LETTER doi:10.3.038/nature11467

LETTER doi:10.3.038/nature11467 Spontaneous giving and calculated greed David G. Randt•2-', Joshua D. Greene2+ & Martin A. Nowakm•ss Cooperation is central to human social behaviour". However, choosing to cooperate requires individuals to incur a personal cost to benefit others. Here we explore the cognitive basis of cooperative decision-making in humans using a dual-process framework" -". We ask whether people are predisposed towards selfishness, behav- ing cooperatively only through active self-control; or whether they are intuitively cooperative, with reflection and prospective reason- ing favouring 'rational' self-Interest. To investigate this issue, we perform ten studies using economic games. We find that across a range of experimental designs, subjects who reach their decisions more quiddy are more cooperative. Furthermore, forcing subjects to decide quickly increases contributions, whereas instructing them to reflect and forcing them to decide slowly decreases con- tributions. Finally, an induction that primes subjects to trust their intuitions increases contributions compared with an induction that promotes greater reflection. To explain these results, we propose that cooperation is intuitive because cooperative heuristics are developed in daily life where cooperation is typically advantageous. We then validate predictions generated by this proposed mechanism. Our results provide convergent evidence that intuition supports coopera- tion in social dilemmas, and that reflection can undermine these cooperative Impulses. Many people are willing to make sacrifices for the common good". Here we explore the cognitive mechanisms underlying this cooperative behaviour. We use a dual-process framework in which intuition and reflection interact to produce decisions'° -".". Intuition is often associated with parallel processing, automaticity, effortlessness, lack of insight into the decision process and emotional influence. Reflection is often associated with serial processing, effortfulness and the rejection of emotional influence" -1".". In addition, one of the psychological features most widely used to distinguish intuition from reflection is processing speed: intuitive responses are relatively fast, whereas reflective responses require additional time for deliberation". Here we focus our attention on this particular dimension, which is closely related to the distinction between automatic and controlled processing". Viewing cooperation from a dual-process perspective raises the following questions are we intuitively self-interested, and is it only through reflection that we reject our selfish impulses and force ourselves to cooperate? Or are we intuitively cooperative, with reflection upon the logic of self-interest causing us to rein in our cooperative urges and instead act selfishly? Or, alternatively, is there no cognitive conflict between intuition and reflection? Here we address these questions using economic cooperation games. We begin by examining subjects' decision times. The hypothesis that self-interest is intuitive, with prosociality requiring reflection to override one's selfish impulses, predicts that faster decisions will be less cooperative. Conversely, the hypothesis that intuition preferentially supports prosocial behaviour, whereas reflection leads to increased selfishness, predicts that faster decisions will be more cooperative. As a first test of these competing hypotheses, we conducted a one- shot public goods game" (PGG) with groups of four participants. I 8 0 We recruited 212 subjects from around the world using the online labour market Amazon Mechanical Turk (AMT)". AMT provides a reliable subject pool that is more diverse than a typical sample of college undergraduates (see Supplementary Information, section 1). In accordance with standard AMT wages, each subject was given USS0.40 and was asked to choose how much to contribute to a common pool. Any money contributed was doubled and split evenly among the four group members (see Supplementary Information, section 3, for experimental details). Figure la shows the fraction of the endowment contributed in the slower half of decisions compared to the faster half. Faster decisions result in substantially higher contributions compared with slower decisions (rank sum test, P= 0.007). Furthermore, as shown in Fig. lb, we see a consistent decrease in contribution amount with a 75% 65% - 55% - 45% - 35% b 100% 80% 60% 40% 20% 0% I I Faster decisions Sower decisions 1-10s >10s 4 56 12 I6 4 02 0.6 1 1.4 1.8 2.2 Decision tine (loa,,N) Figure 1 I Faster decisions are more cooperative Subjects who reach their decisions more quickly contribute more in a one-shot PCG (n = 212). This suggests that the intuitive response is to be cooperative. a, Using a median split on decision time, we compare the contribution levels of the faster half versus slower half of decisions. The average contribution is substantially higher for the faster decisions. b, Plotting contribution as a function of 168w- transformed decision time shows a negative relationship between decision time and contribution. Dot size is proportional to the number of observations, listed next to each dot. Error bars, mean = s.e.m. (see Supplementary Information, sections 2 and 3, for statistical analysis and further details). 