memory test. During the memory test, we presented 36 faces, 24 of them old (i.e., target items from the first Public Goods game, learning phase) and 12 of them new (distractor set 2, no previous encounters, 6 trustworthy and 6 untrustworthy looking faces). For each target, participants first had to choose whether they “know this person from the first part of the study” via mouse click on one of the two buttons differentiating old and new items (= item memory). Whenever participants endorsed that they recognised the target, they subsequently had to indicate whether this target had behaved “fair” or “unfair” during the Public Goods game (remembering behaviour of the target = source memory).

Design

A 3 × 2 design was used with behavioural history (cheater vs. cooperative vs. new) and trustworthiness (trustworthy vs. untrustworthy) as within-subject factors. Investments during the decision tasks, old-new recognition, and source memory were the dependent variables.

Results

Investments in the learning phase

In the learning phase, participants invested more in the game when the opponent looked trustworthy than when he looked untrustworthy, $F (1, 67) = 54.35, p < . 001, η_{p}^{2} = . 45$ . Again, our findings suggest that participants had more positive expectancies toward faces with high a priori trustworthiness. As there were not any previous encounters yet, there was no effect of behavioural history, $F (1, 67) = 0.75, p = . 39, η_{p}^{2} = . 01$ , and no interaction, $F (1, 67) = 0.03, p = . 86, η_{p}^{2} < . 001$ .

Investments in the testing phase

A 3 (behavioural history: cheater vs. cooperative vs. new) × 2 (trustworthiness) repeated measurement ANOVA revealed a significant main effect of behavioural history, F(1.77,118.54) = 18.60, $p < . 001, η_{p}^{2} = . 22$ As can be seen in the middle panel of Figure 1, participants spent less money when playing against a cheater in comparison to a cooperator, $t (67) = - 5.43, p < . 001, d = - 0.66$ , or a new opponent, $t (67) = - 5.12, p < . 001, d = - 0.62$ . There was no difference in investment between cooperators and new opponents, $t (67) = 0.31, p = . 76, d = 0.04$ .⁶ In addition, participants invested more in trustworthy looking opponents than in untrustworthy looking opponents, F(1,67) = 38.80, $p < . 001, η_{p}^{2} = . 37$ . The two main effects were accompanied by a significant interaction, F(1.83,122.62) = 3.33, p = .04, $η_{p}^{2} = . 05$ . Post hoc tests showed that given trustworthy looking opponents, participants spent less money to cheaters than to cooperators $(t (67) = - 5.90, p < . 001, d = - 0.72)$ or new opponents $(t (67) = - 5.11, p < . 001, d = - 0.62)$ , but no difference between cooperators and new opponents $(t (67) = 0.79, p = . 99, d = 0.095)$ . For untrustworthy looking opponents, there was a difference in investments between cheaters and cooperators $(t (67) = - 2.95, p = . 02, d = - 0.36$ ), cheaters and new opponents $(t (67) = - 3.24, p = . 008, d = - 0.39)$ , but no difference between cooperators and new opponents $(t (67) = - 0.29, p = . 99, d = - 0.04)$ .

Old-new recognition

Old-new recognition is reported in terms of P_r, a sensitivity measure of the two-high threshold model. It is calculated by subtracting the false alarm rate from the hit rate. P_r as a sensitivity measure was evaluated in validation studies (Snodgrass & Corwin, 1988) and avoids the problem of undefined values arising from the use of d.’

A 2 (behavioural history: cheater vs. cooperative) × 2 (trustworthiness) repeated measurement ANOVA revealed no significant main effect of behavioural history, $F (1, 67) = 0.40, p = . 53, η_{p}^{2} = . 006$ . Trustworthiness did not affect old–new discrimination, F(1,67) < 0.001, p = .96, $η_{p}^{2} < . 001$ . The interaction between both variables was also not significant, $F (1, 67) = 2.23, p = . 14, η_{p}^{2} = . 03$ (see Figure 2, top panel).

Source memory and behaviour

There are different ways to measure source memory. However, many ad hoc measures are known to confound source memory with guessing biases (Bayen & Murnane, 1996; Bröder & Meiser, 2007). Therefore, we analysed our data using the multinomial source monitoring model of Bayen et al. (1996). Using this kind of models, it is possible to measure probabilities of underlying cognitive processes assumed to be involved in source monitoring and to estimate independent parameters of item memory, source memory, and guessing from the observed response frequencies in the memory test (Erdfelder et al., 2009; Moshagen, 2010). Multinomial models have been used in many studies examining memory for defectors (e.g., Bell et al., 2012; Buchner et al., 2009; Kroneisen et al., 2015). We used the same model as in Bell et al. (2012, Experiment 1; submodel 5d in the classification of identifiable source-memory submodels of Bayen et al., 1996, which incorporates the assumptions D_{cheatTrustworthy} = D_{coopTrustworthy} = D_{NewTrustworthy} = D_{cheatUntrustworthy} = D_{coopUntrustworthy} = D_{NewUntrustworthy}; a = g). Furthermore, similar to Schaper et al. (2019), instead of aggregating response frequencies across participants, we fitted a Bayesian hierarchical extension of the MPT model (Klauer, 2010; Matzke et al., 2015). This method allows to estimate individual parameters from an overarching group distribution (Heck et al., 2018; Klauer, 2010) and can therefore account for heterogeneity of the participants’ memory and guessing tendencies. Furthermore, this allows us to investigate the correlation between individual parameters and participants’ decisions in the decision task directly.

Hierarchical multinomial modelling is an increasingly popular approach in memory research (Marevic et al., 2018; Schaper et al., 2019) but is also used in other research fields (Kroneisen et al., 2021; Kroneisen & Heck, 2020). In line with Schaper et al. (2019), we used a two-step strategy to test (1) whether source memory and guessing processes differed between trustworthy and untrustworthy looking cheaters and cooperators and (2) whether there is a relationship between these source-monitoring processes and decision making during the decision-task. For all analyses, we used the R package TreeBUGS (Heck et al., 2018), which fits Bayesian hierarchical MPT models using Markov-chain Monte-Carlo methods (Plummer, 2003).

