Working Papers
Generative AI Triggers Welfare-Reducing Decisions in Humans (with Fabian Dvorak, Regina Stumpf, and Urs Fischbacher), [link].
Abstract Generative artificial intelligence (AI) is poised to reshape the way individuals communicate and interact. While this form of AI has the potential to efficiently make numerous human decisions, there is limited understanding of how individuals respond to its use in social interaction. In particular, it remains unclear how individuals engage with algorithms when the interaction entails consequences for other people. Here, we report the results of a large-scale pre-registered online experiment (N = 3,552) indicating diminished fairness, trust, trustworthiness, cooperation, and coordination by human players in economic twoplayer games, when the decision of the interaction partner is taken over by ChatGPT. On the contrary, we observe no adverse welfare effects when individuals are uncertain about whether they are interacting with a human or generative AI. Therefore, the promotion of AI transparency, often suggested as a solution to mitigate the negative impacts of generative AI on society, shows a detrimental effect on welfare in our study. Concurrently, participants frequently delegate decisions to ChatGPT, particularly when the AI’s involvement is undisclosed, and individuals struggle to discern between AI and human decisions.
Social Norms and Female Labor Force Participation in Bangladesh: The Role of Social Expectations and Reference Networks (with Luna Bellani, Kumar Biswas, Paul Marx, Shwetlena Sabarwal, Syed Rashed Al-Zayed Josh), R&R at the Journal of Development Studies, [link].
Abstract About 50% of Bangladesh’s female youth working-age population is not in employment, education, or training (NEET). Reducing this number is an important policy goal. However, there is a broad consensus that pervasive gender norms hamper this goal in Bangladesh and other countries from the Global South. In this study, we analyze the social basis of support for young working women. It departs from a theoretical understanding of norms as conditional upon expectations in one’s reference network. Based on vignette experiments, we show that manipulating expectations about acceptance of female employment by others influences personal support for women taking up work. Moreover, we address the question of whose views matter. Manipulating the expectation that fathers (or husbands in the case of married NEETs) support the employment of their daughters (wives) has a particularly strong effect on respondents’ support. In contrast, the stance of religious authorities and peers has surprisingly little relevance. Our evidence suggests that (expectations about) traditional views of fathers and husbands regarding the role of females are a key obstacle to a higher labor force participation of young women in Bangladesh.
Similarity and Consistency in Algorithm-Guided Exploration (with Yongping Bao, Ludwig Danwitz, Fabian Dvorak, Lars Hornuf, Hsuan Yu Lin, and Bettina von Helversen), [link].
Abstract Algorithm-based decision support systems play an increasingly important role in decisions involving exploration tasks, such as product searches, portfolio choices, and human resource procurement. These tasks often involve a trade-off between exploration and exploitation, which can be highly dependent on individual preferences. In an online experiment, we study whether the willingness of participants to follow the advice of a reinforcement learning algorithm depends on the fit between their own exploration preferences and the algorithm’s advice. We vary the weight that the algorithm places on exploration rather than exploitation, and model the participants’ decision-making processes using a learning model comparable to the algorithm’s. This allows us to measure the degree to which one’s willingness to accept the algorithm’s advice depends on the weight it places on exploration and on the similarity between the exploration tendencies of the algorithm and the participant. We find that the algorithm’s advice affects and improves participants’ choices in all treatments. However, the degree to which participants are willing to follow the advice depends heavily on the algorithm’s exploration tendency. Participants are more likely to follow an algorithm that is more exploitative than they are, possibly interpreting the algorithm’s relative consistency over time as a signal of expertise. Similarity between human choices and the algorithm’s recommendations does not increase humans’ willingness to follow the recommendations. Hence, our results suggest that the consistency of an algorithm’s recommendations over time is key to inducing people to follow algorithmic advice in exploration tasks.
Missing the target: Does increased capacity of the local government improve beneficiary selection? (with Viola Asri, Kumar Biswas, Urs Fischbacher, Katharina Michaelowa, and Atonu Rabbani), [link].
Abstract To improve targeting of social policies, research has focused on incentives and ac- countability of local decision makers. This paper identifies the capacity of the local government as another key constraint which has received little attention. We examine whether and how a locally implemented capacity building intervention in- cluding training and data provision for the national Old Age Allowance program in Bangladesh can improve the selection of beneficiaries. The results of a large-scale clustered randomized controlled trial in 80 rural municipalities show that the in- tervention enhanced the knowledge of eligibility criteria but did not improve the targeting performance. We further document that improvements in the targeting performance depend on selection committee’s willingness to improve beneficiary selection and that the illegitimate collection of fees is prevalent.
Beliefs about Others: A Striking Example of Information Neglect (with Baiba Renerte, and Irenaeus Wolff), [link].