'Program kw Evolutionary Dynarnics.Karrard Univers4y.Cambridge.Massachusells02138.USA. 2Deperlma4 ol Psycholori.HanrardUniversty.Combdidge.Abssachuselts02138.USA. 30eparlmentor Psychdogy.YaleUroversily.NewHaren.Connecleul 0652Q USA 'Deportment ol lAathanalics. Harvard LInnerstly.Carribridge.Massachuselts02138.USODepartnenl ol Organismic and Evolutionary Bool3gy. Harvard Unwersrty.Cambndge.i.lassechusetts0213a USA •Theseauthen ccobibuled equally to the work SEPTEMBER 2012 I VOL 489 I NATURE I 4 2 7 02012 Macmillan Publishers Limited. All rights reserved EFTA01146514 RESEARCH LETTER increasing decision time (Tobit regression, coefficient = -15.84, P=0.019; see Supplementary Information, sections 2 and 3, for statistical details). These findings suggest that intuitive responses are more cooperative. Next we examined data from all of our previously published social dilemma experiments for which decision time data were recorder -22. In these studies, conducted in the physical laboratory with college students, the experimental software automatically recorded decision times, but these data had not been previously analysed. To examine the psychology that subjects bring with them into the laboratory, we focused on play in the first round of each experimental session. In a one-shot prisoner's dilemma (ri = 48)20, a repeated prisoner's dilemma with execution errors (it = 278)31, a repeated prisoner's dilemma with and without costly punishment (pi = 104)", and a repeated PGG with and without reward and/or punishment (n = 192)7, we find the same negative relationship between decision time and cooperation (see Supplementary Information, section 4, for details). These results show the robustness of our decision-time findings: across a range of experi- mental designs, and with students in the physical laboratory as well as with an international online sample, faster decisions are associated with more prosociality. We now demonstrate the causal link between intuition and coop- eration suggested by these correlational studies. To do so, we recruited another 680 subjects on AMT and experimentally manipulated their decision times in the same one-shot PGG used above. In the 'time pressure' condition, subjects were forced to reach their decision quickly (within 10 s). Subjects in this condition have less time to reflect than in a standard PGG, and therefore their decisions are expected to be more intuitive. In the 'time delay' condition, subjects were instructed to carefully consider their decision and forced to wait for at least lOs before choosing a contribution amount. Thus, in this condition, decisions are expected to be driven more by reflection (see Supplementary Information, section 5, for experimental details). The results (Fig. 2a) are consistent with the correlational observa- tions in Fig 1. Subjects in the time-pressure condition contribute sig- nificantly more money on average than subjects in the time-delay condition (rank sum, P< 0.001). Moreover, we find that both manip- ulation conditions differ from the average behaviour in the baseline experiment in Fig. 1, and in the expected directions: subjects under time-pressure contribute more than unconstrained subjects (rank sum, P= 0.058), whereas subjects who are instructed to reflect and delay their decision contribute less than unconstrained subjects (rank sum, P = 0.028), although the former difference is only marginally significant. See Supplementary Information, section 5, for regression analyses. Additionally, we recruited 211 Boston-area college students and replicated our time-constraint experiment in the physical laboratory with tenfold higher stakes (Fig. 2b). We find again that subjects in the time-pressure condition contribute significantly more money than subjects in the time-delay condition (rank sum, P= 0.032). We also assessed subjects' expectations about the behaviour of others in their group, and find no significant difference across conditions (rank sum, P= 0.360). Thus, subjects forced to respond more intuitively seem to have more prosocial preferences, rather than simply contributing more because they are more optimistic about the behaviour of others (see Supplementary Information, section 6, for experimental details and analysis). We next used a conceptual priming manipulation that explicitly invokes intuition and reflection'. We recruited 343 subjects on AMT to participate in a one-shot PGG experiment. The first condition promotes intuition relative to reflection: before reading the PGG instructions, subjects were assigned to write a paragraph about a situ- ation in which either their intuition had led them in the right direction, or careful reasoning had led them in the wrong direction. Conversely, the second condition promotes reflection: subjects were asked to write about either a situation in which intuition had led them in the wrong S O a 75% 65% 55% S 45% 35% b 75% 65%- 55% - S O 45% 35% O 75% 65% - 55% - Time pressure 45% - 35% I I I Unconstrained Time delay Constraint condition Contribution Prediction of others' contribution Time pressure Time delay Constraint condition Promote intuition or Promote reflection or inhibit reflection inhibit intuition Priming conckhon Figure 2 I Inducing intuitive thinking promotes cooperation. a. Forcing subjects to decide quickly (10 s or less) results in higher contributions, whereas forcing subjects to decide slowly (more than 10 s) decreases contributions (n = 680). This demonstrates the causal link between decision time and cooperation suggested by the correlation shown in Fig. 1. b, We replicate the finding that forcing subjects to decide quickly promotes cooperation in a second study run in the physical laboratory with tenfold larger stakes = 211). We also find that the time constraint has no significant effect on subjects' predictions concerning the average contributions of other group members. Thus, the manipulation acts through preferences rather than beliefs. c. Priming intuition (or inhibiting reflection) increases cooperation relative to priming reflection (or inhibiting intuition) (n = 343). 