Model fit was assessed with posterior-predicted p values and indicated a satisfactory fit both with respect to the mean (p = .41) and the covariance structure (p = .47) of the observed individual frequencies as tested by T₁ and T₂ statistics proposed by Klauer (2010, see also Heck et al., 2018).

To test for differences in source monitoring depending on behavioural history and trustworthiness, we have to look at the parameter differences of interest. A statistically reliable difference is indicated if the 95% Bayesian Credibility Interval (BCI) for the difference estimate does not contain zero. The corresponding parameter estimates and 95% BCIs are reported in Table 1. Source memory is reflected in the parameter d.⁷ To test for differences in source memory between trustworthy looking cheaters and trustworthy looking cooperators, we sampled the parameter difference Δd_trustworthy = d_Cheat_Trustworthy − d_{CoopTrustworthy}. Source memory did not differ between trustworthy looking cheaters and trustworthy looking cooperators, Δd_trustworthy = .03, 95% BCI = [−0.16, 0.22]. To test for differences in source memory between untrustworthy looking cheaters and cooperators, we sampled the parameter difference Δd_{untrustworthy} = d_CheatUnt_rustworthy − d_{CoopUntrustworthy}. Source memory did not differ between untrustworthy looking cheaters and untrustworthy looking cooperators, Δd_{untrustworthy} = .19, 95% BCI = [−0.08, 0.61]. There was also no difference in source memory between trustworthy looking cheaters and untrustworthy looking cheaters (Δd_cheat = −.15, 95% BCI = [−0.56, 0.14]) or trustworthy looking cooperators and untrustworthy looking cooperators (Δd_coop = .01, 95% BCI = [−0.11, 0.17]).

Table 1.

Parameter estimates (D = probability of recognising a face as old or new, d = conditional probability of source memory in the sense of remembering the context in which a face was encountered, g = conditional probability of guessing that the face belonged to a cheater rather than to a cooperator, b = conditional probability of guessing that an unrecognised face is old) for a priori trustworthiness of the faces (trustworthy [TW] vs. untrustworthy [UTW]) in Experiments 1 and 2.

	Parameter	Parameter estimate	[BCI]
Experiment 1	D	0.78	[.74, .81]
	d _CheatTW	0.09	[.003, .25]
	d _CheatUTW	0.24	[.01, .64]
	d _CoopTW	0.06	[.001, .20]
	d _CoopUTW	0.05	[.001, .14]
	g _TW	0.47	[.39, .55]
	g _UTW	0.72	[.63, .82]
	b _TW	0.28	[.13, .43]
	b _UTW	0.53	[.38, .66]
Experiment 2	D _TW	0.61	[.56, .66]
	d _CheatTW	0.35	[.06, .54]
	d _CheatUTW	0.56	[.10, .94]
	d _CoopTW	0.22	[.01, .49]
	d _CoopUTW	0.05	[.001, .15]
	g _TW	0.42	[.32, .53]
	g _UTW	0.62	[.55, .70]
	b _TW	0.27	[.19, .33]
	b _UTW	0.32	[.24, .40]

BCI: Bayesian Credibility Intervals.

Bayesian Credibility Intervals are reported in the brackets.

We also analysed the guessing parameter g which represents the conditional probability of guessing that a face belonged to a cheater given that the face and/or its context is not remembered. Parameter g can be seen as a measure of participants’ expectancies towards the stimulus faces. Parameter g clearly differed between untrustworthy looking faces and trustworthy looking faces, (Δg = .25, 95% BCI = [0.13, 0.37]). This finding suggests expectancy-congruent guessing, such that participants had a tendency towards guessing that trustworthy looking faces belonged to cooperators and untrustworthy looking faces belonged to cheaters. We also tested if there was a difference in guessing between trustworthy and untrustworthy unrecognised faces. Parameter b represents the conditional probability of guessing that an unrecognised face was shown before. Parameter b differed between untrustworthy looking and trustworthy looking faces, (Δb = .25, 95% BCI = [0.06, 0.44]). For untrustworthy looking faces, participants had a greater tendency to guess that they had seen it during the learning phase.

As mentioned above, we also wanted to test whether source-monitoring processes were correlated with adaptive decision making in the later decision task. Similar to Schaper et al. (2019), we correlated the individual posterior parameter estimates with the proportion of cooperation with trustworthy and untrustworthy looking cooperators, cheaters, and new partners in the decision task,⁸ within the hierarchical model. Table 2 summarises the results.

Table 2.

Correlations between parameter estimates (D = probability of recognising a face as old or new, d = conditional probability of source memory in the sense of remembering the context in which a face was encountered, g = conditional probability of guessing that the face belonged to a cheater rather than to a cooperator, b = conditional probability of guessing that an unrecognised face is old) and the proportion of cooperative decisions as a function of partner type (cooperator, cheater, new) and a priori facial trustworthiness (trustworthy vs. untrustworthy) in Experiment 1.

