research

Published

Adverse Reactions to the Use of Large Language Models in Social Interactions

with Regina Stumpf, Sebastian Fehrler, Urs Fischbacher
PNAS Nexus | paper | online appendix | replication package | preregistration

Details

Large language models are poised to reshape the way individuals communicate and interact. While this form of artificial intelligence has the potential to efficiently make many human decisions, there is limited understanding of how individuals will respond to its use in social interactions. In particular, it remains unclear how individuals interact with large language models when the interaction has consequences for other people. Here, we report the results of a large-scale, pre-registered online experiment (N = 3,552) showing that human players’ fairness, trust, trustworthiness, cooperation, and coordination in economic two-player games decrease when the decision of the interaction partner is taken over by ChatGPT. On the contrary, we observe no adverse reactions when individuals are uncertain whether they are interacting with a human or a large language model. At the same time, participants often delegate decisions to the large language model, especially when the model’s involvement is not disclosed, and individuals have difficulty distinguishing between decisions made by humans and those made by artificial intelligence.

Reference: Dvorak, F., Stumpf, R., Fehrler, S., and Fischacher, U. (2025). Adverse Reactions to the Use of Large Language Models in Social Interactions. PNAS Nexus, pgaf112, https://doi.org/10.1093/pnasnexus/pgaf112

Strategic Conformity or Anticonformity to Avoid Punishment and Attract Reward

with Urs Fischbacher and Katrin Schmelz
The Economic Journal | paper | online appendix | replication package | video creativity task

Details

We provide systematic insights on strategic conformist - as well as anticonformist - behavior in situations where people are evaluated, i.e., where an individual has to be selected for reward (e.g., promotion) or punishment (e.g., layoffs). To affect the probability of being selected, people may attempt to fit in or stand out in order to affect the chances of being noticed or liked by the evaluator. We investigate such strategic incentives for conformity or anticonformity experimentally in three different domains: facts, taste, and creativity. To distinguish conformity and anticonformity from independence, we introduce a new experimental design that allows us to predict participants’ independent choices based on transitivity. We find that the prospect of punishment increases conformity, while the prospect of reward reduces it. Anticonformity emerges in the prospect of reward, but only under specific circumstances. Similarity-based selection (i.e., homophily) is much more important for the evaluators’ decisions than salience. We also employ a theoretical approach to illustrate strategic key mechanisms of our experimental setting.

Reference: Dvorak, F., Fischbacher, U., and Schmelz, K. (2025). Strategic conformity or anticonformity to avoid punishment and attract reward. The Economic Journal, 135(666), 556–583. https://doi.org/10.1093/ej/ueae085

Negotiating Cooperation under Uncertainty: Communication in Noisy, Indefinitely Repeated Interactions

with Sebastian Fehrler
American Economic Journal: Microeconomics | paper | online appendix | replication package | cooperation barcode

Details

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 conduct a laboratory experiment to study how communication affects cooperation under different monitoring structures. Pre-play communication reduces strategic uncertainty and facilitates very high cooperation rates at the beginning of an interaction. Under perfect monitoring, this is sufficient to reach a high and stable cooperation rate. However, repeated communication is important to maintain a high level of cooperation under imperfect monitoring, where players face additional uncertainty about the history of play.

Reference: Dvorak, F., Fehrler, S. (2024). Negotiating cooperation under uncertainty: communication in noisy, indefinitely repeated interactions. American Economic Journal: Microeconomics, 16(3), 232–258. https://doi.org/10.1257/mic.20210117

stratEst: A Software Package for Strategy Frequency Estimation

Details

stratEst is a software package for strategy frequency estimation in the freely available statistical computing environment R (R Development Core Team, 2022). The package aims at minimizing the start-up costs of running the modern strategy frequency estimation techniques used in experimental economics. Strategy frequency estimation (Stahl and Wilson, 1994; Stahl and Wilson, 1995) models the choices of participants in an economic experiment as a finite-mixture of individual decision strategies. The parameters of the model describe the associated behavior of each strategy and its frequency in the data. stratEst provides a convenient and flexible framework for strategy frequency estimation, allowing the user to customize, store and reuse sets of candidate strategies. The package includes useful functions for data processing and simulation, strategy programming, model estimation, parameter testing, model checking, and model selection.

