About five decades ago the first pioneers began exploring human behaviour in economic games using the experimental method, sparking a revolution in microeconomics. These early experiments showed that human behaviour often diverged from the predictions of expected monetary maximisation. Yet these deviations were not random. Instead, they exhibited consistent patterns. The exploration and understanding of these patterns have led to the development of one of the most dynamic and influential fields of research in the last half-century: behavioural economics.
One revolutionary insight brought to light by behavioural economics is that humans frequently rely on heuristics – mental shortcuts to navigate a complex world where calculating the exact outcomes of available actions is often beyond the capabilities of the finite human mind. While generally effective, these heuristics can introduce systematic biases. For instance, individuals tend to value losses more than equivalent gains, prefer their own possessions over others’, discount future rewards more heavily than expected utility theory would suggest, and exhibit a reluctance to change the status quo. The classification of heuristics and biases, along with their application in policymaking, has inspired thousands of research papers, including seminal works that have earned at least three Nobel Prizes.
Another groundbreaking insight was the recognition that humans not only care about their own monetary payoff but also care about the rewards received by others. Take, for instance, the well-known Dictator Game. In this scenario, a participant is given a certain amount of money and has to decide how much, if any, to give to another participant. The recipient has no active role and can only accept whatever the dictator chooses to give. Hundreds of experiments have demonstrated that a considerable portion of participants opt to share their endowment, even under conditions of anonymity, even when it is explicitly stated that the recipient cannot affect the dictator’s future earnings, and even after dictators are asked which course of action would maximise their monetary gain. Similar findings have also emerged in decision-making scenarios designed to capture other social interactions, such as cooperation, trust, and truth-telling.
These findings have inspired the development of alternative models of human behaviour. To explain social behaviours, scholars have introduced social preferences. According to these preferences, people do not merely maximise their monetary payoff but consider some combination of their payoff and that of others. This general idea can be detailed in various ways. For example, Ledyard (Reference Ledyard, Kagel and Roth1995) suggested that decision-makers weigh their monetary gain against the aggregate gains of all players. Conversely, Fehr and Schmidt (Reference Fehr and Schmidt1999) proposed that decision-makers compare their monetary reward with the reward of other players. A critical observation is that, despite these specific differences, all social preferences share one foundational ‘consequentialist assumption’: the utility a decision-maker assigns to an action depends solely on its monetary impact on all participants involved.
This consequentialist assumption is deeply ingrained in the economic modelling of human behaviour. It is indeed met also by the primary formal models introduced to account for human heuristics and biases. Take, for example, prospect theory, arguably the most widely adopted theory to explain human decisions under uncertainty. According to this theory, the utility function of an individual is steeper for losses than for gains, encapsulating the well-documented observation that ‘losses loom larger than gains’. This function is also concave for gains and convex for losses, formalising another human tendency: diminished sensitivity. Furthermore, the objective probabilities of events are modified through a weighting function designed to formalise another well-documented human bias: the tendency to overestimate the likelihood of highly unlikely events and underestimate that of highly probable ones. Setting these details aside, it is crucial to recognise that this utility function, too, is ‘economically consequentialist’: it is determined exclusively by the economic outcomes of the available actions.
Economic consequentialism is also met by formal models designed to explain three classic human biases: the endowment effect, the status-quo bias, and the present bias. The endowment effect refers to the tendency to value one’s property more highly than that of others. The standard approach to modelling the endowment effect involves introducing a reference point. This allows connecting the endowment effect with loss aversion: valuing one’s property more because losing it is perceived as worse than acquiring someone else’s property. This model is economically consequentialist, relying solely on the economic outcomes of available actions. Similarly, the status-quo bias is typically explained by attributing a cost to deviation from a reference point, the status quo: once again, an economically consequentialist explanation. Finally, the predominant formalisation of present bias involves the assumption that individuals discount future payoffs hyperbolically rather than exponentially, as predicted by standard utility theory. But hyperbolic discounting is a function of monetary payoffs, making this modelling as well economically consequentialist in nature.
To summarise, it is not an exaggeration to state that behavioural economics is founded upon the consequentialist assumption, considering that this assumption underpins so many economic models of human behaviour. But then, a critical question emerges. Is this assumption valid, or do people frequently violate the consequentialist assumption? Clearly, if a significant number of individuals violate this assumption, it would necessitate a radical revision of nearly all economic models of human behaviour.
The first thesis of this book posits that the consequentialist assumption is indeed consistently violated. The second objective contends that large language models can facilitate a paradigm shift within behavioural economics from outcome-based utility functions to a more general class of functions that better describe human behaviour.