	Probability of cooperation with. . .
	Cheaters				Cooperators				New partners
	Trustworthy		Untrustworthy		Trustworthy		Untrustworthy		Trustworthy		Untrustworthy
	Corr	[BCI]	Corr	[BCI]	Corr	[BCI]	Corr	[BCI]	Corr	[BCI]	Corr	[BCI]
D	−0.04	[−.22, .14]	−0.03	[−.25, .19]	0.09	[−.09, .27]	0.02	[−.21, .24]	0.08	[−.09, .25]	0.02	[−.18, .21]
d _CheatNTW	−0.09	[−.30, .15]	−0.32	[−.59, .13]	−0.01	[−.21, .22]	−0.30	[−.58, .15]	−0.02	[−.22, .19]	−0.18	[−.42, .15]
d _CheatTW	−0.05	[−.29, .20]	0.01	[−.41, .40]	−0.01	[−.25, .23]	0.07	[−.41, .47]	0.02	[−.20, .24]	0.01	[−.31, .31]
d _CoopNTW	−0.01	[−.26, .24]	0.09	[−.39, .49]	−0.01	[−.25, .24]	0.12	[−.39, .53]	0.01	[−.21, .22]	0.06	[−.28, .36]
d _CoopTW	−0.01	[−.24, .22]	−0.03	[−.41, .35]	0.05	[−.20, .30]	0.00	[−.44, .40]	0.02	[−.20, .24]	0.00	[−.30, .29]
g _NTW	−0.13	[−.27, .02]	−0.56	[−.67, −.43]*	−0.08	[−.24, .07]	−0.62	[−.72, −.50]*	−0.10	[−.24, .05]	−0.37	[−.49, −.23]*
g _TW	−0.12	[−.32, .09]	0.17	[−.10, .42]	−0.17	[−.37, .05]	0.17	[−.12, .43]	−0.04	[−.23, .16]	0.04	[−.19, .27]
b _NTW	0.08	[−.10, .26]	0.12	[−.13, .35]	0.06	[−.12, .23]	0.21	[−.04, .43]	0.09	[−.08, .26]	0.13	[−.05, .31]
b _TW	−0.02	[−.20, .17]	0.09	[−.15, .33]	0.02	[−.16, .21]	0.19	[−.05, .42]	0.02	[−.15, .20]	0.06	[−.13, .26]

BCI: Bayesian Credibility Intervals.

The values in brackets correspond to the Bayesian Credibility Intervals.

Signifies that the credibility interval excluded zero.

We found no relationship between old–new recognition or source memory⁹ and behaviour in the decision task. None of the memory parameters were associated to cooperation, irrespective of a priori trustworthiness. However, source guessing (parameter g) was related to cooperation: The more participants guessed that an untrustworthy looking face belonged to a cheater, the less they cooperated with untrustworthy looking opponents. There was no reliable relationship between old-new guessing (parameter b) and cooperation, irrespective of trustworthiness.

Discussion

In this experiment, expectations were formed by two aspects: facial trustworthiness of the opponent and memory for former behaviour of this opponent. Based on the incongruity effect (Bell et al., 2012; Kroneisen et al., 2015), both factors, trustworthiness and behavioural history, should interact: participants should tend to defect even more often when a formerly learned negative opponent looked trustworthy. Moreover, participants should tend to cooperate more often when a formerly learned good opponent looked untrustworthy. However, replicating the results of our pilot study, in Experiment 1, again, participants invested more money when playing against trustworthy looking formerly learned cooperators and invested the least when playing against untrustworthy looking formerly learned cheaters.

The source-memory data showed no differences in memory between cheaters and cooperators. Instead, participants showed a bias towards guessing that an untrustworthy looking person had cheated in the Public Good game and a trustworthy looking person had cooperated. This challenges the stability of the formerly found source-memory advantage for trustworthy looking cheaters (incongruenty effect; e.g., Bell et al., 2012).

Furthermore, Experiment 1 was designed to test if the investment decisions of participants are connected to source memory. Based on the idea that memory evolved to inform decisions (adaptive memory framework and recent results from Schaper et al., 2019), a correlation between source memory (parameter d) and behaviour in the decision game should be detected. Our experiment showed no influence of source memory, but a clear relationship between source guessing, a priori facial trustworthiness and decisions: For untrustworthy looking opponents, the more participants guessed that a face belonged to a cheater, the less they cooperated with this person.

However, Experiment 1 differed from other experiments (Bell et al., 2012; Schaper et al., 2019) in so far that we included an additional task between learning and memory phases. It is unclear if the time between learning and testing is a critical moderator of the source-memory advantage for unexpected information and of an effect of memory on decisions. To test the hypothesis that the order of memory test and decision task is critical, we replicated Experiment 1, but using the reversed order of memory test before decision task.

Experiment 2

Method

Participants

Participants were recruited via electronic mailing-lists of the University of Mannheim and the University of Koblenz-Landau. Eighty-nine students (69 females) participated. Participants’ age ranged from 18 to 42 years (M = 22.60, SD = 4.10). They received a participation credit at the end of the experiment.

Apparatus and Materials, Procedure, and Design were as in Experiment 1, except that the order of the memory task and the decision task was changed with memory being tested before the decision task. Given N = 89, α = .05, and 36 responses in the source-memory test, w = .06 of the behavioural history variable could be detected with a probability of 1−β = .92.

Results

Investments in the learning phase

In line with the pilot study and experiment 1, the 2 (behavioural history) × 2 (trustworthiness) repeated measurement ANOVA revealed a main effect of trustworthiness: Participants invested more in the game when the opponent looked trustworthy than when he looked untrustworthy $F (1, 88) = 80.39, p < . 001, η_{p}^{2} = . 48$ . There was also a significant main effect of behavioural history, $(F (1, 88) = 4.51, p = . 04, η_{p}^{2} = . 05)$ . Participants spend more money to opponents that cooperated later. There was no significant interaction, $F (1, 88) = 0.72, p = . 40, η_{p}^{2} = . 008$ .

Investments in the testing phase

In line with Experiment 1, a 3 (behavioural history: cheater vs. cooperative vs. new) × 2 (trustworthiness) repeated measurement ANOVA showed a significant main effect of behavioural history, $F (2, 176) = 40.38, p < . 001, η_{p}^{2} = . 32$ . Participants spent less money when playing against a cheater in comparison to a cooperator, $t (88) = - 6.93, p < . 001, d = - 0.74$ , or a new opponent, $t (88) = - 8.42, p < . 001, d = - 0.89$ (see Figure 1, lower panel). There was no difference in investment between cooperators and new opponents, t(88) = −1.49, p = .14, $d = - 0.16$ .¹⁰ In addition, participants invested more in trustworthy looking opponents than in untrustworthy looking opponents, $F (1, 88) = 90.25, p < . 001, η_{p}^{2} = . 51$ . The two main effects were accompanied by a significant interaction, $F (2, 176) = 8.33, p < . 001, η_{p}^{2} = . 09$ . Post hoc tests revealed that given trustworthy looking opponents, participants spent less money to cheaters than to cooperators, t(88) = −4.33, $p < . 001, d = - 0.46$ or new opponents, t(88) = −8.11, $p < . 001, d = - 0.86$ , and spent less to cooperators than new opponents, $t (88) = - 3.78, p < . 001, d = - 0.40$ . For untrustworthy looking opponents, there was a difference in investments between cheaters and cooperators, t(88) = −5.91, $p < . 001, d = - 0.63$ , between cheaters and new opponents, $t (88) = - 4.33, p < . 001, d = - 0.46$ , but no difference between cooperators and new opponents, t(88) =1.58, p = .29, $d = 0.17$ .

