2.1. Two problems

Any theory of decision-making must account for how beliefs and values yield action. We divide this question into two problems – mediation and combination.

2.1.1. The mediation problem

Since data must be interpreted to be useful, decision-making requires a mental representation – a currency of thought – mediating between the external world and our actions (dashed lines in Fig. 1). When we face a decision, we form beliefs – based on prior knowledge and new data – to characterize what will likely happen given potential actions. Those beliefs must be represented in a format that can be combined with our values to guide action (Baumeister & Masicampo, Reference Baumeister and Masicampo2010). Put differently, external data or raw facts do not become actionable information until interpreted in conjunction with our broader knowledge (Tuckett, Holmes, Pearson, & Chaplin, Reference Tuckett, Holmes, Pearson and Chaplin2020). The burglar must consider, were he to burgle, the likely outcomes (beliefs) and their desirability (values). Some mental representation must simultaneously be the output of the reasoning process that judges what will happen and an input to the decision-making process that combines those beliefs with each outcome's desirability.

Figure 1.

The logic of decision. Decisions reflect both data picked up from the external world – including the social environment – and internally derived goals. The mediation problem (dashed lines) reflects the need for an internal representation – a currency of thought – that can mediate between data from the external world and actions decided internally. The combination problem (gray lines) reflects the need for a process – a driver of action – that can combine beliefs and goals to yield actions. In classical decision theory, the currency of thought is probability and the driver of action is expected utility maximization. In CNT, the currency of thought is narratives, and the driver of action is affective evaluation.

In classical decision theory, the currency of thought is probability – continuous values that quantify risk. The burglar's decision depends on his perceived chance he will be caught (C) or not caught (NC). But this assessment depends potentially on many things – the police presence, burglar's skill, odds the inhabitants are home, etc. These data must be aggregated through Bayesian inference (Section 4.1), combining prior knowledge with new evidence.

The burglar weighs the evidence, assigning 0.2 probability to C and 0.8 to NC. These probabilities summarize all relevant data about the external world in a format used internally to combine these beliefs about outcomes with values about their desirability. Probabilities solve the mediation problem because a single representation can be the output of belief-formation and the input to decision-making.

2.1.2. The combination problem

Decision-making requires a process – a driver of action – that combines beliefs and values to yield action. The burglar must not only assess the likelihood of being caught or not, but how bad or good that would be. If the mediation problem has been solved, we have a suitable representation of likelihood to combine with value judgments. Yet a further principle must govern this combination.

In classical decision theory, the driver of action is utility-maximization: Disparate sources of value are aggregated into an outcome's utility, multiplied by each outcome's probability to yield an option's expected utility. The decision rule is simply to maximize this quantity. The burglar would consider the sources of (dis)utility associated with being caught (C) – social stigma, financial costs, prison – and with not being caught (NC) – newfound wealth, perhaps guilt. The utility of C and NC might be −8 and +3, respectively. Then, the expected utility is each state's utility, weighted by its probability:

$$U( C ) \times P( C ) + U( {NC} ) \times P( {NC} ) = ( {0.2} ) ( {-8} ) + ( {0.8} ) ( 3 ) = { + } 0.8$$

Crime is expected to pay, so the rational burglar would attempt the burglary. Although expected utility maximization is not the only justifiable decision rule, philosophers and economists have marshalled powerful arguments for its rationality (Savage, Reference Savage1954; Von Neumann & Morgenstern, Reference Von Neumann and Morgenstern1944).

2.2. Two puzzles

The reader may already feel uneasy about these admittedly ingenious solutions to the mediation and combination problems: Where do these numbers come from? Varieties of this puzzle afflict both probabilities and utilities.

4.1. Rational approaches

Bayesian cognitive science models go beyond classical decision theory in showing how probabilities can be calculated and applied to many tasks (Tenenbaum, Kemp, Griffiths, & Goodman, Reference Tenenbaum, Kemp, Griffiths and Goodman2011). First, such models specify the hypothesis space. For example, the CEO considering a new blender model might entertain three hypotheses: “We cannot engineer the blender,” “We engineer the blender but cannot successfully market it,” “The blender expansion succeeds.” The reasoner assigns prior probabilities to the hypotheses, evaluates each hypothesis's fit with the data, and combines these using Bayesian inference. For example, the CEO might assign priors of 0.4, 0.35, and 0.25 to these hypotheses. As evidence accumulates – engineers develop a prototype, marketers run focus groups – she considers the likelihood of this evidence under each hypothesis, using Bayes’ theorem to combine likelihoods with priors. Despite the simplicity of this toy example, Bayesian cognitive science models make a range of substantive and interesting claims about how people think in a vast array of contexts.

Many critiques have been written of these models (Bowers & Davis, Reference Bowers and Davis2012; Jones & Love, Reference Jones and Love2011; Marcus & Davis, Reference Marcus and Davis2013), and we do not endorse all points of criticism. From our perspective, Bayesian approaches potentially work, both in principle and practice, quite well under risk. But their inescapable limitation is the same as for classical decision theory – probabilities cannot be modeled under radical uncertainty, which can interfere with each modeling step:

• • Hypothesis space. Many problems resist an enumeration of possible outcomes. The CEO has neglected many other possibilities – competitors entering the blender market, engineers generating a prototype no better than competitors’, regulators blocking the expansion. This is due to both aleatory uncertainty (some possibilities cannot be imagined even in principle because the future is unknown – the “unknown unknowns”) and to generation limits (our finite cognitive capacity to imagine possibilities in open-ended problems).

• • Priors. It is often unclear even in principle how to rationally set priors. How did the CEO assign a 0.4 probability to her engineering team's failure? Why not 0.2, 0.35, or 0.7? Is it reasonable to use the base rate of engineering team failures given that this product has never before been designed? This is a specification limit – one cannot non-arbitrarily favor one value over another within a range of plausible values. This not only limits the psychological plausibility of these models, but can be problematic for the models themselves since the specific priors chosen sometimes drive the model's fit (Marcus & Davis, Reference Marcus and Davis2013).