Abstract In many games of imperfect information, players can make Bayesian inferences about other players’ types based on the information that is contained in their own type. Several behavioral theories of belief-updating even start from the assumption that players project their own type onto others also when it is not rational. We investigate such inferences in a simple laboratory task, in which types are drawn from one out of two states of the world and participants have to guess the type of another participant. We find little evidence for irrational (over-)projection. Instead, between 50% and 70% of the participants in our experiment completely neglect the information contained in their own type and base their choices only on the prior probabilities. Using several experimental interventions, we show that this striking neglect of information is very robust.
Publications
Refereed Journals
Negotiating Cooperation under Uncertainty: Communication in Noisy, Indefinitely Repeated Interactions (with Fabian Dvorak), American Economic Journal: Microeconomics, forthcoming, [link] [WP].
Abstract Case studies of cartels and recent theory suggest that communication is a key factor for cooperation under imperfect monitoring, where actions can only be observed with noise. We present results from a laboratory experiment which show that communication facilitates cooperation by reducing two types of uncertainty. Pre-play communication reduces strategic uncertainty, which boosts cooperation at the beginning of an interaction. Under perfect monitoring, this is sufficient to reach a high and stable cooperation rate. However, repeated communication is necessary to maintain a high level of cooperation under imperfect monitoring, where players face additional uncertainty about the history of play.
Committee Decision-Making under the Threat of Leaks (with Volker Hahn), Journal of Politics, 85 (3), 1107-1122, 2023, [link] [WP].
Abstract Leaks are pervasive in politics. Hence, many committees that nominally operate under secrecy de facto operate under the threat that information might be passed on to outsiders. We study theoretically and experimentally how this possibility affects the
behavior of committee members and the decision-making accuracy. Our theoretical analysis generates two major predictions. First, a committee operating under the threat of leaks is equivalent to a formally transparent committee in terms of the probabilities of supporting the adoption of a new policy. Second, the threat of leaks leads to a status-quo bias. In our experimental analysis of a committee with possible leaks, individual behavior is often less strategic than theoretically predicted, which leads to frequent leaks. However, despite these deviations on the individual level, our experiment confirms the two major theoretical predictions.
Delegation to a Group (with Moritz Janas), Management Science, 67, 3714-3743, 2021, [link] [article].
Abstract We study the choice of a principal to either delegate a decision to a group of careerist experts, or to consult them individually and keep the decision-making power. Our model predicts a trade-off between information acquisition and information aggregation. On the one hand, the expected benefit from being informed is larger in case the experts are consulted individually. Hence, the experts either acquire the same or a larger amount of information, depending on the cost of information, than in case of delegation. On the other hand, any acquired information is better aggregated in case of delegation, where experts can deliberate secretly. To test the model’s key predictions, we run an experiment. The results from the laboratory confirm the predicted trade-off, despite some deviations from theory on the individual level.
Buying Supermajorities in the Lab (with Maik T. Schneider), Games and Economic Behavior, 127, 113-154, 2021, [link] [article].
Abstract Many decisions taken in legislatures or committees are subject to lobbying efforts. A seminal contribution to the literature on vote-buying is the legislative lobbying model pioneered by Groseclose and Snyder (1996), which predicts that lobbies will optimally form supermajorities in many cases. Providing the first empirical assessment of this prominent model, we test its central predictions in the laboratory. While the model assumes sequential moves, we relax this assumption in additional treatments with simultaneous moves. We find that lobbies buy supermajorities as predicted by the theory. Our results also provide supporting evidence for most comparative statics predictions of the legislative lobbying model with respect to lobbies’ willingness to pay and legislators’ preferences. Most of these results carry over to the simultaneous-move set-up but the predictive power of the model declines.
Honesty and Self-Selection into Cheap Talk (with Urs Fischbacher and Maik T. Schneider), The Economic Journal, 130, 2468–2496, 2020, [link] [article].
Abstract In many situations, statements are non-binding and people can lie strategically for their own benefit – political campaigns and sales marketing are obvious examples. Since individuals differ with respect to their willingness to lie, the level of lying and thus the credibility of statements will crucially depend on who self-selects into such situations. We study self-selection into cheap talk in the concrete setting of a two-stage political competition model and test its key predictions in the lab. At the entry stage, potential candidates compete in a contest to become their party’s candidate in an election. At the election stage, the nominated candidates campaign by making promises to voters. Confirming the model’s key prediction, we find in the experiment that dishonest people over-proportionally self-select into the political raceand thereby lower voters’ welfare.
How Transparency Kills Information Aggregation: Theory and Experiment (with Niall Hughes), American Economic Journal: Microeconomics, 10, 181-209, 2018, [link] [WP].
Abstract We investigate the potential of transparency to influence committee decision-making. We present a model in which career concerned committee members receive private information of different type-dependent accuracy, deliberate and vote. We study three levels of transparency under which career concerns are predicted to affect behavior differently and test the model’s key predictions in a laboratory experiment.The model’s predictions are largely borne out – transparency negatively affects information aggregation at the deliberation and voting stages, leading to sharply different committee error rates than under secrecy. This occurs despite subjects revealing more information under transparency than theory predicts.