Old–new recognition

A 2 (behavioural history: cheater vs. cooperative) × 2 (trustworthiness) repeated measurement ANOVA revealed no significant main effect of behavioural history on P_r, $F (1, 88) = 0.01, p = . 92, η_{p}^{2} < . 001$ . Trustworthiness did not affect old–new discrimination, $F (1, 88) = 1.43, p = . 24, η_{p}^{2} = . 02$ . The interaction between both variables was also not significant, F(1,88) =1.02, $p = . 32, η_{p}^{2} = . 01$ (see Figure 2, lower panel).

Source memory and behaviour

Similar to Experiment 1, we again used a two-step strategy to analyse our data: First, we tested whether source memory and guessing were affected by trustworthiness and behavioural history. Second, we tested whether memory and guessing processes influenced later decision making during the decision task.

Model fit was again assessed with posterior-predicted p values and indicated a satisfactory fit, p_T1 = .34 and p_T2 = .79. The parameter estimates and 95% BCIs are reported in Table 1. To test for differences in source memory between trustworthy looking cheaters and trustworthy looking cooperators, we sampled the parameter difference Δd_trustworthy = d_Cheat_Trustworthy—d_{CoopTrustworthy}. Source memory did not differ between trustworthy looking cheaters and trustworthy looking cooperators, Δd_trustworthy = .13, 95% BCI = [−0.35, 0.49]. To test for differences in source memory between untrustworthy looking cheaters and untrustworthy looking cooperators, we sampled the parameter difference Δd_{untrustworthy} = d_CheatUnt_rustworthy—d_{CoopUntrustworthy}. Source memory did differ between untrustworthy looking cheaters and untrustworthy looking cooperators, Δd_{untrustworthy} = .51, 95% BCI = [0.005, 0.91]. Participants had a better source memory for untrustworthy looking cheaters in comparison to untrustworthy looking cooperators. There was no difference in source memory between trustworthy looking cheaters and untrustworthy looking cheaters (Δd_cheat = −.20, 95% BCI = [−0.67, 0.31]) or trustworthy looking and untrustworthy looking cooperators (Δd_coop = .18, 95% BCI = [−0.06, 0.45]). Again, the guessing parameter g clearly differed between untrustworthy looking and trustworthy looking faces, (Δg = .215, 95% BCI = [0.07, 0.21]) which suggests expectancy-congruent guessing. Parameter b (old-new guessing) did not differ between untrustworthy looking and trustworthy looking faces, (Δb = .05, 95% BCI = [−0.04, 0.15]). There was no difference in guessing that an unrecognised face was shown before.

As in Experiment 1, we computed the correlation between the individual posterior parameter estimates and the proportion of cooperation with trustworthy and untrustworthy looking cooperators, cheaters, and new opponents in the decision task. Table 3 summarises the results. There was a positive correlation between old-new recognition (D) and cooperation given trustworthy looking cooperators and untrustworthy looking cooperators, but not given cheaters. Similar to Experiment 1, we found no relationship between source memory (d) and behaviour in the decision task, irrespective of a priori trustworthiness. Source guessing (g) was related to cooperation: The more participants guessed that an untrustworthy looking face belonged to a cheater, the less they cooperated with untrustworthy looking opponents.

Table 3.

	Probability of cooperation with
	Cheaters				Cooperators				New partners
	Trustworthy		Untrustworthy		Trustworthy		Untrustworthy		Trustworthy		Untrustworthy
	Corr	[BCI]	Corr	[BCI]	Corr	[BCI]	Corr	[BCI]	Corr	[BCI]	Corr	[BCI]
D	−0.05	[−.22, .14]	−0.05	[−.18, .07]	0.18	[.06, .29]*	0.14	[.02, .26]*	0.11	[−.01, .22]	0.12	[−.003,.23]
d _CheatNTW	−0.02	[−.23, .20]	−0.16	[−.17, .12]	−0.12	[−.31, .10]	−0.08	[−.29, .16]	−0.03	[−.23, .18]	−0.05	[−.25, .16]
d _CheatTW	−0.09	[−.34, .19]	−0.10	[−.38, .27]	−0.08	[−.33, .22]	−0.05	[−.32, .23]	−0.04	[−.22, .28]	0.00	[−.26, .24]
d _CoopNTW	0.02	[−.25, .26]	0.05	[−.28, .35	0.02	[−.26, .29]	0.05	[−.324 .32]	0.03	[−.28, .22]	0.04	[−.21, .28]
d _CoopTW	0.04	[−.22, .29]	0.04	[−.27, .33]	0.06	[−.23, .31]	0.01	[−.27, .28]	0.04	[−.20, .24]	0.00	[−.30, .29]
g _NTW	−0.10	[−.26, .06]	−0.35	[−.48, −.18]*	−0.06	[−.23, .11]	−0.26	[−.41, −.10]*	−0.11	[−.26, .05]	−0.25	[−.39, −.10]*
g _TW	−0.17	[−.35, .07]	−0.17	[−.38, .10]	−0.23	[−.42, .05]	−0.07	[−.287, .17]	−0.17	[−.35, .06]	−0.01	[−.21, .19]
b _NTW	0.14	[−.04, .32]	0.11	[−.10, .32]	0.06	[−.14, .26]	0.16	[−.04, .34]	0.15	[−.04, .32]	0.15	[−.04, .32]
b _TW	0.15	[−.08, .34]	0.07	[−.18, .30]	0.09	[−.13, .30]	0.18	[−.05, .38]	0.13	[−.09, .32]	0.08	[−.14, .28]

BCI: Bayesian Credibility Intervals.