• • Likelihoods. Likelihoods reflect the probability of the evidence conditional on each hypothesis. This raises three problems in realistic contexts. First, information limits: How do we know what evidence is relevant? The CEO may scan the newspaper, consumer research, and marketing reports to grasp the blender market, but will struggle to know which pieces of information bear on her hypotheses. Second, as for priors, specification limits: How do we assign probabilities non-arbitrarily? How is she supposed to estimate the likelihood of particular focus group feedback conditional on the product's future success? Third, capacity limits: With very many pieces of evidence and hypotheses, the amount of calculation rubs up against memory and attention limits.

• • Updating. Bayesian inference itself quickly approaches capacity limits as the number of hypotheses increases, since priors and likelihoods for each hypothesis must be stored and combined. This is especially problematic for the chains of inference that are common in real-world problems. The probability of successful roll-out depends on the quality of the product, the performance of the marketing team, word-of-mouth, predictions made by analysts and retailers, and countless other factors, often mutually dependent. Moreover, the problems surrounding hypothesis spaces, priors, and likelihoods compound at each step.

Rational approaches – both classical and Bayesian varieties – are nonetheless valuable. They can characterize “small-world” problems, such as risky decisions with enumerable outcomes and probabilities; sometimes provide normative benchmarks for assessing human decisions; provide valuable insights for designing artificial systems (Lake, Ullman, Tenenbaum, & Gershman, Reference Lake, Ullman, Tenenbaum and Gershman2017); and provide insight at Marr's (Reference Marr1982) computational level by characterizing the goals of a cognitive system. And although probabilistic approaches cannot capture cognition under radical uncertainty, they have inspired some of the boundedly rational approaches discussed next.

4.2. Boundedly rational approaches

Both classical and Bayesian approaches have been criticized for their lack of psychological realism, leading to several varieties of bounded rationality as amendments. Their core insight is that, although our minds are limited and prone to error, we get quite far with these limited resources: We can be rational within the bounds of our finite minds.

The dominant theoretical style in behavioral economics is surprisingly continuous with neoclassical modeling. Traditional models assume that economic agents are rational, then specify the institutional environment through which the agents interact (e.g., firms and consumers in a competitive market with priced goods) and examine the resulting equilibrium. Behavioral models use the same steps, but tweak the agents to incorporate biases or non-standard preferences, as explained in Section 2.

Like classical approaches, we believe these models produce valuable insights, particularly how small changes to the assumed psychology of economic agents qualify the results of standard models. Yet such models struggle with radical uncertainty and fuzzy evaluation. The same problems that plague classical models with probabilities apply to behavioral models with “decision weights”: Such models may capture real psychological biases in how people process probabilities, yet assume probabilities exist to be processed. This makes sense for some formal models and laboratory tasks, but not when probabilities do not exist (Section 2.2.1). Likewise, models that stuff the utility function with goodies can capture genuine trends in preferences, but create an illusion of precision when evaluation is fuzzy and options are incommensurable (Section 2.2.2).

Several important principles of bounded rationality, however, do not depend on the intelligibility of optimization:

• • Resource rationality. In coining the term “bounded rationality,” Simon (Reference Simon1957) did not view humans as capriciously irrational, but as managing the best we can given our cognitive and environmental limitations. This approach has been refined in sophisticated models of resource rationality (Lieder & Griffiths, Reference Lieder and Griffiths2020), emphasizing the rationality of simplifying strategies such as sampling (Sanborn, Griffiths, & Navarro, Reference Sanborn, Griffiths and Navarro2010). Rationing limited resources is one reason to adopt simplifications – such as narrative thinking – in the face of the calculational difficulties of uncertainty. Equally importantly, probabilistic strategies under uncertainty are not always capable of giving any answer at all.

• • Heuristics. A heuristic is a fallible-but-useful shortcut. Heuristics are often discussed in contexts where correct answers exist but algorithmic approaches are infeasible or knowledge too limited to provide an optimal answer: Some researchers (“heuristics-and-biases”) emphasize the “fallible” part of “fallible-but-useful,” and others (“fast-and-frugal”) the “but-useful” part (Gigerenzer, Reference Gigerenzer2008; Tversky & Kahneman, Reference Tversky and Kahneman1974). For our purposes, we note that heuristics also may be useful in situations where no objectively correct answer exists, yet some answers are more reasonable than others.

  A parallel debate has raged in normative and descriptive ethics. Utilitarianism emphasizes calculation (Bentham, Reference Bentham1907/1789), positing a duty to maximize social utility. Yet just as Bayesian calculations are often impossible in principle, utilitarian calculations often fail in real-world situations. Aristotle (Reference Irwin1999/350 BCE) bemoans the impossibility of a complete theory of ethics, instead urging us to cultivate habit and virtue to do the right thing in particular situations. Indeed, people distinguish between “rational” and “reasonable” behaviors (Grossmann, Eibach, Koyama, & Sahi, Reference Grossmann, Eibach, Koyama and Sahi2020), with the former characterized by optimization and abstract universalism, the latter by pragmatism and context-sensitivity (Rawls, Reference Rawls and Kelly2001; Sibley, Reference Sibley1953). This is likely why descriptive accounts of moral decision-making point to tools such as rules (Greene, Sommerville, Nystrom, Darley, & Cohen, Reference Greene, Sommerville, Nystrom, Darley and Cohen2001; Kant, Reference Kant2002/1796; Mikhail, Reference Mikhail2007), norms (Nichols, Reference Nichols2002), sacred or protected values (Baron & Spranca, Reference Baron and Spranca1997; Tetlock, Reference Tetlock2003), and character virtues (Johnson & Ahn, Reference Johnson and Ahn2021; Uhlmann, Pizarro, & Diermeier, Reference Uhlmann, Pizarro and Diermeier2015), which often act like heuristics (De Freitas & Johnson, Reference De Freitas and Johnson2018; Sunstein, Reference Sunstein2005). For moral decisions, like many everyday choices, often no clearly correct option exists, yet some are more readily justifiable.