Reference: Dvorak, F. (2023). stratEst: a software package for strategy frequency estimation. Journal of the Economic Science Association, 9, 337–349. https://doi.org/10.1007/s40881-023-00141-7

Genetic Modulation of Oxytocin Sensitivity: A Pharmacogenetic Approach

with Frances Chen, Robert Kumsta, Gregor Domes, Onn Yim, Richard Ebstein, and Markus Heinrichs
Translational Psychiatry | paper | online appendix

Details

Intranasal administration of the neuropeptide oxytocin has been shown to influence a range of complex social cognitions and social behaviors, and it holds therapeutic potential for the treatment of mental disorders characterized by social functioning deficits such as autism, social phobia and borderline personality disorder. However, considerable variability exists in individual responses to oxytocin administration. Here, we undertook a study to investigate the role of genetic variation in sensitivity to exogenous oxytocin using a socioemotional task. In a randomized, double-blind, placebo-controlled experiment with a repeated-measures (crossover) design, we assessed the performance of 203 men on an emotion recognition task under oxytocin and placebo. We took a haplotype-based approach to investigate the association between oxytocin receptor gene variation and oxytocin sensitivity. We identfied a six-marker haplotype block spanning the promoter region and intron 3 that was significantly associated with our measure of oxytocin sensitivity. Specifically, the TTCGGG haplotype comprising single-nucleotide polymorphisms rs237917-rs2268498-rs4564970-rs237897-rs2268495-rs53576 is associated with increased emotion recognition performance under oxytocin versus placebo, and the CCGAGA haplotype with the opposite pattern. These results on the genetic modulation of sensitivity to oxytocin document a significant source of individual differences with implications for personalized treatment approaches using oxytocin administration.

Reference: Chen, F. S., Kumsta, R., Dvorak, F., Domes, G., Yim, O.-S., Ebstein, R. P. & Heinrichs, M. (2015). Genetic modulation of oxytocin sensitivity: a pharmacogenetic approach. Translational Psychiatry, 5, e664. https://doi.org/10.1038/tp.2015.163

Working Paper

Similarity and Consistency in Algorithm-Guided Exploration

with Yongping Bao, Ludwig Danwitz, Sebastian Fehrler, Lars Hornuf, Hsuan Yu Lin and Bettina von Helversen
R&R @ Journal of Behavioral Decision Making | working paper

Details

Algorithmic advice has the potential to significantly improve human decision-making, especially in dynamic and complex tasks that require a balance between exploration and exploitation. This study examines conditions under which individuals are willing to accept advice from algorithms in such scenarios, focusing on the interaction between participants’ exploration preferences and those of the advising algorithm. In an online experiment, we designed reinforcement learning algorithms to prioritize either exploration or exploitation and observed participants’ decision-making behavior, modeled using a cognitive framework analogous to the algorithm. Contrary to expectations, participants did not show a preference for algorithms that matched their own exploration tendencies. In particular, participants were more likely to follow the advice of exploitative, consistent algorithms, possibly interpreting consistency as an indicator of competence. Although participants also benefited from the advice of exploratory algorithms, their relative reluctance to follow exploratory advice highlights a potential challenge in promoting effective human-algorithm collaboration. Explorative algorithms have the potential to foster behavioral diversification, but this effect is negated if humans disregard explorative advice. In such cases, algorithmic guidance may inadvertently reduce behavioral diversity by reinforcing established patterns.

Reappraising Stress-Related Arousal Enhances Prosocial Behaviour in Individuals Perceiving the Intervention as Effective

with Laura Oswald, Jeremy P. Jamieson, Markus Heinrichs and Bastian Schiller
R&R @ Anxiety, Stress & Coping

Details

While psychosocial stress is ubiquitous and on the rise in society, the conditions under which stress facilitates or hinders prosocial behaviour remain elusive. To resolve heterogeneous findings, it is essential to consider the modulatory effects of individuals’ appraisals of stress as challenging versus threatening. Here, we tested the hypothesis that strengthening an individual’s challenge appraisal under stress will enhance prosocial behaviour. 121 individuals (62 women, 59 men) participated in the Trier Social Stress Test for Groups. Participants were instructed to either reappraise their upcoming arousal as adaptive (n = 61) or they performed a reading exercise without instructions on how to cope with the psychosocial stressor (n = 60). Participants then engaged in incentivised resource allocation games. The study was preregistered as a clinical trial (DRKS00027518). We observed no general effect of arousal reappraisal on prosocial behaviour. However, reappraising stress-related arousal as a functional resource did enhance prosocial behaviour in individuals who internalised the message and found this intervention to be effective. Our findings suggest that cognitive stress appraisals modulate behavioural stress effects, and appraisals can be leveraged to help improve stress-related response and outcomes.

Sustaining Cooperation with Correlated Information: An Experimental Test

with Yongping Bao and Sebastian Fehrler
submitted | working paper

Details

In infinitely or indefinitely repeated games with noisy signals about others’ actions, sustaining cooperation is difficult. Theoretical work shows that cooperation can be maintained if the signals are correlated and the degree of correlation depends on the actions. In this study, we implement such an information structure in a laboratory experiment and investigate whether subjects are able to sustain cooperation by conditioning their behavior on it. A substantial number of subjects adopt strategies accounting for the correlation, but this does not increase cooperation compared to a control treatment without correlation, as behavior with independent signals is more lenient.