The values in brackets correspond to the Bayesian Credibility Intervals.

Signifies that the credibility interval excluded zero.

Discussion

In line with the pilot study and Experiment 1, participants invested more money when playing against trustworthy looking formerly learned cooperators and invested the least when playing against untrustworthy looking formerly learned cheaters. The source-memory data showed a difference in memory between untrustworthy looking cheaters and untrustworthy looking cooperators: Participants remembered more untrustworthy looking cheaters. There was no such difference given trustworthy old opponents. Furthermore, facial trustworthiness again strongly biased guessing in the source-memory test replicating other studies (Bell et al., 2012).

Consistent with Experiment 1, there was a reliable relationship between source guessing, a priori facial trustworthiness, and decisions: When opponents looked untrustworthy, participants guessed that this face belonged to a cheater and cooperated less with this person. However, again, no correlation between source memory and later behaviour could be found. In line with Schaper et al. (2019), we found a relationship between old-new recognition and cooperation in the cooperator condition. The better old–new recognition was, the more our participants cooperated with previously encountered cooperative opponents. As Schaper et al. (2019) already pointed out, that the recognition of a face seems to be useful to approach cooperators, but not to avoid cheaters.

General discussion

Several studies were able to show that source memory was better for cheaters than other types of behaviour (e.g., Barclay, 2008; Bell & Buchner, 2010; Bell et al., 2010; Buchner et al., 2009; Kroneisen, 2018; Kroneisen & Bell, 2013; Suzuki & Suga, 2010). This has led to the hypothesis of a cheater detection module that is a highly specific cognitive mechanism assumed to support social exchange by facilitating the detection of cheaters. The results of more recent studies, however, showed a more general influence of social expectations on memory (e.g., Bell et al., 2012; Kroneisen & Bell, 2013; Kroneisen et al., 2015). Cheating persons were not per se remembered better than cooperative persons, but source memory for expectancy-incongruent information was superior in comparison to expectancy-congruent information. These findings are inconsistent with the assumption of a specialised cognitive module. A mechanism, however, that is sensitive to expectancy violations, independent of the expectations’ origin (e.g., stereotypes, social rules, facial appearance), may be an efficient way of optimal information processing in social exchange situations (Bell et al., 2012).

Beside the effects on memory, moral beliefs about persons also influence people`s decision making (see, for example, Delgado et al., 2005). First impressions, reputational information about a person, and personal experience influence economic decision making (Chang et al., 2010; Delgado et al., 2005; Frank et al., 1993). It can be concluded that social cooperation depends fundamentally on expectations about other people’s behaviours (Mieth et al., 2016).

But how are these expectations about the trustworthiness of a person formed? Several studies have shown that (1) initial impressions and (2) previous interactions influence the amount of trust people place in a partner (Chang et al., 2010). Trustworthiness, for example, is often rapidly inferred from several first impressions. People can detect very subtle signals of trustworthiness by simply viewing faces (Winston et al., 2002). Facial trustworthiness as an implicit social signal also influences initial judgements of trustworthiness and therefore economic decision making. van’t Wout and Sanfey (2008), for example, showed that judgements of facial trustworthiness predict the financial risk someone is willing to take. Combining these different strands of research, in one pilot study and two main experiments, we examined the interaction of appearance–based expectations and memory on economic decisions. More precisely, we were interested in (1) how much the behaviour in a two-player Public Goods game depends on memory for previous experiences with the opponents and (2) what influence moral expectations like facial trustworthiness have on both, source memory and economic decisions.

Interestingly, the source-memory results from both main experiments did not replicate the often-found source-memory advantage for unexpected behaviour. Our data showed no differences in source memory between cheaters and cooperators (independently of facial trustworthiness) in Experiment 1, and only a source-memory advantage for untrustworthy looking cheaters in comparison to cooperators in Experiment 2, challenging previous results. However, when looking at other studies, these findings may not be completely inconsistent with the literature. So far, a clear source-memory advantage for cheaters was found when using third-hand information in the learning phase. This is usually done by presenting participants with pictures of faces paired with short statements describing the pictured persons as cheating, cooperative, or neutral. Other researcher used game-theoretical paradigms in the learning phase (Bell et al., 2010, 2016; Giang et al., 2012; Schaper et al., 2019). When using this paradigm, no reliable evidence for a source memory advantage of cheaters over cooperators could be found. One possible explanation for this effect is that people focus on the information that is relevant in a given situation. When giving third-hand information about cooperative behaviour, there is no need to reciprocate it. This changes when playing a “game” with different partners. Here, a cooperative opponent must be rewarded. Furthermore, an explicit interaction with a person suggests that future encounters are likely. Therefore, it is important to remember the cooperators as well as the cheaters to give an appropriate reaction to every opponent (Bell et al., 2010, 2016; Schaper et al., 2019). Furthermore, also the prioritisation of the processing of unexpected information and its association to better source memory was often found in third-hand information paradigms (Bell et al., 2012, 2015; Kroneisen & Bell, 2013; Kroneisen et al., 2015). But only a few experiments used game-theoretical paradigms to test this (Bell et al., 2012). There is the possibility that a similar argumentation also holds for the lack of the incongruity effect in our experiments: When using a paradigm that encourages a direct involvement of the participants, it is important to remember cooperators as well as cheaters. Furthermore, in such situations other cues that influence moral expectations about other people, like facial trustworthiness, may become increasingly important for the participants. Apparently, behavioural relevance determines the processing of social information. Kroneisen (2018), for example, found that behavioural relevance impacted source memory resulting in better source memory for cheaters who showed a behaviour with high probability of occurrence for a student population (e.g., “F.A. secures a spot in the library early in the morning during the busy exam prep period, but then shows up three hours later to study.”) in contrast to cheaters who showed a behaviour with only low behavioural relevance for a student population (e.g., “P.O. fulfils his military service. Because he has free access to the armoury, he steals ammunition and sells it on the black market.”). In the current experiments, explicit interaction and the need to remember cooperators as well as cheaters might have triggered participants to use other cues and to overuse facial trustworthiness as an important diagnostic cue for later behaviour, increasing its behavioural relevance.