• • Ecological rationality. A crucial point made by some researchers from boundedly rational traditions is that decision-making is adapted to real environments, so seemingly irrational behaviors observed in atypical contexts may be manifestations of more deeply rational – or at least adaptive – behaviors (Todd & Gigerenzer, Reference Todd and Gigerenzer2007). If most everyday decisions are taken under radical uncertainty, then behaviors that may be adaptive in the real world may manifest as demonstrably suboptimal decisions in the context of risky (often lab-based) contexts.

Narrative approaches to decision-making are compatible with these insights, and can be considered a species of bounded rationality – albeit, at least for CNT, one for which the appropriate benchmark is reasonableness rather than rationality.

4.3. Narrative approaches

Several researchers in both psychology and economics have argued that narratives guide decision-making.

From early days in the heuristics-and-biases tradition, causal thinking was thought to play a privileged role in judgment (Kahneman & Tversky, Reference Kahneman, Tversky, Kahneman, Slovic and Tversky1982), such as our ability to use base rates (Ajzen, Reference Ajzen1977; Krynski & Tenenbaum, Reference Krynski and Tenenbaum2007; Tversky & Kahneman, Reference Tversky, Kahneman and Fishbein1980; cf. Barbey & Sloman, Reference Barbey and Sloman2007; Gigerenzer & Hoffrage, Reference Gigerenzer and Hoffrage1995; Koehler, Reference Koehler1996). However, the first decision-making model to consider detailed cognitive mechanisms underlying narrative thought was the Story Model of Pennington and Hastie (Reference Pennington and Hastie1986, Reference Pennington and Hastie1988, Reference Pennington and Hastie1992, Reference Pennington and Hastie1993), most famously applied to juror decisions. In their model, jurors reach verdicts by constructing causal stories and assigning the story to the most appropriate verdict category (e.g., manslaughter, not-guilty). In their studies, participants generated verdicts based on realistic trial evidence. When describing their reasoning, participants overwhelmingly supported their verdicts with causal stories (describing intentions and behaviors) rather than unelaborated lists of facts, with these stories differing greatly across individuals depending on their verdict (Pennington & Hastie, Reference Pennington and Hastie1986). Manipulating the ease of constructing coherent stories (scrambling evidence order) dramatically shift participants’ verdicts (Pennington & Hastie, Reference Pennington and Hastie1988, Reference Pennington and Hastie1992). Although the Story Model's legal applications are best-known, it has also been applied to other contexts including economic decisions (Hastie & Pennington, Reference Hastie, Pennington, Connolly, Arkes and Hammond2000; Mulligan & Hastie, Reference Mulligan and Hastie2005).

Research since this seminal work has developed in two directions. In cognitive science, increasingly sophisticated theories model how people think about networks of causal relationships (Gopnik et al., Reference Gopnik, Glymour, Sobel, Schulz, Kushnir and Danks2004). For example, Sloman and Hagmayer (Reference Sloman and Hagmayer2006) argue that people conceptualize their decisions as interventions on a causal network – an idea in sympathy with CNT, wherein choice points in a narrative are opportunities to select among different imagined futures implied by the narrative. A separate but kindred line of work on the Theory of Narrative Thought (Beach, Reference Beach2010; Beach, Bissell, & Wise, Reference Beach, Bissell and Wise2016) emphasizes the pervasive role of narratives in memory and cognition, and, like CNT, highlights the importance of narratives for forecasting (Beach, Reference Beach2020).

A second direction (Abolafia, Reference Abolafia2020; Akerlof & Shiller, Reference Akerlof and Shiller2009; Akerlof & Snower, Reference Akerlof and Snower2016; Shiller, Reference Shiller2019) emphasizes the role of shared narratives in economic outcomes. Shiller argues that when shared narratives go viral, they influence expectations about the future, shaping macroeconomic activity. Shiller's view also provides a powerful role for contagious emotions, especially excitement and panic.

CNT develops these ideas in a third direction: Incorporating ideas about narrative decision-making into a broader framework that elaborates processes and mechanisms, explains how narratives and emotion combine to drive and support action, and accounts for the role of cultural evolution of narratives to render decision-making adaptive even under radical uncertainty. We see CNT as complementing rather than contradicting these perspectives, developing these approaches to the next stage in their evolution.

5.1. Narratives are structured, higher-order representations

In a structured mental representation, relations are represented among the objects it represents. For example, a feature list is an unstructured representation of a category, whereas a propositional, sentence-like representation with explicit predication (i.e., assigning attributes to specific elements and specifying relations among elements) is highly structured. Similarly, causal models are highly structured, as are representations of categories whose features are related to one another through analogy (Gentner, Reference Gentner1983).

Narratives may have been difficult to pin down in past work because they are structured, like causal models, but contain richer information that is not typically represented in causal models. Specifically, we argue that narratives can represent causal, temporal, analogical, and valence structure. Complicating things further, not all types of structure are necessarily invoked in a given narrative. Narratives are higher-order representations that flexibly include lower-order representations.

5.1.1. Causal structure

Narratives represent at least two types of causal relationships: event-causation and agent-causation. Event-causation can further be sub-divided based on the kind of event (individuals or categories) and how the events are connected (e.g., colliders, chains, or webs).

Event-causation refers to dependency relationships between either individual events (the Central Bank lowered interest rates, causing investment to increase) or event categories (lower interest rates cause increased investment). Both kinds of event-causal relations are important since narratives incorporate information both about individual events (e.g., the course of my marriage) and event categories (e.g., how relationships work generally), including analogical links between these knowledge types. We are agnostic about the representational format of event-causation, and indeed these representations may involve aspects of networks, icons, and schemas (Johnson & Ahn, Reference Johnson, Ahn and Waldmann2017). For familiarity, the diagrams we use to depict narratives are elaborated from causal network conventions (Figs. 3–6).

Figure 3.