'This finding prosides further evidence for the specific role of intuition versus reflection in motivating cooperation. as suggested by the decision time studies. Error bars, mean ± s.c.m. (see Supplementary Information, sections 5-7, for statistical analysis and further details). direction, or careful reasoning had led them in the right direction. Consistent with the seven experiments described above, we find that contributions are significantly higher when subjects are primed to promote intuition relative to reflection (Fig. 2c; rank sum, P= 0.011; see Supplementary Information, section 8, for experimental details and analysis). These results therefore raise the question of why people are intuitively predisposed towards cooperation. We propose the follow- ing mechanism: people develop their intuitions in the context of daily life, where cooperation is typically advantageous because many important interactions are repeated"" 22, reputation is often at 428 1 NATURE 1 VOL 489 20 SEPTEMBER 201: 02012 Macmillan Publishers Limited. All rights reserved EFTA01146515 LETTER RESEARCH stake's" and sanctions for good or bad behaviour might exist9b Thus, our subjects develop cooperative intuitions for social interactions and bring these cooperative intuitions with them into the laboratory. As a result, their automatic first response is to be cooperative. It then requires reflection to overcome this cooperative impulse and instead adapt to the unusual situation created in these experiments, in which cooperation is not advantageous. This hypothesis makes clear predictions about individual difference moderators of the effect of intuition on cooperation, two of which we now test. First, if the effects described above result from intuitions formed through ordinary experience, then greater familiarity with laboratory cooperation experiments should attenuate these effects. We test this prediction on AMT with a replication of our conceptual priming experiment. As predicted, we fmd a significant interaction between prime and experience it is only among subjects naive to the experimental task that promoting intuition increases cooperation (Fig. 3a; see Supplementary Information, section 9, for experimental details and statistical analysis). This mechanism also predicts that subjects will only fmd coopera- tion intuitive if they developed their intuitions in daily-life settings in which cooperation was advantageous. Even in the presence of repe- tition, reputation and sanctions, cooperation will only be favoured if enough other people are similarly cooperative". We tested this pre- diction on AMT with a replication of our baseline correlational study. As predicted, it is only among subjects that report having mainly cooperative daily-life interaction partners that faster decisions are a 75% b I 65% 55% 45% 35% 75%. 65% - 55%. 45% - 35% Naive ■ Primed to promote intuition • Primed to promote reflection I _I_ Experienced Previous experience with experimental setting ■ Faster dectsicns ■ Slower decisions ■Cooperative Uncooperative Opinion of daily-life interaction partners Figure 3 I Evidence that cooperative intuitions from daily lift spill over into the laboratory. Two experiments validate predictions of our hypothesis that subjects develop their cooperative intuitions in the context ofdaily life, in which cooperation is advantageous. a, Priming that promotes reliance on intuition increases cooperation relative to priming promoting reflection,but only among naive subjects that report no previous experience with the experimental setting where cooperation is disadvantageous (ra = 256). b, Faster decisions arc associated with higher contribution levels, but only among subjects who report having cooperative daily-life interaction partners = 341). As in Fig. la, a median split is carried out on decision times, separating decisions into the faster versus slower half. Error bars, mean ± s.e.m. (see Supplementary Information• sections 9 and 10. for statistical analysis and further details). associated with higher contributions (Fig. 3b; see Supplementary Information, section 10, for experimental details and statistical analysis). Thus, there are some people for whom the intuitive response is more cooperative and the reflective response is less cooperative; and there are other people for whom both the intuitive and reflective responses lead to relatively little cooperation. But we find no cases in which the intuitive response is reliably less cooperative than the reflective res- ponse. As a result, on average, intuition promotes cooperation relative to reflection in our experiments. By showing that people do not have a single consistent set of social preferences, our results highlight the need for more cognitively com- plex economic and evolutionary models of cooperation, along the lines of recent models for non-social decision-making' ra'-16. Furthermore, our results suggest a special role for intuition in promoting coopera- tion". For further discussion, and a discussion of previous work exploring behaviour in economic games from a dual-process perspec- tive, see Supplementary Information, sections 12 and 13. On the basis of our results, it may be tempting to conclude that cooperation is 'innate' and genetically hardwired, rather than the product of cultural transmission. This is not necessarily the case: intuitive responses could also be shaped by cultural evolution" and social learning over the course of development. However, our results are consistent with work demonstrating spontaneous helping behaviour in young children". Exploring the role of intuition and reflection in cooperation among children, as well as cross-culturally, can shed further light on this issue. Here we have explored the cognitive underpinnings of cooperation in humans. Our results help to explain the origins of cooperative behaviour, and have implications for the design of institutions that aim to promote cooperation. Encouraging decision-makers to be maximally rational may have the unintended side-effect of making them more selfish. Furthermore, rational arguments about the import- ance of cooperating may paradoxically have a similar effect, whereas interventions targeting prosocial intuitions may be more successful 30. Exploring the implications of our findings, both for scientific under- standing and public policy, is an important direction for future study: although the cold logic of self-interest is seductive, our first impulse is to cooperate. METHODS SUMMARY Across studies 1.6, 8,9 and 10,a total of 1.955 subjects were recruited using AMR'" to participate in one of a series of variations on the one-shot PGG, played through an online survey website. Subjects received $030 for participating. and could earn up to SI more based on the PGG. In the PGG, subject were given S0A0 and chose how much to contribute to a 'common project'. All contributions were doubled and split equally among four group