Interestingly, in line with Schaper et al. (2019), we found a clear positive relationship between old-new recognition and cooperation with previously encountered cooperative partners in Experiment 2. Participants cooperated more often with formerly learned cooperators, the better old-new recognition was. Schaper et al. (2019) already pointed out that face recognition seemed to be useful to approach cooperators only. No negative relationship between old–new recognition and cooperation with previously encountered cheating partners could be found.

In contrast to Schaper et al. (2019), we did not find a relationship between source memory and the decision to cooperate. Instead, we found that the correlation between cooperation and the classifications in the memory test was based on strategic guessing, but only when the opponent looked untrustworthy. The more participants guessed that an untrustworthy looking face belonged to a cheater, the less they cooperated with untrustworthy looking opponents. However, there are some differences between the study reported by Schaper and colleagues and our studies. First, instead of two learning rounds, we only had one. Second, we used pictures from trustworthy and untrustworthy looking opponents. Why are these differences important? Previous research showed that expectations about the trustworthiness of other people are often formed automatically on the basis of appearance, for example, based on facial cues (Todorov et al., 2009, 2015). The results from our decision tasks indicate that both, facial trustworthiness as well as experienced behaviour of the opponent can influence later actions. Similar to other studies (Mieth et al., 2016; van’t Wout & Sanfey, 2008), participants invested less money in the two-player Public Goods game when their opponent looked untrustworthy than when he looked trustworthy. Contrary to our expectations, our participants invested more when playing against trustworthy looking formerly learned cooperators and least when playing against untrustworthy looking formerly learned cheaters. A finding that was consistently found in all our experiments. This confirms that facial cues have a strong effect on trust, social expectations, and thereby decision making. In line with this notion, the effect of facial trustworthiness was strongest when playing against new partners. This is not a completely unreasonable behaviour: As Sparks et al. (2016) showed, people can predict each other’s Prisoner’s Dilemma decisions after talking with later opponents for only a short time. Our experiments extend these results. Facial trustworthiness seems to be a very important factor for investment behaviour. Our results indicate that this cue can even overwrite previous encounters with a person. Given facial trustworthiness as cue is available, people seem to overweight it in their investment decisions. This is also consistent with our finding that source guessing is associated with investment behaviour.

However, memory of the formerly shown behaviour can also influence decisions. Knowledge about one’s opponent reduces the influence of facial trustworthiness. Chang et al. (2010) showed that trustworthiness and experience interact: trustworthy looking opponents who also showed trustworthy behaviour were entrusted with the most money. However, contrary to our experiments, in the experiment from Chang et al. participants played repeatedly with their opponents and in Schaper et al. (2019) there were two learning rounds. Thus, the more people know about their opponents, the less influence their facial appearance has on investment strategies.

To summarise, our present experiments extend the current research on memory and guessing on adaptive decision making. We were able to demonstrate that besides previous interactions and the memory for these interactions, there are other factors influencing cooperation. Facial trustworthiness seems to be a very important diagnostic cue for people to decide whether they can trust someone or not. This cue is so important that it may even overwrite previous experiences with the same person. However, one caveat in this research field remained also in our study. Psychological studies often ignore the possibility of within-person variations in appearance (Jenkins et al., 2011). We used the same image of a person in the learning phase as well as in the memory test and the decision task. From an ecological validity perspective, this is an unrealistic choice. Different pictures from a person can vary substantially and it is unclear if the recognition of one image of a person can be seen as proof that the identity of this person was remembered (Jenkins et al., 2011). Therefore, future research should try to increase the ecological validity of their experiments.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) received no financial support for the research, authorship, and/or publication of this article.

ORCID iD

Meike Kroneisen

Notes

References

Axelrod

(1984). The evolution of cooperation. Basic Books.

Axelrod

Hamilton

(1981). The evolution of cooperation. Science, 211, 1390–1396.

Ayal

Klar

(2014). Detecting varieties of cheating: How do people navigate between different cheating ploys? Thinking & Reasoning, 20, 51–76.

Barclay

(2008). Enhanced recognition of defectors depends on their rarity. Cognition, 107, 817–828.

Bayen

U. J.

Murnane

(1996). Aging and the use of perceptual and temporal information in source memory tasks. Psychology and Aging, 11, 293–303. https://doi.org/10.1037/0882-7974.11.2.293

Bayen

U. J.

Murnane

Erdfelder

(1996). Source discrimination, item detection, and multinomial models of source monitoring. Journal of Experimental Psychology: Learning, Memory, and Cognition, 22, 197–215. https://doi.org/10.1037/0278-7393.22.1.197

Bell

Buchner

(2010). Valence modulates source memory for faces. Memory & Cognition, 38, 29–41.

Bell

Buchner

Kroneisen

Giang

(2012). On the flexibility of social source memory: A test of the emotional incongruity hypothesis. Journal of Experimental Psychology: Learning, Memory, and Cognition, 38, 1512–1529. https://doi.org/10.1037/a0028219

Bell

Buchner

Musch

(2010). Enhanced old-new recognition and source memory for faces of cooperators and defectors in a social-dilemma game. Cognition, 117, 261–275.

10.

Bell

Giang

Mund

Buchner

(2013). Memory for reputational trait information: Is social-emotional information processing less flexible in old age? Psychology and Aging, 28, 984–995. https://doi.org/10.1037/a0034266

11.

Bell

Mieth

Buchner

(2015). Appearance-based first impressions and person memory. Journal of Experimental Psychology: Learning, Memory, and Cognition, 41, 456–472. https://doi.org/10.1037/xlm0000034

12.

Bell

Sasse

Möller

Czernochowski

Mayr

Buchner

(2016). Event-related potentials in response to cheating and cooperation in a social dilemma game. Psychophysiology, 53, 216–228. https://doi.org/10.1111/psyp.12561

13.

Buchner

Bell

Mehl

Musch

(2009). No enhanced recognition memory, but better source memory for faces of cheaters. Evolution and Human Behavior, 30, 212–224. https://doi.org/10.1016/j.evolhumbehav.2009.01.004

14.