Common causal structures in narratives. In panel A (a causal collider), multiple potential causes (A or B) could explain an event (C); a typical inference problem would be to evaluate A and B as potential explanations of observation C, which may in turn license other inferences about effects of A or B (not depicted here). In panel B (a causal chain), a sequence of causally related events (D, E, F, G) is posited; typical inference problems would be to evaluate whether the overall sequence of events is plausible, or whether an intermediate event (E) is plausible given that the other events (D, F, G) were observed. In panel C (a causal web), many event types (H–N) are thought to be related to one another, with some relationships positive and others negative, and some bidirectional; typical inference problems would be to evaluate the plausibility of individual links or to infer the value of one variable from the others. In panel D (agent-causation), an agent (Q) considers taking an action (P), based partly on reasons (O) and their judgment of the action itself (P); typical inference problems would be to predict the agent's action based on the available reasons, or infer the agent's reasons based on their actions. [Circles and squares depict events and agents, respectively; straight arrows depict causal relationships, which could be unidirectional or bidirectional, positive (default or with a “+” sign) or negative (with a “–” sign); curved, diamond-tipped arrows depict reasons. For causation among events and agents, but not event-types (panels A, B, and D), left–right orientation depicts temporal order.]

Patterns of event-causation also vary in their topology and inference patterns (Johnson, Reference Johnson2019). Three common types of causal patterns in narrative contexts are colliders, chains, and webs (Fig. 3A–C).

In a causal collider (Fig. 3A), we can observe evidence and seek an explanation for it, which may in turn generate further predictions. For example, if a central banker makes some statement, this licenses inferences about the banker's intention, which may yield predictions about the bank's future policies.

In a causal chain (Fig. 3B), a sequence of events is causally and temporally ordered. Mrs O'Leary went to milk her cow; the cow objected and kicked a lantern; the lantern started a fire; and so began the Great Chicago Fire (supposedly). A mysterious individual invents blockchain technology; it fills an important economic niche; it gains value; it becomes widely adopted. We can think about the plausibility of these sequences overall, fill in missing events from the chains, and predict what will happen next.

In a causal web (Fig. 3C), we ask how a set of variables relate to one another – an intuitive theory (Shtulman, Reference Shtulman2017). Some intuitive theories probably have innate components (Carey, Reference Carey2009a), but many decision-relevant intuitive theories are learned. For example, investors must have mental models of how macroeconomic variables such as exchange rates, inflation, unemployment, economic growth, and interest rates are linked (Leiser & Shemesh, Reference Leiser and Shemesh2018), and voters have intuitions about trade, money, and profits (Baron & Kemp, Reference Baron and Kemp2004; Bhattacharjee, Dana, & Baron, Reference Bhattacharjee, Dana and Baron2017; Johnson, Zhang, & Keil, Reference Johnson, Zhang, Keil, Rogers, Rau, Zhu and Kalish2018, Reference Johnson, Zhang and Keil2019). These intuitions likely drive much political and economic behavior, yet differ strikingly from economists’ consensus (Caplan, Reference Caplan2007; Leiser & Shemesh, Reference Leiser and Shemesh2018). This may reflect both divergences between the modern world and evolved intuitions (Boyer & Petersen, Reference Boyer and Petersen2018), and the difficulty of correctly extracting causal structure from causal systems with more than a few variables (Steyvers, Tenenbaum, Wagenmakers, & Blum, Reference Steyvers, Tenenbaum, Wagenmakers and Blum2003).

Narratives often center around the intentional actions of human agents, and such agent-causation appears to be a very different way of thinking about causation. Rather than representing events as causing one another (e.g., Alan Greenspan's forming the intention to decrease interest rates caused interest rates to decrease), people sometimes appear to think of agents as causing events directly (e.g., Alan Greenspan caused interest rates to decrease). This reflects the intuitive notion that agents have free will; that our choices, when construed as agent-causes rather than event-causes, are themselves uncaused (Hagmayer & Sloman, Reference Hagmayer and Sloman2009; Nichols & Knobe, Reference Nichols and Knobe2007). Agents can act for reasons (Malle, Reference Malle1999); they make intentional choices based on their beliefs and desires which are assumed to be rational (Gergely & Csibra, Reference Gergely and Csibra2003; Jara-Ettinger, Gweon, Schulz, & Tenenbaum, Reference Jara-Ettinger, Gweon, Schulz and Tenenbaum2016; Johnson & Rips, Reference Johnson and Rips2015). Complicating things further, reasons are often anticipations of the likely effects of one's action.

5.1.3. Analogical structure

Both the power and peril of narrative thinking compound when people perform inference not only by causal thinking within a single domain, but across different domains through analogies. Structure-mapping theory (Gentner, Reference Gentner1983) is a model of how people select analogies and use them to make inferences, emphasizing that matches in the relationships within a domain make it a good or bad analogy for another domain. Thus, analogy is especially powerful when combined with other relational systems such as causal systems (Holyoak, Lee, & Lu, Reference Holyoak, Lee and Lu2010) (Fig. 4A).

Figure 4.

Analogical, valence, and causal structure. In panel A (analogical structure), one causal chain (R₁, S₁, T₁) is analogized to another (R₂, S₂, T₂); typical inference patterns would be to reason from a known sequence (R₁, S₁, T₁) of specific events or schematized depiction of a general causal mechanism to infer the causal–temporal order of a new sequence (R₂, S₂, T₂) or to infer missing events (T₂) given that all other events are observed. In panel B (valence structure), positive event types (U, V, W) are seen as bidirectionally and positively related to each other, negative event types (X, Y, Z) are seen as bidirectionally and positively related to each other, whereas negative and positive events are seen as bidirectionally and negatively related to each other (Leiser & Aroch, Reference Leiser and Aroch2009). (Dashed lines represent analogical correspondences; white and black circles represent “good” and “bad” events or event types, respectively.)