members. Once all subjects in the experiment had made their decisions, groups of four were randomly matched and the resulting payoffs were calculated. Each subject was then paid accordingly through the AMT payment system. and was informed about the average contribution of the other members of his or her group. No deception was used. In study 7, a total of 21I subjects were recruited from the Boston. Massachusetts. metropolitan area through the Harvard University Computer Laboratory for Experiment Research subject pool to participate in an experiment at the Harvard Decision Science Laboratory. Participation was restricted to students under 35 years of age. Subjects received a $5 show-up fee for arriving on time and had the opportunity to earn up to an additional S12 in the experiment. Subjects played a single one-shot PGG through the same website interface used in the AMT studies, but with tenfold larger stakes (maximum earnings of 510). Subjects were then asked to predict the average contribution of their other group numbers and had the chance to win up to an additional S2 based on their accuracy. These experiments were approved by the Harvard University Committee on the Use of Human Subjects in Research. For further details of the experimental meihods.set Supplementary Information. Received 13 December 2011; accepted 2 August 2012. 1. Trivers. R. The evolution of reciprocal altruism. Q. Rev. Biol. 46, 35-57 (1971) 12 SEPTEMBER 2012 I VOL 499 I NATURE 1 429 02012 Macmillan Publishers Limited. All rights reserved EFTA01146516 RESEARCH LETTER 2. Fudenberg.D.& Maskin. E. The folk theorem in repeated games with discounting or with incomplete information. Economefrica 54, 533-554 (1986} 3. Nowak M. A.& Sigmund. K Evolution of indirect reciprocity. Nature 437, 1291-1298(2005} 4. Boyd. R. Gintis. H.. Bowles. S.& Richerson. P. J. The evolution of altruistic punishment. Proc. Nat Acad. Sci. USA 100, 3531-3535 (2003). 5. Mil inski.M..Semrrenn. D.& KrambeckH.J. Reputation helps solve the 'tragedy of the commons. Nature 415.424-426 (2002). 6. Rockenbach. & Milinsld. M. The efficient interaction of indirect reciprocity and costly punishment. Nature 444, 718-723 (2006). 7. Rand, D.G.. Dreher, A.. Ellirgsen.T, Fudenberg. a& Nowak. M.A Positive interactions promote public cooperation. Science 325, 1272-1275 (2009). & Fehr, E& Gachter.S. Altruistic punishment in humans. Nature 415,137-140 (2002). 9. Rand, D. G..Arbesman. & & Christakis N. A. Dynamic social networks promote cooperation in experiments with humans. Proc Nay Aced Sci USA 108, 19193-19198 (2011). la Sloman, S.A. The empirkal case for two systems of reasonirg.Psychol Bull 119, 3-22 (1996). 11. Stanovich. K. E& West. R. F. Individual differences in rational thougbt. J. Exp. Psychal 127, 161-188(1998} 12 Chaiken.S.&Trope.Y. Dual-ProcessTheories in SocialPsycholog/(Gu Word, 1999} 13. Kahneman, D. A perspective on judgment and choice: mapping bounded rationality. Am. Psycho). 58, 697-720 (2003). 14. Plessner. Betsch. C. & Betsch. T. Intuition in Judgment and Decision Making (Lawrence Erlbaum. 2008} 15. Kahneman, D. Thinking Fast and Sfotv (Straus and Giroux. 2011} 16. Shiffrin.R_ M.& Schneider. W.Controlled and automatic information processing: II. Perceptual teaming, automatic attending, and a general theory. Psycho,. Rev. 84, 127-190(1977} 17. Milkw.E K&Ccben J. D.An integrative theory of prefrontal cortex function.Annu. Rev. Neumsci. 24,167-202 (2001). 18. Frederick S,Ccgnitive reflection and decision making.J.Econ.Parmect 19.25-42 (2005). 19. Horton. J. J., Rand, D.G. & Zeckhauser, R. J. The online laboratory: conducting experiments in a real labor market. Erp. Econ. 14, 399-425 (2011). 20. Pfeiffer. T.. Tran, L Krumme. C.& Rand. D. G. The value of reputation.J. R. Soc. Interface http://dxdoiorg/10.1098/rsif2012.0332 (20 June 2012). 21. Fudenberg. D.. Rand, a G.& Dreber.A Slow to anger and fast to forgive: cooperation in an uncertain world. Am. Econ. Rev. 102, 720-749 (2012} 22. Dreher. A, Rand, D.C.. Fudenberg.D.& Nowak M.A Winners don't punish. Nature 452, 348-351(2008). 23. Shenhay.A.. Rand. D. G. & Greene. J. D. Divine intuition: ccgnitive style influences belief in God. J. Exp. PsychoL Gen. 143.423-428 (2012). 24. Benhabib.J. & Bisin, A. Modeling internal commitment mechanisms and self- control: a neuroeconomics approach to consumption -saving decisions. Games Econ. BEAN. 52.460-492 (2005). 25. Fudenberg.D.& Levine. D. KA Dual-self model of impulse confrci.Am. Econ. Rev. 96, 1449-1476 (2C06} 2& McClure, S. M. Laibson. D. I, Loewenstein. G. & Cohen J. D. Separate neural systems value immediate and delayed monetary rewards. Science 306, 503-507 (2004). 27. Bowies S. & Gintis H. in The Economyas a EvolvingComplex System 3 (eds Blume, L and Durlaut S. N.) 339-364 (2002} 26 Richerson. P.J. & Boyd, R. Not byGenes Alone: How Culture Transformed Human Evolution. (Univ. Chicago Press 2005). 29. Wameken, F. & Tomasello. M. Altnistic helping in human infants and young chimpanzees. Science 311, 1301-1303 (2006). 30. Bowies S.Policies designed for seff-interested citizens mayundermine "the moral sentiments": evidence from economic experiments. Science 320, 1605-1609 f2003). Supplementary Information is available in the online version of the paper. Acknowledgements We thank H. Ahlblad, O. Amir. F. Fu. O. Hauser.). Horton and R Kane for assistance with carrying out the experiments. and P. Blake. S. BmvIes. N. Christakis. F. Cushman. A. Dreher. T. Ellingsen. F. Fu. D. Fudenberg. 0. Hauser. J. Jordan. M. Johannesson, M. Manapat J. Paxton. A Peysakhovkh. A Shenhay. J. Sirlin-Rand. M. van Veelen and 0. Wurzbacher for discussion and comments This work was supported in part bya National Science Foundation grant (SES-082197/3 to JD.G.). D.G.R and MAN. are supported by grants from the John Templeton Foundation. Author Contributions D.G.R..J.D.G. and MAN. designed the experiments D.G.R carried out the experiments and statistical analyses.ancl D.G.R.JD.G.and MAN. wrote the paper. Author Information Reprints and permissions information is available at bwnvnature.comireprints. The authors declare no competing financial interests Readers are welcome to comment on the online version of the paper. Correspondence and requests for materials should be addressed to D.G.R.