Bröder

Meiser

(2007). Measuring source memory. Zeitschrift für Psychologie/Journal of Psychology, 215, 52–60. https://doi.org/10.1027/0044-3409.215.1.52

15.

Carlisle

Shafir

(2005). Questioning the cheater-detection hypothesis: New studies with the selection task. Thinking and Reasoning, 11, 97–122. https://doi.org/10.1080/13546780442000079

16.

Chang

L. J.

Doll

van’t Wout

Frank

Sanfey

A. G.

(2010). Seeing is believing: Trustworthiness as a dynamic belief. Cognitive Psychology, 61, 87–105. https://doi.org/10.1016/j.cogpsych.2010.03.001

17.

Cosmides

(1989). The logic of social exchange: Has natural selection shaped how humans reason? Studies with the Wason Selection Task. Cognition, 31, 187–276.

18.

Delgado

M. R.

Frank

R. H.

Phelps

E. A.

(2005). Perceptions of moral character modulate the neural systems of reward during the trust game. Nature Neuroscience, 8, 1611–1618. https://doi.org/10.1038/nn1575

19.

Engell

A. D.

Haxby

J. V.

Todorov

(2007). Implicit trustworthiness decisions: Automatic coding of face properties in the human amygdala. Journal of Cognitive Neuroscience, 19, 1508–1519. https://doi.org/10.1162/jocn.2007.19.9.1508

20.

Erdfelder

Auer

T.-S.

Hilbig

B. E.

Aßfalg

Moshagen

Nadarevic

(2009). Multinomial processing tree models. A review of the literature. Zeitschrift für Psychologie/Journal of Psychology, 217, 108–124. https://doi.org/10.1027/0044-3409.217.3.108

21.

Faul

Erdfelder

Lang

A. G.

Buchner

(2007). G*Power 3: A flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behavior Research Methods, 39, 175–191.

22.

Fehr

Fischbacher

(2004). Social norms and human cooperation. Trends in Cognitive Sciences, 8, 185–190.

23.

Frank

R. H.

Gilovich

Regan

D. T.

(1993). The evolution of one-shot cooperation: An experiment. Ethology and Sociobiology, 4, 247–256.

24.

Giang

Bell

Buchner

(2012). Does facial resemblance enhance cooperation? PLOS ONE, 7, Article e47809. https://doi.org/10.1371/journal.pone.0047809

25.

Gigerenzer

Hug

(1992). Domain-specific reasoning: Social contracts, cheating, and perspective change. Cognition, 43, 127–171.

26.

Heck

D. W.

Arnold

N. R.

Arnold

(2018). TreeBUGS: An R package for hierarchical multinomial-processing-tree modeling. Behavior Research Methods, 50, 264–284. https://doi.org/10.3758/s13428-017-0869-7

27.

JASP Team. (2020). JASP (Version 0.12.2.0) [Computer software]. https://jasp-stats.org/

28.

Jenkins

White

Van Montfort

Mike Burton

(2011). Variability in photos of the same face. Cognition, 121, 313–323. https://doi.org/10.1016/j.cognition.2011.08.001

29.

Klauer

K. C.

(2010). Hierarchical multinomial processing tree models: A latent-trait approach. Psychometrika, 75, 70–98. https://doi.org/10.1007/s11336-009-9141-0

30.

Kroneisen

(2018). Is he important to me? Source memory advantage for personally relevant cheaters. Psychonomic Bulletin & Review, 25, 1129–1137. https://doi.org/10.3758/s13423-017-1345-1

31.

Kroneisen

Bell

(2013). Sex, cheating, and disgust: Enhanced source memory for trait information that violates gender stereotypes. Memory, 21, 167–181. https://doi.org/10.1080/09658211.2012.713971

32.

Kroneisen

Bell

(2018). Remembering the place with the tiger: Survival processing can enhance source memory. Psychonomic Bulletin & Review, 25, 667–673. https://doi.org/10.3758/s13423-018-1431-z

33.

Kroneisen

Heck

D. W.

(2020). Interindividual differences in the sensitivity for consequences, moral norms, and preferences for inaction: Relating basic personality traits to the CNI model. Personality and Social Psychology Bulletin, 46(7), 1013–1026. https://doi.org/10.1177/0146167219893994

34.

Kroneisen

Steghaus

(2021). The influence of decision time on sensitivity for consequences, moral norms, and preferences for inaction: Time, moral judgments, and the CNI model. Journal of Behavioral Decision Making, 34(1), 140–153. https://doi.org/10.1002/bdm.2202

35.

Kroneisen

Woehe

Rausch

L. S.

(2015). Expectancy effects in source memory: How moving to a bad neighborhood can change your memory. Psychonomic Bulletin & Review, 22, 179–189. https://doi.org/10.3758/s13423-014-0655-9

36.

Marevic

Arnold

N. R.

Rummel

(2018). Item-method directed forgetting and working memory capacity: A hierarchical multinomial modeling approach. Quarterly Journal of Experimental Psychology, 71, 1070–1080. https://doi.org/10.1080/17470218.2017.1310270

37.

Matzke

Dolan

C. V.

Batchelder

W. H.

Wagenmakers

E.-J.

(2015). Bayesian estimation of multinomial processing tree models with heterogeneity in participants and items. Psychometrika, 80, 205–235. https://doi.org/10.1007/s11336-013-9374-9

38.

Mieth

Bell

Buchner

(2016). Cognitive load does not affect the behavioral and cognitive foundations of social cooperation. Frontiers in Psychology, 7, Article 1312. https://doi.org/10.3389/fpsyg.201601312

39.

Minear

Park

D. C.

(2004). A lifespan database of adult facial stimuli. Behavior Research Methods, Instruments, & Computers, 36, 630–633.

40.

Moshagen

(2010). MultiTree: A computer program for the analysis of multinomial processing tree models. Behavior Research Methods, 42, 42–54. https://doi.org/10.3758/BRM.42.1.42

41.