Analogies are important to narratives for at least two reasons. First, they allow us to use familiar domains to make sense – if imperfectly – of less familiar domains. For example, people have highly impoverished mental models of central banking, but more detailed mental models of cars. In a car, stepping on the gas pedal causes more gasoline to enter the engine, increasing the car's speed. People often use this analogy for discussing and understanding central banking; the central bank prints more money, causing more money to enter the economy, causing the economy to go faster. Second, abstract and gist-like representations apply to a broader set of future situations, particularly when making decisions about the distant future (Schacter & Addis, Reference Schacter and Addis2007; Trope & Liberman, Reference Trope and Liberman2003). Thus, forming analogical links among specific past events and, ultimately, between specific events and more abstract event categories is crucial for generating generalizable knowledge. This is how our representational system incorporates some aspects of narratives’ hierarchical structure.

5.1.6. What narratives are not

Narratives are a flexible representational format, but they are not infinitely flexible. We (tentatively) suggest the following test for whether a representation is a narrative: It must (i) represent causal, temporal, analogical, or valence information, and (ii) for any of these it does not represent, it must be possible to incorporate such information.

This distinguishes narratives from several other kinds of representations, including probabilities, spatial maps, associative networks, images, categories, and logical relations: Such formats do not necessarily include any of the four structured information types. (However, elements of narratives may be linked to such other representations in memory. Indeed, this may be required for narrative simulation to generate iconic representations of imagined futures.)

Figure 5 provides additional examples of possible narratives that might underlie decision-making in the context of a global pandemic, as in one of our running examples.

Figure 5.

Possible narratives around a global pandemic. Panel A depicts one possible individual's narrative around a global pandemic, which aligns largely with the mainstream view. Infections and deaths (which are bad) are negatively related to interventions such as social distancing, masks, and vaccines, which are themselves results of government action. The government chose these actions for the reason that it would have a preventative effect on deaths. The causal links between each intervention and infection is supported by an analogy to other diseases, such as influenza (i.e., staying home from work, covering one's mouth when coughing, and vaccines all help to prevent flu infections). Panel B depicts one possible conspiratorial narrative around a pandemic. In this narrative, global elites control the government, and are acting so as to increase their profits, which can be accomplished by several channels including economic distress, population reduction, and mind control. These causal links to profitability are supported by their own analogies (e.g., the global financial crisis and subliminal messaging being ways that bankers, corporations, and other elites are thought to increase their profits), as is the idea that the government is captured by unelected elites such as lobbyists for big business. In this narrative, social distancing has little effect on the spread of disease but a strong link to intentional economic distress; masks and vaccines increase infection and death rather than preventing it. For this reason, interventions that are seen as good in the mainstream narrative (because they have a preventive relationship with death) are seen as bad in the conspiracy narrative. These hypothetical narratives will be supported by different social and informational environments, yield conflicting forecasts about the future, and motivate distinctive actions.

5.2. Narratives characterize real-world decisions under radical uncertainty

Lab experiments are ill-suited for testing the prevalence of narrative thinking in everyday decision-making. Thus, we bring linguistic and qualitative data to bear.

5.2.1. Linguistic data about macroeconomic narratives

Shiller (Reference Shiller2019) uses time-series data from Google N-Grams to track language linked to particular shared narratives. Shiller emphasizes that “viral” narratives, even if false, can affect macroeconomic events. We would add that shared narratives held true within one's own social network (rather than those held true somewhere else) may be the only way for many to make sense of complex macroeconomic causation.

The shared narrative that labor-saving machinery creates unemployment is perennial. Shiller traces it from Aristotle, through worker riots during the Industrial Revolution, to economic depressions, to present-day concerns about artificial intelligence displacing humans. (Most economists disagree: Since machines increase productivity, wages rise and labor is redeployed to higher-valued uses.) Shiller traces the frequency of “labor-saving machinery” in books from 1800 to 2008, with the term peaking during the 1870s depression and again in the lead-up to the Great Depression, with the new term “technological unemployment” reaching epidemic proportions throughout the 1930s. Plausibly, the fear produced by such narratives exacerbated the underlying problems causing the Depression. Shiller notes that this dovetails with the then-popular folk theory – never accepted by economists – that machines would produce such plentiful products that we could never consume it all, generating unemployment. Accordingly, the term “underconsumption” skyrockets during the Depression. A simplified version of this narrative can be seen in Figure 6A.

Figure 6.

Economic narratives from linguistic and interview data. Panels A–C depict simplified versions of three narratives drawn from Shiller's (Reference Shiller2019) linguistic studies of viral economic narratives and Tuckett's (Reference Tuckett2011) interview studies of money managers. In panel A, a generic causal mechanism of machinery generally leading to increased efficiency (Ma ₁ and Ef ₁) is analogized to machinery in one's particular industry leading to increased efficiency in that industry (Ma ₂ and Ef ₂). Efficiency is thought to cause unemployment (UN) directly by displacing human workers and indirectly through underconsumption (UC). Because unemployment is seen as bad, all other variables in the causal chain are inferred to be bad too. In panel B, greedy businessmen (GB) are inspired by the opportunity of World War I (W1) to increase prices (HP), which leads to inflation (Inf). A boycott (By) is thought to reduce demand (RD), which would in turn push prices back down (negative effect on HP). Since inflation and the greedy businessmen who cause it are bad, the countervailing boycott chain is perceived as good. In panel C, negative news about a company (NN) is thought to affect its stock price at an initial time (P ₁), but only the company's fundamentals (F) affect its stock price later (P ₂). Other investors (Other) are less observant and only act based on the negative news, but our investment firm (Us) is more observant and sees the fundamentals, creating a profit opportunity.

Another of Shiller's examples concerns boycotts following World War I (Fig. 6B). The US dollar experienced 100% inflation after the war, contributing to an anti-business shared narrative, with mentions of “profiteer” in newspapers peaking at the start of the subsequent depression. According to this narrative, businesses were raising prices to achieve “excess profits” during the war, explaining the inflation. (Although economists now reject this view of inflation, similar narratives were proposed by some American politicians during the inflation episode in 2022.) Protests ensued, resulting in boycotts on the theory that if consumers did not buy products beyond minimum necessities, the drop in demand would force prices back to “normalcy.” Although prices never declined to pre-war levels, deflation did indeed result in the 1920–1921 depression.