Before discussing how large language models can play a crucial role, it is essential to review how the consequentialist assumption has been contested. A major critique has been advanced by experimental research indicating that the specific words used to describe the decision context and the available actions significantly influence decision-making, impacting beyond the monetary outcomes for individuals involved in the interaction. In a pioneering series of studies, Ross and colleagues demonstrated that linguistic labels could affect behaviour in the Prisoner’s Dilemma (Ross & Samuels, Reference Ross and Samuels1993; Kay & Ross, Reference Kay and Ross2003; Liberman, Samuels, & Ross, Reference Liberman, Samuels and Ross2004). Specifically, participants were more likely to cooperate when the game was described as a ‘community game’ rather than a ‘Wall Street game’. Numerous other studies have since shown that linguistic descriptions impact a variety of human behaviours, including altruism and cooperation, as well as how people resolve conflicts between equitable and efficient distributions of resources. Some studies have also uncovered a darker aspect of linguistic framing. For instance, Ścigała and colleagues (Reference Ścigała, Zettler, Pfattheicher and Capraro2022) discovered that individuals with higher levels of the Honesty-Humility personality trait (typically associated with moral behaviour) could be persuaded to accept a bribe simply by labelling it a ‘cooperation act’. Together, these studies challenge the consequentialist assumption, showing that utility functions cannot solely rely on the economic outcomes of available actions. Instead, utility functions must also consider the language used to describe these actions and the overall context. Consequently, in a recent review in the Journal of Economic Literature joint with Joe Halpern and Matjaz Perc, we have argued that ‘behavioral economics is in the midst of a paradigm shift from outcome-based to language-based preferences’ (Reference Capraro, Halpern and Perc2024a, 115).
To be fair, economists have long recognised that the linguistic content used to describe available actions can influence decisions. Experimentalists typically put considerable effort into writing ‘neutral’ instructions for their experiments, labelling any deviation from this neutrality as a potential source of ‘demand effect’. However, the pursuit of truly neutral instructions is like chasing a chimera: perfectly neutral instructions are unattainable. Many words commonly used to describe economic decisions (e.g., ‘give’, ‘take’) inherently carry connotations and thus cannot be considered neutral. Like a chimera, neutral instructions might be approximated but can never be fully realised. And even more importantly, striving for neutral experimental instructions is a misplaced effort, which moves us away from, rather than towards, a better understanding of human behaviour. In reality, people ‘demand’ things from others all the time, often through language. The ‘demand effect’ should not be feared as if it were the worst nightmare; it should instead be studied and quantified.
And here we arrive at large language models. The second thesis of this book is that large language models can transform behavioural economics for two primary reasons.
The first is that people are increasingly using large language models as support for decision-making, a trend already manifesting itself across various sectors. For example, large language models are being employed to analyse market trends and assist in medical diagnoses. They are becoming an integral part of complex decision-making processes, where they can offer predictions, generate insights, and suggest actions. This is reflected in emerging research at the interface between artificial intelligence and behavioural economics. This research has started exploring how large language models behave in economic games as well as how they predict humans would behave in these games. In a 2023 article in the Proceedings of the National Academy of Sciences of the USA, Chen and colleagues (Reference Chen, Liu, Shan and Zhong2023) arrived at the conclusion that ‘GPT could have the potential in assisting human decision-making.’ However, this support is fundamentally based on language, meaning that people will increasingly engage in language-based, human–machine interactions to aid decision-making. Consequently, the rise of large language models and their integration into decision-making support systems will amplify the demand for language-based utility functions capable of accurately reflecting human behaviour.
This increasing demand for language-based utility functions would remain unmet if there were not, simultaneously, a growing capability to quantitatively assess textual content. Arguably, one of the reasons economists have historically been ‘afraid’ of the demand effect and the influence of language in economic decisions has been the absence of methods to quantify it. However, this challenge is now increasingly addressable through advancements in text analysis tools, which for the first time offer the ability to quantify what was once considered unquantifiable.
Sentiment analysis represents a set of tools developed by computational linguists to quantify the sentiment tenor of a piece of text. Basic sentiment analysis tools can determine whether a piece of text expresses a positive or negative sentiment. More advanced sentiment analysis tools can break down the sentiment into more fundamental emotions, such as joy, fear, anger, or disgust. These tools have found applications in numerous contexts, ranging from online reviews to political communication. However, sentiment analysis has not been extensively applied in understanding human behaviour in economic games. This is because, while sentiment analysis represents a powerful tool for quantifying the emotional tenor of a single piece of text, explaining and modelling human behaviour in decision-making contexts requires an additional, hierarchical, step: assigning a sentiment value to a piece of text within the context described by another piece of text. In other words, it is not sufficient to quantify the linguistic description of a decision-making context; one needs to quantify also the linguistic description of each available action within the linguistic description of that context.