Cognitive Models of Bayesian Anchoring in Discrete Choice Experiments

with Klaus Glenk, Ivana Logar and Jürgen Meyerhoff
submitted | working paper

Details

Discrete choice experiments are an important method to derive willingness-to-pay estimates for non-market goods. Several studies have shown that willingness-to-pay estimates derived from discrete choice experiments can be sensitive to the order of the presented choice tasks or the size of the presented costs, which raises concerns about the validity of such estimates. In this paper, we present cognitive models of Bayesian anchoring that control for choice-task ordering and cost-vector anomalies in discrete choice experiments. We show that ordering and cost-vector effects arise if respondents update their marginal utility of money based on the costs presented during the experiment and introduce novel models based on Bayesian updating that correct for anchoring processes at the individual level. Using data from a discrete choice experiment on micro- and nanoplastic pollution of freshwater ecosystems in Switzerland, we demonstrate how cognitive modeling can be used to correct willingness-to-pay estimates and discuss the implications for welfare analysis and policy design.

Social Learning with Intrinsic Preferences

with Urs Fischbacher
working paper

Details

Despite strong evidence for peer effects, little is known about how individuals balance intrinsic preferences and social learning in different choice environments. Using a combination of experiments and discrete choice modeling, we show that intrinsic preferences and social learning jointly influence participants’ decisions, but their relative importance varies across choice tasks and environments. Intrinsic preferences guide participants’ decisions in a subjective choice task, while social learning determines participants’ decisions in a task with an objectively correct solution. A choice environment in which people expect to be rewarded for their choices reinforces the influence of intrinsic preferences, whereas an environment in which people expect to be punished for their choices reinforces conformist social learning. We use simulations to discuss the implications of these findings for the polarization of behavior.

Work in Progress

Conformity in Moral Judgments

with Urs Fischbacher, Katrin Schmelz and Georg Sator
work in progress

Details

Morality is one of the key features of the human species. It enables us to live in and benefit from social groups and thereby constitutes the foundation of social coexistence and civilization. The starting point of our study is the notion that many important moral choices are not made in social isolation, but in the presence of other people and other people’s moral judgments. This poses the question whether moral judgment and decision-making is prone to social influence. We conduct an online experiment which allows us to identify the effect of social information on people’s publicly stated moral convictions. The experimental design uses a battery of moral trilemmas allowing use to separate conformity, anticonformity and independence in moral judgments without drawing on deceptive methods.

Public Preferences for Low-Carbon Energy Systems

with researchers of WP6 of SCENE
work in progress

Details

The future decarbonized energy system is expected to rely increasingly on electricity. This will require efficient technologies for heating and energy storage as well as advanced flexibility mechanisms in the residential sector to reduce the load on the electricity system. As part of Work Package 6 of SCENE, we conduct two nationwide discrete choice experiments in Switzerland to identify regional differences in homeowners’ preferences for the adoption of low-carbon energy technologies and tenants’ preferences for demand-side flexibility measures, which will be integrated into energy system modeling to identify socially acceptable pathways to the low-carbon energy system of the future.

Demand for Carbon-Neutral Products

with Stefano Carattini, Ivana Logar and Begüm Özdemir Oluk
work in progress

Details

Corporate social responsibility and the private provision of (global) public goods are of key interest to economists and policymakers alike. Increasingly, private companies are making their operations carbon neutral, often leading their own products to also be certified accordingly. It is an empirical question how consumers value carbon-neutral products, which we address as follows. First, we provide a meta-analysis of the literature analyzing demand for products with carbon-neutral labels, based on an overall sample of 27,241 participants. In this analysis, the focus is on average willingness to pay for carbon reductions as well as on the characteristics of the underlying literature, including the use of stated preferences and population samples, and their association with willingness to pay. Second, we leverage information on prices and product characteristics from one of the largest online marketplaces, Amazon’s, to infer from revealed preferences on consumers’ valuation of carbon-neutral products, through a hedonic approach. The staggered process of carbon-neutral certification leads to a series of quasi-natural experiments, which we use for identification purposes. We find that the literature, which is mainly based on survey studies, suggests a positive willingness to pay for carbon neutrality of products that exceeds most estimates of the social cost of carbon. However, this finding is not supported by the hedonic approach, which is based on market prices, where we do not find evidence for a positive willingness to pay for carbon neutrality for a wide range of products sold on Amazon.

Eliciting Strategies in Repeated Games of Strategic Substitutes and Complements

Matthew Embrey, Friederike Mengel and Ronald Peeters
work in progress

Details

We introduce a novel method to elicit strategies in indefinitely repeated games and apply it to games of strategic substitutes and complements. We find that out of 256 possible unit recall machines (and 1024 full strategies) participants could use, only five machines are used more than 5 percent of the time. Those are ‘static Nash’, ‘myopic best response”, ’Tit-for-Tat’ and two ‘Nash reversion” strategies. We compare outcome data with ’hot’ treatments and find that the fact that we elicit strategies did not affect the path of play. We further compare the frequencies of the elicited strategies with results of the strategy frequency estimation method. We also discuss applications to IO literature and compare insights to previous literature on strategy elicitation mostly focused on the prisoner’s dilemma.