6.1. We use a suite of explanatory heuristics to evaluate narratives

Successful explanation under radical uncertainty requires strategies that circumvent probabilistic reasoning, but instead exploit other forms of structure available in a given situation. People use a variety of heuristics and strategies satisfying this criterion (e.g., Horne, Muradoglu, & Cimpian, Reference Horne, Muradoglu and Cimpian2019; Lombrozo, Reference Lombrozo2016; Zemla, Sloman, Bechlivanidis, & Lagnado, Reference Zemla, Sloman, Bechlivanidis and Lagnado2017). These strategies are useful, not because they are infallible or optimal (which may not even be meaningful under radical uncertainty), but because they do not require explicit probabilistic reasoning, yet exploit regularities that often make these strategies broadly truth-tracking.

Above, we highlighted how temporal, analogical, and valence information are incorporated into narrative representations. Heuristics exploiting such information are valuable because they capitalize on regularities we naturally attend to, such as event structure and causal mechanisms, while powerfully constraining which explanations are deemed plausible (Einhorn & Hogarth, Reference Einhorn and Hogarth1986). For example, people use temporal order (Bramley, Gerstenberg, Mayrhofer, & Lagnado, Reference Bramley, Gerstenberg, Mayrhofer and Lagnado2018), the delay between cause and effect (Buehner & May, Reference Buehner and May2002), and the part–whole structure of events (Johnson & Keil, Reference Johnson and Keil2014) to disambiguate causal directionality, determine which events are causally relevant, and assign causal responsibility. Such strategies are adaptive because they exploit regularities in the environment that are less susceptible to the problems of radical uncertainty, even if they are not infallible.

Likewise, analogies allow us to extend hard-won knowledge of one domain into another. Much of our causal knowledge appears to be stored in the format of stereotyped causal schemas (Johnson & Ahn, Reference Johnson and Ahn2015, Reference Johnson, Ahn and Waldmann2017). We can analogically reason from a known story (e.g., my cousin's marriage; the Spanish Flu) to the current situation (e.g., my marriage; COVID); or at a more abstract level, we reason from known causal mechanisms when evaluating an unfamiliar domain, as when we compare the economy to a stalled car or a bureaucratized corporation to an arthritic giant. The notion that “good begets good” (Leiser & Aroch, Reference Leiser and Aroch2009) can be thought of a highly generalized causal schema, with narratives failing to match this schema (e.g., actions that decrease inflation are likely to increase unemployment) fighting an uphill battle for plausibility. In other cases, we have domain-specific expectations (Johnston, Sheskin, Johnson, & Keil, Reference Johnston, Sheskin, Johnson and Keil2018), such as the belief that physical causation follows more linear causal pathways compared to the more web-like structure of social causation (Strickland, Silver, & Keil, Reference Strickland, Silver and Keil2016).

Other heuristics derive from causal structure itself, accounting for the anomalies above. People often substitute the vague and challenging question of an explanation's prior probability for the more straightforward question of an explanation's simplicity (Lombrozo, Reference Lombrozo2007). Yet, because simpler explanations often explain less of the data, people also use an opponent complexity heuristic to estimate an explanation's fit to the data or Bayesian likelihood (Johnson, Valenti, & Keil, Reference Johnson, Valenti and Keil2019). Likewise, people prioritize explanations that account for a wider range of observations (Read & Marcus-Newhall, Reference Read and Marcus-Newhall1993) and attempt to make inferences about evidence that has not actually been observed (Johnson, Kim, & Keil, Reference Johnson, Kim, Keil, Papafragou, Grodner, Mirman and Trueswell2016a; Johnson, Rajeev-Kumar, & Keil, Reference Johnson, Rajeev-Kumar and Keil2016b; Khemlani et al., Reference Khemlani, Sussman and Oppenheimer2011).

Explanatory heuristics are used widely across cognition – for problems such as categorization, causal attribution, theory-of-mind, stereotyping, and even some visual tasks (Johnson, Reference Johnson2016; Johnson et al., Reference Johnson, Jin, Keil, Bello, Guarini, McShane and Scassellati2014; Johnson, Merchant, & Keil, Reference Johnson, Merchant, Keil, Noelle, Dale, Warlaumont, Yoshimi, Matlock, Jennings and Maglio2015a; Johnson, Rajeev-Kumar, & Keil, Reference Johnson, Rajeev-Kumar, Keil, Noelle, Dale, Warlaumont, Yoshimi, Matlock, Jennings and Maglio2015b; Johnson et al., Reference Johnson, Kim, Keil, Papafragou, Grodner, Mirman and Trueswell2016a, Reference Johnson, Rajeev-Kumar and Keil2016b; Sussman, Khemlani, & Oppenheimer, Reference Sussman, Khemlani and Oppenheimer2014) and emerge early in development (Bonawitz & Lombrozo, Reference Bonawitz and Lombrozo2012; Cimpian & Steinberg, Reference Cimpian and Steinberg2014; Johnston, Johnson, Koven, & Keil, Reference Johnston, Johnson, Koven and Keil2017). They are also linked to action: People favor explanations that license task-relevant and high-utility actions, and which highlight stable causal relationships likely to apply across contexts (Johnson et al., Reference Johnson, Rajeev-Kumar, Keil, Noelle, Dale, Warlaumont, Yoshimi, Matlock, Jennings and Maglio2015b; Vasilyeva, Wilkenfeld, & Lombrozo, Reference Vasilyeva, Wilkenfeld and Lombrozo2017, Reference Vasilyeva, Blanchard and Lombrozo2018). Explanatory heuristics, though imperfect, aid people in circumventing specification limits and information limits (Section 2.2.1).

6.2. Explanatory fit is experienced affectively

Although these heuristics may function to push us toward more useful or probable portions of the hypothesis space, their phenomenology is often more affective than cognitive. Emotions rapidly summarize information not readily available to consciousness (Rolls, Reference Rolls2014; Rozin & Fallon, Reference Rozin and Fallon1987; Todorov, Reference Todorov2008). Because emotions have an intelligence of their own (Nussbaum, Reference Nussbaum2001), we make inferences from them (Cushman, Reference Cushman2020; Schwarz, Reference Schwarz, Sorrentino and Higgins1990) and often rely on “gut” feelings to assess situations (Klein, Reference Klein1998). This can be a broadly rational strategy despite leading to some mistakes.