This is the second reason why large language models can play a key role in the transition towards language-based preferences: they could facilitate the process of assigning hierarchical sentiment values to actions within a context. The first step in this direction has been conducted by Rathje and colleagues (Reference Rathje, Mirea, Sucholutsky, Marjieh, Robertson and Van Bavel2024), who demonstrated that GPT-4 accurately detected sentiments with a precision comparable to fine-tuned machine learning models. Going a step further, in a joint work with Roberto Di Paolo, Matjaz Perc, and Veronica Pizziol (Reference Kuang and Bicchieri2024c), we have shown that GPT-4 can assign hierarchical values to actions within a context in a manner that significantly predicts human behaviour within that context. Specifically, we prompted GPT-4 with the textual instructions of 61 Dictator Game experiments. For each experimental instruction (which we interpret as the context in which a decision takes place), we asked GPT-4 to estimate the sentiment value of the available actions in that context, ranging from ‘extremely negative’ to ‘extremely positive’. The results demonstrated that these sentiment scores could explain human behaviour, beyond just the economic outcomes. Of course, the predictions are not perfect but, still, they represent the first instance where textual content has been quantified to predict human behaviour in economic games more accurately than solely through monetary outcomes.
This research represents a starting point because the information conveyed by a piece of text is generally far more complex than a simple binary classification into negative or positive emotions. People may experience a multitude of emotions, each affecting behaviour in distinct ways. For example, Motro, Ordóñez, Pittarello, and Welsh (Reference Motro, Ordóñez, Pittarello and Welsh2018) have demonstrated that inducing anger tends to increase unethical behaviour, while inducing guilt tends to promote ethical behaviour. Therefore, if different emotions can influence behaviour in different ways, it becomes crucial for future research to quantify the diverse emotions associated with the available actions in a given context.
But this is only one layer of complexity. A piece of text can also convey information about norms of behaviour, such as what other people typically do in a situation, or what is approved or disapproved of. Additionally, a piece of text may activate personal moral values. There is considerable debate among moral psychologists, philosophers, and anthropologists regarding the fundamental moral dimensions. However, there is a clear consensus: morality is not monolithic but decomposes into several dimensions. According to (the basic version of) moral foundations theory, these dimensions are care, fairness, loyalty, authority, and purity. According to morality-as-cooperation theory, the dimensions are family, reciprocity, loyalty, heroism, authority, fairness, and property. These dimensions influence human behaviour in diverse ways in important critical ethical issues, such as the decision to vaccinate children. Here, purity leans individuals towards vaccine hesitancy, while authority tends towards vaccine acceptance.
The influence of moral values on behaviour suggests that sentiment analysis should be complemented by ‘normative analysis’: language models capable of capturing the normative tenor of text along various dimensions. Recent research is beginning to explore this area. In the last few years, some corpora of annotated texts have been collected based on moral foundations theory, marking an initial step towards developing normative analysis.
And of course, emotions and norms can interact in complex ways. For example, collective emotions may forge social norms, while, conversely, the enforcement of certain norms can evoke emotions. These complex interactions between various emotions and norms influence the final strategic decisions of a decision-maker, adding another web of complexity. A useful qualitative model to organise these complexities is the LENS model, deriving its name from the key pathways of the model: the Linguistic content triggers Emotions and suggests Norms, which interact and determine Strategy choice. Large language models can play a crucial role in transitioning this qualitative model into a quantitative one, through sentiment and normative analysis applied to actions within a context.
Finally, it is important to recognise that integrating large language models into behavioural economics has implications beyond the academic world. It can reshape everyday decision-making, business strategies, and even policymaking. On the one hand, by using large language models to understand and predict human behaviour more accurately, individuals can make more informed choices, businesses can tailor products and services to better meet consumer needs, and policymakers can design more effective interventions to promote public welfare. On the other hand, there are also major ethical considerations to consider. Over-reliance on large language models could lead to manipulation, privacy infringements, and algorithmic biases that can perpetuate or even exacerbate existing inequalities. Moreover, ensuring that the recommendations of large language models align with human intent may be challenging, especially in situations involving ethical trade-offs between competing moral values. Chapter 9 will engage with these concerns, encouraging the development of transparent models that enhance rather than undermine human agency.
The journey ahead is complex yet exciting, crossing several disciplines, including psychology, computational linguistics, artificial intelligence, and ethics. Some paths are yet to be charted, while others may lead to dead ends that require revaluation. And though this exploration might resemble the pursuit of a chimera, and we may never fully decipher human behaviour, it nevertheless brings behavioural economics closer to its foundational purpose: understanding human behaviour in economic interactions.
