1. Introduction

Individuals rely on various reasoning styles to articulate their preferences regarding an inherent ambivalence, presenting advantages and disadvantages without a definitive objective resolution (Kaplan, 1972). Our primary focus is on the reasoning style known as critical thinking (Halpern, 2013), in which individuals recognize the ambivalence of the issue and, through introspection, derive a reasoned preference.

This paper presents a simple incentivized experiment to identify, and classify transitions between different reasoning styles. It emphasizes the role of storytelling formats in shifting individuals from Consequently, this research contributes to the growing experimental literature that aims to identify such reasoning styles and examine their impact on policy (List, 2022). It further expands the spectrum of reasoning styles investigated in the behavioral literature, which is typically focused on motivated reasoning (Kunda, 1990; Bénabou & Tirole, 2006).

In addition, our study emphasizes the role of tailored storytelling formats, where factual information is conveyed through a specific visual design and writing style -- commonly known as "UX design" in the marketing and communication literature. This is instrumental in shifting individuals from stereotypical to critical thinking when hard facts are absent. The interplay of information quantity, quality, and personal cognitive styles is crucial in shaping preferences. In the digital economy, media acts as a significant catalyst, encouraging critical thinking. When confronted with dilemmas, individuals find that mere reliance on 'objective facts' is insufficient. Instead, critical thinking is key to developing well-reasoned preferences.

Consequently, we adopt the terms "story" and "storytelling formats" to represent "media content" and "media format," respectively. The manner in which an issue is presented, ranging from a simplistic tweetstorm to a detailed newspaper article, can influence individual awareness of an issue's ambivalence. Our main findings indicate that two similar interpretations of the same fact, presented through different visual formats and writing styles, elicit different behavioral responses in participants. This expands the conventional definition of "narratives" in economics, generally defined as a specific interpretation of a fact (Shiller, 2017; Eliaz & Spiegler, 2020), by considering the specific format through which the interpretation is presented. In summary, when it comes to dilemmas, the influence of information on individual preferences lies not just in the content, but also in the delivery.

Our experimental design involves three primary stages. Initially, we categorize participants as stereotypical or critical thinkers on a contentious topic using a combination of self-report measures and incentivized elicitation techniques. Subsequently, we expose participants to one of three storytelling interventions, each revolving around the same set of pros and cons related to the issue. The storytelling strategies range from a concise, simplistic style presented in a Twitter-like format, to a medium-level complexity style in a Facebook-like format, to an intricate, detailed style in a newspaper-like format.

In considering social media through the lens of these storytelling perspectives, we contribute to a growing body of literature that increasingly focuses on the impact of specific formats on shaping individual political behaviors, such as voting (Gorodnichenko et al., 2021; Munir, 2018; Falck et al., 2014). More generally, our findings highlight an additional channel (altering the share of critical thinkers) through which social networks can affect welfare, contributing to the rapidly expanding literature on social networks and welfare (Allcott et al., 2020).

After exposure, participants are prompted to write an incentivized critical thinking essay following specific guidelines. Completing this task is incentivized using Large Language Models (LLM), particularly GPT-3, which offers an automatic comparative ranking against a U.S. average score. Given the absence of a definitive "critical thinking" measure, we instituted a secondary experiment to collect expert human feedback. This practice aligns with the model-based reinforcement learning techniques commonly employed by AI-oriented companies, such as OpenAI. Expert feedback was obtained from cognitive psychologists with Ph.D. or higher degrees and experience in ambivalence and critical thinking. The essays authored by the participants were randomly assigned to three independent expert labelers tasked with grading the submissions as pass or fail, depending on their judgment of clear indications of critical thinking in the content. The data from these evaluations were then used to reclassify participants into stereotypical or critical thinkers after the storytelling interventions.

Using psychological literature on cognitive sophistication, we measured participants' need for cognition (Cacioppo & Petty, 1982) and cognitive flexibility (Martin & Rubin, 1995) throughout our experiment. These metrics enabled us to conduct a heterogeneity analysis. Intriguingly, our results indicate that individuals with a higher need for cognition transition more quickly from stereotypical to authentic preferences upon exposure to a medium level of storytelling (i.e., Facebook) than a lower (Twitter) or higher one (Newspaper).

Upon establishing the role of storytelling as a catalyst for critical thinking, we examined its implications for industrial organization and political economy. Specifically, we explore how storytelling techniques impact the efficiency of surveys and elections, rendering them crucial for social welfare in industrial and political decision-making contexts. Indeed, decision-makers who use surveys and run elections may find critical thinking preferences to offer more reliable data than raw preferences.

Consider a public figure or organization whose social image or economic returns hinge on the public endorsement of their position on a particular issue. Such a principal needs to anticipate the public's expected stance, as public endorsements serve as reputational commitments and "focusing events" that prompt individuals to evaluate their raw preferences and establish reasoned preferences critically. Therefore, the principal should gauge the public's reasoned preferences before declaring a stance, minimizing the risk of sustained backlash. Suppose that such an estimate is based on a poll. In that case, its precision depends on the respondents reporting their reasoned preferences, which requires that these preferences have been formed in the first place.

Furthermore, consider an institutional principal, such as a policymaker, tasked with formulating an economic policy on a societal issue that presents a binary dilemma. The principal can select from a wide spectrum of policy alternatives. The optimal policy is a function of the distribution of reasoned preferences, i.e., the proportion of individuals who prefer one alternative over the other after engaging in critical thinking. This establishes the need for the principal to anticipate (and incentivize) the formation of agents' reasoned preferences before making a decision, since the reasoned preference distribution forms its normative criterion.

In both instances, we identify a principal who is interested in the distribution of reasoned preferences: either out of fear that their actions will provoke a backlash if they deviate excessively from the target or because they use such a distribution "for lack of anything better" as an appropriate normative criterion for the social aggregation of preferences. Elections would be efficient if all individuals reported their reasoned preferences at the poll, enabling precise estimation of the relevant unknown. However, individuals arrive at their reasoned preferences only after participating in a critical thinking process, a process not all individuals may have completed by the time the election is held.

The remainder of this paper is organized as follows. Section 1 details our experimental design. Section 2 discusses our empirical findings. Section 3 presents our behavioral model, accompanied by its key positive and normative results. Finally, Section 4 provides a conclusion in which we discuss potential limitations and future extensions of our model and experiment.

Read the rest of the paper