In the case of explanations, we feel “satisfied” when we achieve a sense of understanding (Gopnik, Reference Gopnik1998; Lipton, Reference Lipton2004). Despite the adaptive basis of many of these heuristics, we do not feel like we are performing rational computations when using them. Instead, good explanations often are accompanied by positive emotions or even aesthetic beauty (Johnson & Steinerberger, Reference Johnson and Steinerberger2019), as when scientists and mathematicians claim to prioritize beauty in constructing their theories. Conversely, explanations that conflict with prior beliefs produce cognitive dissonance and may therefore be rejected (Festinger, Reference Festinger1962).

Although heuristics often lead to error (Kahneman, Reference Kahneman2002), they can be adaptive in solving problems given cognitive and environmental limits (Gigerenzer & Goldstein, Reference Gigerenzer and Goldstein1996; Simon, Reference Simon1955). Good thing, too: Under radical uncertainty, they often are all we have.

7.1. Imagined futures are simulated by projecting a narrative forward

The brain mechanisms involved in prospective thought about the future overlap with those used for episodic memory about the past (Schacter, Addis, & Buckner, Reference Schacter, Addis and Buckner2008) and may even be subsystems of a broader mental time travel faculty (Suddendorf & Corballis, Reference Suddendorf and Corballis1997, Reference Suddendorf and Corballis2007). This is consistent with our view that the same representations – narratives – underlie explanations of the past and simulations of the future (Aronowitz & Lombrozo, Reference Aronowitz and Lombrozo2020). Moreover, simulation can rely on step-by-step reasoning using causal mechanisms. For example, when shown a set of interlocked gears and asked which direction one gear will turn given that another gear is turned, people solve this problem by mentally turning one gear at a time (Hegarty, Reference Hegarty2004). Likewise, thoughts about how reality might be different – counterfactuals – operate over causal structures (Rips, Reference Rips2010; Sloman, Reference Sloman2005) and are central to imagination (Markman, Gavanski, Sherman, & McMullen, Reference Markman, Gavanski, Sherman and McMullen1993).

However, less work has looked at how particular features of narratives manifest in simulations or how these simulations manifest in choices. Our recent research program has studied the role of narrative-based simulation in financial decision-making.

Our strategy relies on the idea that cognitive processes have specific signatures associated with their limitations or biases. To use an example from a different area (Carey, Reference Carey2009b), a signature of analog magnitude representations is that the ability to discriminate two magnitudes is proportional to their ratio. Since discriminations for large (but not small) numbers have this property, this implies that representations of large numbers are analog.

Analogously, we look for signature limitations associated with using narrative representations for predictions and decisions under uncertainty. Since narratives can incorporate causal, valence, temporal, and analogical structure (Section 5.1), we designed experiments to examine whether introducing these structural features into forecasting problems produces signature biases relative to standard financial theory. Although any one study does not necessarily implicate narratives, their combination triangulates on the conclusion that people project narratives to simulate the future.

7.2. Imagined futures are simulated one at a time

We have focused so far on how narratives circumvent limitations on probabilistic thinking. Yet narrative thinking has limits of its own: Once we have adopted a particular narrative, we are often blind to alternative possibilities. We simulate narratives one at a time.

For example, different government policies lead to different predictions about market prices. If the central bank raises interest rates, this is likely to depress market prices. But central bankers often speak opaquely. If an investor assigns a 75% chance to the story that the banker plans to raise rates but a 25% chance to the story that the banker plans to leave them alone, does she account for both possibilities or rely on just one? A Bayesian would calculate the likely effect on markets if each story is true, taking a weighted average when estimating future prices. But an investor who “digitizes” and treats these stories as either certainly true or false would “round up” the 75% chance to 100% and “round down” the 25% chance to 0%. Only the dominant narrative resulting from explanatory reasoning would be retained for downstream computations such as forecasting.

We found that investors are not Bayesians, instead digitizing (Johnson & Hill, Reference Johnson, Hill, Gunzelmann, Howes, Tenbrink and Davelaar2017). In one study, participants were given information leading them to think that one government policy was likelier than another (in one variation, they were even given these probabilities directly). Comparing conditions identical except the effect of the more-likely (75% chance) policy, there was a large difference in predictions. But comparing conditions identical except the effect of the less-likely (25% chance) policy, predictions do not differ at all. Investors take account of the implications of more-likely narratives, but ignore entirely the implications of less-likely narratives: They adopt a single narrative as true, treating it as certain rather than probable.

Digitization is a broad feature of cognition; similar effects have been found in causal reasoning (Doherty, Chadwick, Caravan, Barr, & Mynatt, Reference Doherty, Chadwick, Caravan, Barr and Mynatt1996; Fernbach, Darlow, & Sloman, Reference Fernbach, Darlow and Sloman2010; Johnson, Merchant, & Keil, Reference Johnson, Merchant and Keil2020) and categorization (Lagnado & Shanks, Reference Lagnado and Shanks2003; Murphy & Ross, Reference Murphy and Ross1994). Yet it has boundary conditions: People do reason across multiple hypotheses in cases where one of the hypotheses invokes a moral violation (Johnson, Murphy, Rodrigues, & Keil, Reference Johnson, Murphy, Rodrigues, Keil, Goel, Seifert and Freksa2019) or danger (Zhu & Murphy, Reference Zhu and Murphy2013), and expertise in a domain may promote multiple-hypothesis use (Hayes & Chen, Reference Hayes and Chen2008).

Despite boundary conditions, simulations produce, by default, a single imagined future. Electrons may exist as probability clouds rather than in one definite state. But stories resist Heisenberg's principle – they take only one state at a time.