([email protected]). 430 I NATURE I VOL 489 I 20 SEPTEMBER 2012 02012 Macmillan Publishers Limited. All nghts reserved EFTA01146517 SUPPLEMENTARY INFORMATION doi:10.1038/nature11467 1. Online recruitment procedure using Amazon Mechanical Turk 2 2. Log-transforming decision times 3 3. Study 1: Correlational decision time experiment on AMT 4 4. Studies 2 - 5: Reanalysis of previously published experiments run in the physical laboratory 6 5. Study 6: Time pressure / time delay experiment on AMT 12 6. Study 7: Time pressure / time delay experiment with belief elicitation in the physical laboratory 14 7. Behavior on AMT versus the physical laboratory (Study 6 vs Study 7) 17 8. Study 8: Conceptual priming experiment on AMT 18 9. Study 9: Conceptual priming experiment with experience measure and decision times on AMT 22 10. Study 10: Correlational experiment on AMT with moderators, individual differences in cognitive style, and additional controls 26 12. Implications for economic and evolutionary models 36 13. Previous dual-process research using economic games 37 14. Supplemental study: Experiment on AMT showing that detailed comprehension questions induce reflective thinking and reduce cooperation 38 15. Experimental instructions 40 References 47 WWW.NATURE-COM/NATUREI I EFTA01146518 doi:10.1038/nature11467 RESEARCH SUPPLEMENTARY INFORMATION 1. Online recruitment procedure using Amazon Mechanical Turk Subjects for many of the experiments in this paper were recruited using the online labor market Amazon Mechanical Turk (AMT)". AMT is an online labor market in which employers can employ workers to complete short tasks (generally less than 10 minutes) for relatively small amounts of money (generally less than $1). Workers receive a baseline payment and can be paid an additional bonus depending on their performance. This makes it easy to run incentivized experiments: the baseline payment is a `show-up fee,' and the bonus payment is determined by the points earned in the experiment. One major advantage of AMT is it allows experimenters to easily expand beyond the college student convenience samples typical of most economic game experiments. Among American subjects, AMT subjects have been shown to be significantly more nationally representative than college student sampled/. Furthermore, workers on AMT are from all around the world: in our experiments, 37% of the subjects lived outside of the United States, with more than half of the non-American subjects living in India. In our statistical analyses below, we show that there is no significant difference in the effects we are studying between US and non-US subjects. This diversity of subject pool participants is particularly helpful in the present study, given our focus on intuitive motivations that may vary based on life experience. Of course, issues exist when running experiments online that do not exist in the traditional laboratory. Running experiments online necessarily involves some loss of control, since the workers cannot be directly monitored as in the traditional lab; hence, experimenters cannot be certain that each observation is the result of a single person (as opposed to multiple people making joint decisions at the same computer), or that one person does not participate multiple times (although AMT goes to great lengths to try to prevent this, and we use filtering based on IP address to further reduce repeat play). Moreover, although the sample of subjects in AMT experiments is more diverse than samples using college undergraduates, we are obviously restricted to people that participate in online labor markets. To address these potential concerns, recent studies have explored the validity of data gathered using AMT (for an overview, see ref I). Most pertinent to our study are two quantitative direct replications using economic games. The first shows quantitative agreement in contribution behavior in a repeated public goods game between experiments conducted in the physical lab and those conducted using AMT with approximately 10-fold lower stakes2. The second replication again found quantitative agreement between the lab and AMT with 10-fold lower stakes, this time in cooperation in a one-shot Prisoner's Dilemmas. The latter study also conducted a survey on the extent to which subjects trust that they will be paid as described in the instructions (a critical element for economic game experiments) and found that AMT subjects were only slightly less trusting than subjects from a physical laboratory subject pool at Harvard University (trust of 5.4 vs 5.7 on a 7-point Likert scale). A third study compared behavior on AMT in games using $1 stakes with unincentivized games, examining the public goods game, the dictator game, the ultimatum game and the trust games. Consistent with previous research in the physical laboratory, adding stakes was only found to affect play in the dictator game, where subjects were significantly more generous in the unincentivized dictator game compared to the $1 dictator game. Furthermore, the average behavior in these games on AMT was within the range of WWW.NATURE.CONVNATURg 12 EFTA01146519 doi:10.1038/nature 11467 RESEARCH SUPPLEMENTARY INFORMATION averages reported from laboratory studies, demonstrating further quantitative agreement between AMT and the physical lab. In additional studies, it has also been shown that AMT subjects display a level of test-retest reliability similar to what is seen in the traditional lab on measures of political beliefs, self- esteem, Social Dominance Orientation, and Big-Five personality traits'', as well as belief in God, age, gender, education level and incomeI•6; and do not differ significantly from college undergraduates in terms of attentiveness or basic numeracy skills, as well as demonstrating similar effect sizes as undergraduates in tasks examining framing effects, the conjunction fallacy, and outcome bias7. The present studies add another piece of evidence for the validity of experiments run on AMT by comparing our AMT studies with decision time data from previous laboratory experiments (Main text Figure 2): Both online and in the lab, subjects that take longer to make their decisions are less cooperative. 