Choosing a Partner for Social Exchange: Charitable Giving as a Signal of Trustworthiness (with Wojtek Przepiorka), Journal of Economic Behavior & Organization, 129, 157-171, 2016, [link] [WP].
Abstract People benefit from being perceived as trustworthy. Examples include sellers trying to attract buyers, or candidates in elections trying to attract voters. In a laboratory experiment using exchange games, in which the trustor can choose the trustee, we study whether trustees can signal their trustworthiness by giving to charity. Our results show that donors are indeed perceived as more trustworthy and they are selected significantly more often as interaction partners. As a consequence of this sorting pattern, relative payoffs to donors and non-donors differ substantially with and without partner choice. However, we do not find donors to be significantly more trustworthy than non-donors. Our findings suggest that publicly observable generosity, such as investments in corporate social responsibility or donations to charity during a political campaign, can induce perceptions of trustworthiness and trust.
Pro-Social Missions and Worker Motivation: An Experimental Study (with Michael Kosfeld), Journal of Economic Behavior & Organization, 100, 99-110, 2014, [link] [WP].
Abstract Do employees work harder if their job has the right mission? In a laboratory labor market experiment, we test whether subjects provide higher effort if they can choose the mission of their job. We observe that subjects do not provide higher effort than in a control treatment. Surprised by this finding, we run a second experiment in which subjects can choose whether they want to work on a job with their preferred mission or not. A subgroup of agents (roughly one third) is willing to do so even if this option is more costly than choosing the alternative job. Moreover, we find that these subjects provide substantially higher effort. These results suggest that some workers can be motivated by missions and that selection into mission-oriented organizations is an important factor to explain empirical findings of lower wages and high motivation in these organizations.
Can You Trust the Good Guys? Trust Within and Between Groups with Different Missions (with Michael Kosfeld), Economics Letters, 121, 400-404, 2013, [link] [WP].
Abstract Non-governmental organizations and other non-profit organizations attract workers who strongly identify themselves with their missions. We study whether these “good guys” are more trustworthy, and how such pronounced group identities affect trust and trustworthiness within the groups and towards out-groups. We find that subjects who strongly identify themselves with a non-profit mission are more trustworthy in a minimal group setting but also harshly discriminate against out-groups when subjects are grouped by the missions they identify themselves with.
Charitable Giving as a Signal of Trustworthiness: Disentangling the Signaling Benefits of Altruistic Acts (with Wojtek Przepiorka), Evolution & Human Behavior, 34, 139-145, 2013, [link] [WP].
Abstract It has been shown that psychological predispositions to benefit others can motivate human cooperation and the evolution of such social preferences can be explained with kin or multi-level selection models. It has also been shown that cooperation can evolve as a costly signal of an unobservable quality that makes a person more attractive with regard to other types of social interactions. Here we show that if a proportion of individuals with social preferences is maintained in the population through kin or multi-level selection, cooperative acts that are truly altruistic can be a costly signal of social preferences and make altruistic individuals more trustworthy interaction partners in social exchange. In a computerized laboratory experiment, we test whether altruistic behavior in the form of charitable giving is indeed correlated with trustworthiness and whether a charitable donation increases the observing agents’ trust in the donor. Our results support these hypotheses and show that, apart from trust, responses to altruistic acts can have a rewarding or outcome-equalizing purpose. Our findings corroborate that the signaling benefits of altruistic acts that accrue in social exchange can ease the conditions for the evolution of social preferences.
The Effectiveness of Inputs in Primary Education: Insights from Recent Student Surveys for Sub-Saharan Africa (with Katharina Michaelowa and Annika Wechtler), Journal of Development Studies, 45, 1545-1578, 2009, [link].
Abstract With SACMEQ and PASEC there are now two large data bases available on student achievement, socio-economic background and school and teacher characteristics in both anglophone and francophone sub-Saharan Africa. A joint analysis of PASEC and SACMEQ in a common education production function framework allows us to estimate the impact of educational inputs on student achievement in 21 sub-Saharan African countries and to compare our results with those of earlier empirical studies for education systems in Africa and other world regions. In our analysis we focus on school equipment, teacher quality and class organisation. The issue of teacher and student incentives cannot be adequately addressed with the given data. Our results are based on a traditional retrospective analysis of student achievement in PASEC and SACMEQ countries. In contrast to the ‘nothing works’ result from most industrialized countries’ studies we find robust positive correlations of achievement test scores and the possession of textbooks and negative correlations with teaching in shifts. The most striking result is the weak or even absent correlation of achievement test scores and teacher education and professional training. However, some differences between francophone and anglophone education systems can be observed in this context if differences in the sampling methodology are duly taken into account.