Murty

V. P.

FeldmanHall

Hunter

L. E.

Phelps

E. A.

Davachi

(2016). Episodic memories predict adaptive value-based decision-making. Journal of Experimental Psychology: General, 145, 548–558. https://doi.org/10.1037/xge0000158

42.

Nairne

Pandeirada

(2016). Adaptive memory: The evolutionary significance of survival processing. Perspectives on Psychological Science, 11, 496–511. https://doi.org/10.1177/1745691616635613

43.

Phillips

P. J.

Wechsler

Huang

Rauss

P. J.

(1998). The FERET database and evaluation procedure for face-recognition algorithms. Image and Vision Computing, 16, 295–306. https://doi.org/10.1016/S0262-8856(97)00070-X

44.

Plummer

(2003). JAGS: A program for analysis of Bayesian graphical models using Gibbs sampling. In Hornik

Leisch

Zeileis

(Eds.), Proceedings of the 3rd international workshop on distributed statistical computing (Vol. 124, p. 125). Technische Universität Wien.

45.

Rezlescu

Duchaine

Olivola

Chater

Rustichini

(2012). Unfakeable facial configurations affect strategic choices in trust games with or without information about past behavior. PLOS ONE, 7, Article e34293. https://doi.org/10.1371/journal.pone.0034293

46.

Schaper

M. L.

Mieth

Bell

(2019). Adaptive memory: Source memory is positively associated with adaptive social decision making. Cognition, 186, 7–14. https://doi.org/10.1016/j.cognition.2019.01.014

47.

Snodgrass

J. G.

Corwin

(1988). Pragmatics of measuring recognition memory: Applications to dementia and amnesia. Journal of Experimental Psychology: General, 117, 34–50. https://doi.org/10.1037/0096-3445.117.1.34

48.

Sparks

Burleigh

Barclay

(2016). We can see inside: Accurate predictions of Prisoner’s Dilemma decisions after a brief face-to-face interaction. Evolution and Human Behavior, 37, 210–216.

49.

Stahl

Wiese

Schweinberger

S. R.

(2008). Expertise and own-race bias in face processing: An event-related potential study. NeuroReport, 19, 583–587.

50.

Suzuki

Honma

Suga

(2013). Indelible distrust: Memory bias toward cheaters revealed as high persistence against extinction. Journal of Experimental Psychology: Learning, Memory, and Cognition, 39, 1901–1913. https://doi.org/10.1037/a0033335

51.

Suzuki

Suga

(2010). Enhanced memory for the wolf in sheep’s clothing: Facial trustworthiness modulates face-trait associative memory. Cognition, 117, 224–229.

52.

Todorov

(2008). Evaluating faces on trustworthiness: An extension of systems for recognition of emotions signaling approach/avoidance behaviors. Annals of the New York Academy of Sciences, 1124, 208–224. https://doi.org/10.1196/annals.1440.012

53.

Todorov

Olivola

C. Y.

Dotsch

Mende-Siedlecki

(2015). Social attributions from faces: Determinants, consequences, accuracy, and functional significance. Annual Review of Psychology, 66, 519–545. https://doi.org/10.1146/annurev-psych-113011-143831

54.

Todorov

Pakrashi

Oosterhof

N. N.

(2009). Evaluating faces on trustworthiness after minimal time exposure. Social Cognition, 77, 813–833. https://doi.org/10.1521/soco.2009.27.6.813

55.

Tooby

Cosmides

(1992). The psychological foundations of culture. In Barkow

Cosmides

Tooby

(Eds.), The adapted mind: Evolutionary psychology and the generation of culture (pp. 19–136). Oxford University Press.

56.

Tooby

Cosmides

(2005). Conceptual foundations of evolutionary psychology. In Buss

D. M.

(Ed.), The handbook of evolutionary psychology (pp. 5–67). John Wiley.

57.

Trivers

(1971). The evolution of reciprocal altruism. Quarterly Review of Biology, 46, 35–57.

58.

van’t Wout

Sanfey

A. G.

(2008). Friend or foe: The effect of implicit trustworthiness judgments in social decision-making. Cognition, 108, 796–803. https://doi.org/10.1016/j.cognition.2008.07.002

59.

Wason

(1966). Reasoning. In Foss

(Ed.), New horizons in psychology (pp. 135–151). Penguin Books.

60.

Wiese

Schweinberger

S. R.

Hansen

(2008). The age of the beholder: ERP evidence of an own-age bias in face memory. Neuropsychologia, 46, 2973–2985.

61.

Wiese

Schweinberger

S.-R.

Neumann

M. F.

(2008). Perceiving age and gender in unfamiliar faces: Brain potential evidence for implicit and explicit person categorization. Psychophysiology, 45, 957–969.

62.

Willis

Todorov

(2006). First impressions: Making up your mind after a 100-ms exposure to a face. Psychological Science, 17, 592–598. https://doi.org/10.1111/j.1467-9280.2006.01750.x

63.

Winston

J. S.

Strange

B. A.

O’Doherty

Dolan

R. J.

(2002). Automatic and intentional brain responses during evaluation of trustworthiness of faces. Nature Neuroscience, 5, 277–283.

Remembering the bad ones: Does the source memory advantage for cheaters influence our later actions positively?

Abstract

Keywords

(Source) memory for cheaters

Trustworthiness as cue in economic decisions

Facial appearance as indicator of trustworthiness

Interaction effects of memory and (facial) trustworthiness on economic decisions

Pilot study

Participants

Apparatus and materials

Procedure

Debt game—learning phase

Public goods game—testing phase

Design

Results

Likeability ratings

Investments in the debt game (learning phase)

Public goods game investments 3 (testing phase)

Discussion

Experiment 1

Method

Participants

Apparatus and materials

Procedure

Public goods game—learning phase

Decision task—testing phase

Memory test

Design

Results

Investments in the learning phase

Investments in the testing phase

Old-new recognition

Source memory and behaviour

Discussion

Experiment 2

Method

Participants

Results

Investments in the learning phase

Investments in the testing phase

Old–new recognition

Source memory and behaviour

Discussion

General discussion

Footnotes

Declaration of conflicting interests

Funding

ORCID iD

Notes

References

Public goods game investments³ (testing phase)