2. Log-transforming decision limes In several of our experiments, we predict cooperation as a function of decision times. However, the distribution of decision times (measured in seconds) is heavily right-skewed, as we did not impose a maximum decision time (decision times for the baseline decision time experiment, Study 1, are shown in Figure S la). Thus linear regression is not appropriate using non- transformed decision times, as the few decision times that are extremely large exert undue influence on the fit of the regression. To address this issue, we log10-transform decision times in all analyses (log 10 transformed decision times for the baseline decision time experiment are shown in Figure S lb). As reported below, our main results are qualitatively similar if we instead analyze non-transformed decision times and exclude outliers (subjects with decision times more than 3 standard deviations above the mean decision time). a IL O O L 100 200 300 DeOeRN mme0ecaay b4,2 LL O 1.5 WDOOSIOITim? jscenIsil 2.5 Figure SI. (a) Distribution of decision times in the baseline experiment. (b) Distribution of log10 transformed decision times in the baseline experiment. WWW.NATURE-COWNATURII3 EFTA01146520 doi:10.1038/nature 11467 RESEARCH SUPPLEMENTARY INFORMATION 3. Study 1: Correlational decision time experiment on AMT Methods In the baseline experiment (main text Figure I), subjects were recruited using AMT and told they would receive a $0.50 show-up fee for participating, and would have the chance to earn up to an additional $1.00 based on the outcome of the experiment. After accepting the task, subjects were redirected to website where they participated in the study. First subjects were shown the Instructions Screen, where they read a set of instructions describing the following one-shot public goods game: Players interacted in groups of 4; each player received 40 cents; players chose how many cents to contribute to the group (in increments of 2 to avoid fractional cent amounts) and how many to keep; all contributions were doubled and split equally by all group members. After they were finished reading the instructions, subjects clicked OK and were taken to the Contribution Screen. Here they entered their contribution decision and clicked OK. The website software recorded how long it took each subject to make her decision (in seconds), that is, the amount of time she spent on the Contribution Screen. Time spent on the Instructions Screen did not count towards our decision time measure. (Time spent on the Instructions Screen is examined below in Study 10 and shown not to influence cooperation.) After entering their contribution amount, subjects were taken to the Comprehension Screen in which they answered two comprehension questions to determine whether they understood the payoff structure: "What level of contribution earns the highest payoff for the group as a whole?" (correct answer = 40) and "What level of contribution earns the highest payoff for you personally?" (correct answer = 0). Subjects were then taken to a demographic questionnaire and given a completion code. We included comprehension questions after the contribution decision, rather than before as is typical in most laboratory experiments, because we were concerned about the possibility of pushing all of our subjects into a reflective mindset prior to their decision-making. (In SI Section 14, we discuss a supplemental experiment that validates this concern by demonstrating that subjects who complete comprehension questions, including a detailed payoff calculation, before making their decision choose to contribute significantly less than those who complete the comprehension questions afterward). Importantly, we show that our result is robust to controlling for comprehension, indicating that the negative relationship between decision time and cooperation is not driven by a lack of comprehension among the faster responders. Once the decisions of all subjects had been collected, subjects were randomly matched into groups of 4, payoffs were calculated, and bonuses were paid through AMT. Payoffs were determined exactly as described in the instructions, and no deception was used. WWW.NATURe.COM/NATUREI EFTA01146521 doi:10.1038/nature11467 RESEARCH SUPPLEMENTARY INFORMATION Results We begin with descriptive statistics: N=212 Mean Std Contribution 23.83 15.39 Decision time 15.92 22.96 Log I 0(Decision time) 1.03 0.34 Age 28.02 8.73 Gender (0=M, I=F) 0.42 0.49 US Residency (0=N, I=Y) 0.45 0.49 Failed Comprehension (0=N, 1=Y) 0.28 0.45 In the baseline experiment, we ask how the amount of time a subject takes to make her contribution decision relates to the amount contributed. To do so, we perform a set of Tobit regressions with robust standard errors, taking contribution amount as the dependent variable (Table SI). Tobit regression allows us to account for the fact that contribution amounts were censored at 0 and 40 (the minimum and maximum contribution amounts). In the first regression, we take log-10 transformed decision time as the independent variable, and find a significant negative relationship. In the second regression, we show that this effect remains significant when including controls for age, gender, US residency, and failing to correctly answering the comprehension questions, as well as dummies for education level. In the third regression, we show that this effect also remains significant when excluding extreme decision times for which there was comparatively little data (regression 3 includes only subjects with 0.6 < log10(decision time) < 1.2). We also continue to find a significant negative relationship between decision time and contribution (coeff=-0.497, p=0.018) using non-transformed decision times and excluding outliers (subjects with decision times more than 3 standard deviations above the mean [mean decision time = 15.9, std = 23.0 implies a cutoff of 85 seconds]) and including controls for age, gender, US residency and comprehension. It is worthwhile to note that the average level of contribution (59.6% of the endowment) of our subjects recruited from AMT is well within the range of average contribution levels observed in previous studies. Our PGG uses a marginal per capita return (MPCR) on public good investment of 0.5 (for every cent contributed, each player earns 0.5 cents). We used an MPCR of 0.5, rather than the value of 0.4 used in many previous studies (where contributions are multiplied by 1.6 and split amongst 4 group members), to create more easily divisible numbers and therefore simpler instructions for the AMT workers, many of whom are less sophisticated than university students. Previous lab studies that used an MPCR of 0.5 report average contribution levels of 40%-70% s"12, which are in line with our value of 59.6%. Thus our experiment adds to the growing body of literature demonstrating the validity of data gathered on AMT. WWW.NATURF-COWNATURI I EFTA01146522 doi:10.1038/nature 11467 RESEARCH SUPPLEMENTARY INFORMATION Table SI. PGG contribution regressed against decision time. (1) (2) (3) Decision time (logI0 seconds) -18.42** -15.84** -29.63** (7.285) (6.772) (15.06) US Residency (O=N, 1=Y) 2.829 2.210 (5.113) (5.666) Age 0.695 0.502 • • Gender (CM, 1=F) 0.402 2.598 (4.104) (4.794) Failed Comprehension ((-.N, 1=Y) -5.886 -8.789 (4.459) (5.306) Education dummies No Yes Yes Constant 49.01*** 25.91 25.21 (8.091) (22.99) (24.27) Observations 212 212 156 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 4. Studies 2 - 5: Reanalysis of previously published experiments run in the physical laboratory Here we analyze decision time data from all of our previously published cooperation experiments for which decision times were recorded1346. These experiments were conducted in the physical laboratory with Boston area college student participants, using the experimental software Ztreel I. We leverage the fact that Ztree automatically records decision times. Thus, although these experiments were not originally conducted to explore the role of intuitive versus reflective cognitive processes, that fact that we find the same negative relationship between cooperation and decision time found in our online experiments demonstrates the robustness of the effect to variations in experimental design, subject pool, and online versus physical laboratory recruitment/implementation. We note that unlike our AMT experiments, in these lab studies the subjects completed simple comprehension quizzes prior to beginning the experiment (with the exception of ref 16 which did not have a comprehension quiz). The details of these quizzes varied across experiments, but none involved the multiple detailed payoff calculations sometimes used in PGG laboratory experiments. Typical questions in our experiments, where subjects played Prisoner's Dilemma games, included "What is the probability of a subsequent round after round 1? After round 10?" or reading off entries in a Prisoner's Dilemma payoff matrix, such as "If you chose A and the other person chooses B, how many points do you get?" See SI Section 14 for a supplemental experiment exploring the effect of asking comprehension questions with detailed payoff calculations before versus after the contribution decision. WWW.NATURLCOM/NATUREI 6 EFTA01146523 doi:10.1038/nature11467 RESEARCH SUPPLEMENTARY INFORMATION We begin by analyzing the control treatment from ref 13 in which 48 subjects played a series of one- shot Prisoner's Dilemma games. In each interaction, subjects were randomly paired, and each pair simultaneously chose to either pay 10 units to give their partner 30 units (i.e. cooperate) or to do nothing (i.e. defect). After making a decision and being informed of their partner's decision, subjects were randomly rematched with new partners for another interaction. Players were given no information about their partner's play in previous games. In total, 29 such interactions occurred. We focus on the first decision subjects made in the experimental session (i.e. the first interaction). The first decision most cleanly represents the intuitions subjects bring into the laboratory by minimizing in-game learning, and also maximizes comparability to our one-shot experiments. Examining cooperation in the first interaction (using logistic regression with robust standard errors), we find a significant negative relationship between cooperation probability and decision time (coeff=-3.42, p4).014; Figure S2A). This relationship continues to exist (coeff=-3.37, p=0.062) when excludin decision times with relatively few observations (times less than 10°A seconds or more than 101. seconds). Using logistic regression with robust standard errors clustered on subject and session, we continue to find a significant effect (coeff=-0.95, p=0.047) when considering the first 5 interactions and controlling for interaction number, albeit with a smaller coefficient; but no longer find a significant effect when considering all 29 interactions (coeff=-0.03, p=0.931). Table S2. Cooperation in series of 1-shot PDs (data from Pfeiffer et aL (2012) J Royal Society Interface). Logistic regression. (1) (2) (3) (4) (5) Interaction 1 Ints 1-5 Ints 1-5 All Ints All Ints Decision time (logIO seconds) -3.417** -0.243 M.951** 0.268 -0.0261 (1.394) (0.432) (0.480) (0.306) (0.301) Interaction # -0.342*** -0.0542 (0.115) (0.0384) Constant 2.939** -0.370 1.092* -1.474*** -0.567 (1.308) (0.546) (0.632) (0.401) (0.639) Observations 48 240 240 1,392 1,392 Robust standard errors in paren heses *** <0.01, ** <0.05, * <0.1 When considering ref 13, we focus on the control condition described above because it demonstrates that our effect exists in one-shot games in the physical laboratory. The effect is not restricted, however, to the control condition. If we instead analyze the data from the 176 subjects that played a stochastically repeated indirect reciprocity game, we continue to find a negative relationship between decision time and cooperation. In these experiments, the setup is the same as the control, except that there is a reputation system such that after each PD, subjects' reputations are updated (to be either `good' or 'bad') based on an explicit assignment rule that is known to the subjects. There were three such conditions, with the assignment rule varying across conditions. Furthermore, subjects were allowed to buy and sell their reputations in two of the conditions. See ref 13 for more details. WWW.NATURS.COM/NATURE I] EFTA01146524 doi:10.1038/nature11467 RESEARCH SUPPLEMENTARY INFORMATION Examining cooperation in the first interaction (using logistic regression with robust standard errors), we find a significant negative relationship between cooperation probability and decision time (coeff=-1.85, p=0.002). This relationship continues to hold (coeff=-1.46, ps3.027) when including condition dummies. Using logistic regression with robust standard errors clustered on subject and session, we continue to find a significant effect when considering all 29 interactions (no controls: coeff=-1.31, p<0.001; controlling for round number and condition dummies: coeff = -0.96, p<0.001). Ref 13 also included a set of fixed-length game conditions that we do not reanalyze as the decision time data for those conditions are not available. Next we consider ref 15, where 278 subjects played a series of stochastically repeated 2-player Prisoner's Dilemma games with execution errors. In each round, there was a 1/8 probability of a player's move being switched to the opposite, and a 7/8 probability of a subsequent round occurring. The benefit-to-cost ratio of cooperation was varied across four different conditions, with b/c=[1.5, 2, 2.5 and 4). Examining cooperation in the first round of the first interaction (using logistic regression with robust standard errors and including condition dummies), we find a significant negative relationship between intended cooperation probability and decision time (coeff =-1.43, ps3.005, including controls for b/c ratio; Figure S2B). This relationship continues to exist (coeff=-1.15, ps3.053) when excluding decision times with relatively few observations (times of than 10 seconds). Moreover, we continue to find a significant effect when considering all decisions over the course of the session (standard errors clustered on subject and group, coeff=-0.97, p<0.001, including controls for b/c ratio, interaction number and round number), albeit with a smaller coefficient. Regressions are shown in Table S3. Table S3. Cooperation in stochastically repeated PD with execution errors (data from Fudenberg et al 2012 AER). Logistic regression. (1) (2) (4) (5) (6) 1st decision 1st decision All decisions All decisions All decisions Decision time (log10 seconds) -1.295*** -1.427*** -0.731*** M.777*** A.970*** (0.478) (0.504) (0.161) (0.119) (0.141) Interaction # 0.0199 (0.0124) Round # -0.187*** (0.0122) Condition dummies No Yes No Yes Yes Constant 0.937*** 1.342*** 0.132 0.459*** 1.296*** (0.222) (0.312) (0.156) (0.138) (0.189) Observations 278 278 26,584 26,584 26,584 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Next we analyze ref 14, where 104 subjects played a series of stochastically repeated 2-player Prisoner's Dilemma games (without execution errors). After every round, there was a 3/4 probability WWW.NATURLCOM/NATUREI EFTA01146525 doi:10.1038/nature11467 RESEARCH SUPPLEMENTARY INFORMATION of a subsequent round. The benefit-to-cost ratio of cooperation and the availability of a 3nd option for costly punishment (pay 1 for the other to lose 4) were varied across treatments (4 treatments: low b/c without punishment, low b/c with punishment, high b/c without punishment, high b/c with punishment). Examining cooperation in the first round of the first interaction (using logistic regression with robust standard errors and including dummies for treatment), we again find a significant negative relationship between cooperation probability and decision time (coeff=-2.67, p=3.018; Figure S2C). This relationship continues to hold (coeff=-2.78, p".1.031) when excluding decision times with relatively few observations (times less than 10°A seconds or more than 101 seconds). Furthermore, we continue to find a significant relationship when analyzing all decisions over the course of the session (standard errors clustered on subject and group, coeff=-0.53, p=0.002), although the coefficient is smaller than in the first period. Regressions are shown in Table 54. Table S4. Cooperation in stochastically repeated PD w thAvithout costly punishment (data from Dreher et al 2008 Nature). Logistic regression. (1) (2) (4) (5) (6) 1st decision 1st decision All decisions All decisions All decisions Decision time (logI0 seconds) -2.741** -2.660** 0.254 -0.528*** -0.554*** (1.107) (1.123) (0.203) (0.171) (0.178) Interaction # -0.0128* (0.00752) Round # -0.361*** (0.0313) Condition dummies No Yes No Yes Yes Constant 2.887*** 2.522*** -0.275** 0.568*** 1.741*** (0.882) (0.961) (0.117) (0.210) (0.291) Observations 104 104 8,120 8,120 8,120 Robust standard errors in parentheses *** p<0.01, ** p<0.05, * p<0.1 Finally, we consider ref 16, where 192 subjects played a repeated public goods game with persistent groups and identities. Subjects were given no information regarding the length of the game, which lasted 50 rounds. The possibility of targeted interaction was varied across four conditions: control PGG, PGG with costly punishment, PGG with costly reward, and PGG with both punishment & reward. As in our 1-shot PGG, tobit with robust standard errors find a significant negative correlation between first round contribution (0-20) and decision time (coeff=-26.38, p=0.001, including condition dummies; Figure S2D). This relationship continues to hold (coeff=-23.01, p=1006) when excluding decision times with relatively few observations (times less than 10°° seconds or more than 1012 seconds). WWW.NATURLCOM/NATUREI 9 EFTA01146526 doi:10.1038/nature11467 RESEARCH SUPPLEMENTARY INFORMATION The relationship between contribution and decision time, however, decays with experience: we find a significant effect when analyzing the first 10 periods (linear regression