8.1. Affective evaluations of imagined futures motivate choices

We feel emotions in response not only to the present situation, but to situations we imagine (Loewenstein, Weber, Hsee, & Welch, Reference Loewenstein, Weber, Hsee and Welch2001; Richard, van der Pligt, & de Vries, Reference Richard, van der Pligt and de Vries1996). This is why we experience emotions when understanding literature (Mar, Oatley, Dikic, & Mullin, Reference Mar, Oatley, Dikic and Mullin2010) or empathizing with others (Mitchell, Banaji, & Macrae, Reference Mitchell, Banaji and Macrae2005). CNT accords a central role to the emotional reactions we experience in response to imagined futures generated from narratives. These emotional reactions within a narrative, by motivating approach and avoidance behaviors, drive action in the real world.

Some emotions are inherently future-oriented: If one feels excited (or anxious) about a potential future, one acts to approach (or avoid) that future. But even past-oriented emotions, such as regret (Loomes & Sugden, Reference Loomes and Sugden1982), can influence our choices through simulations of how we would feel. For example, people anticipate more regret over not playing in postcode lotteries (where non-players can learn if they would have won) versus traditional lotteries, which motivate participation (Zeelenberg & Pieters, Reference Zeelenberg and Pieters2004). Many other anticipated emotions are known to guide choices, including guilt, sadness, anger (Baron, Reference Baron1992); pleasure (Wilson & Gilbert, Reference Wilson and Gilbert2005); dread and savoring (Dawson & Johnson, Reference Dawson and Johnson2021; Loewenstein, Reference Loewenstein1987); and envy (Loewenstein, Thompson, & Bazerman, Reference Loewenstein, Thompson and Bazerman1989). Emotions mediate between our predictions of the future and decisions to approach or avoid that future, coloring narratives with emotion (Beach, Reference Beach1998; Bruner, Reference Bruner1986).

8.2. Imagined futures can be appraised on default or ad hoc dimensions

The fuzzy evaluation problem (Section 2.2.2) results from the challenges of summarizing incommensurable attributes as a single utility, especially when our values may change over time. CNT proposes two computational simplification strategies that the affective system can use to address this problem – a default, bottom-up strategy and an ad hoc, top-down strategy.

This is analogous to the distinction between natural categories and ad hoc categories. Natural categories – such as BIRD, TABLE, and MOUNTAIN – roughly capture regularities in the external world (Rosch, Mervis, Gray, Johnson, & Boyes-Braem, Reference Rosch, Mervis, Gray, Johnson and Boyes-Braem1976); by default, objects are classified bottom-up, automatically and effortlessly, into natural categories (Greene & Fei-Fei, Reference Greene and Fei-Fei2014). In contrast, ad hoc categories (Barsalou, Reference Barsalou1983) – such as THINGS TO SELL IN A GARAGE SALE and WAYS TO ESCAPE THE MAFIA – are constructed on the fly to achieve specific goals, using effortful, top-down processes. Whereas bottom-up classification into natural categories proceeds by default and relies on predetermined dimensions, top-down classification into ad hoc categories requires effort and relies on spontaneously determined, goal-derived dimensions.

Analogously, in line with the appraisal tendency framework, one strategy for evaluating imagined futures relies on a default set of dimensions mirroring those for evaluating actually present situations (e.g., controllability, certainty, pleasantness), which determine which emotion is felt. That emotion, in turn, motivates action. Because these dimensions are thought to be evaluated automatically with minimal effort (Lazarus, Reference Lazarus1991), this default appraisal strategy is an appealing solution to the fuzzy evaluation problem. Specific emotions are felt in response to qualitative appraisals of predetermined dimensions. Because the dimensions are predetermined, the computational problem of identifying dimensions is avoided; because the appraisals are qualitative, the need to trade off these dimensions is circumvented. Moreover, although particular preferences may well change over time, our basic emotional architecture does not. Thus, the problems of incommensurable attributes and non-stationary values are averted.

However, these default dimensions often do not suffice when we have specific goals, leading to a second, ad hoc strategy based on the decision-maker's goal hierarchy. A decision-maker's attention will be deployed according to the active goal(s) at the time of decision-making (Van Osselaer & Janiszewski, Reference Van Osselaer and Janiszewski2012). Narratives can be used to generate imagined futures, on an ad hoc basis, that are evaluable with respect to these goals. The compatibility of those imagined futures with those goals produces approach and avoidance emotions that motivate action (Elliot, Reference Elliot2006; Oatley & Johnson-Laird, Reference Oatley and Johnson-Laird1987).

This ad hoc route depends on two claims. First, we assume that although we may have many goals, a small subset are typically active at once, because goals are triggered context-dependently (Panksepp, Reference Panksepp1998; Tuckett, Reference Tuckett, Kirman and Teschiin press) and organized hierarchically. For example, when basic physiological needs are not met, these are likely to supersede less immediately essential needs such as social belonging (Lavoie, Reference Lavoie1994; Maslow, Reference Maslow1943). This is not only helpful for survival, but also a crucial computational simplification: Goal hierarchies allow us to evaluate imagined futures over a much smaller number of dimensions. This is how ad hoc appraisals, like default appraisals, help to resolve the fuzzy evaluation puzzle. This insight also casts new light on multi-attribute choice strategies (Payne, Bettman, & Johnson, Reference Payne, Bettman and Johnson1988). A couple facing divorce faces a dizzying array of potential attributes. Yet, for many such couples, their children's well-being is paramount. If this dimension does not prove decisive, they may move on to other concerns, such as their financial well-being or sexual satisfaction. The hierarchical organization of goals can explain why particular situations call for particular decision rules, such as lexicographic rules (using a single attribute) or elimination-by-aspects (eliminating options beneath a minimum criterion on a key attribute, then iteratively considering other dimensions; Tversky, Reference Tversky1972).

Second, we assume that we can generate imagined futures containing details relevant to evaluating the required dimensions. For example, our married couple might first imagine their future with respect to their child's well-being, then elaborate this image to consider their romantic prospects, then their finances. Although these different imagined aspects may well emanate from a shared narrative of the couple's married life, it is unlikely that this complete conception of their future would emerge fully formed, but instead must be simulated one piece at a time. Although we do not know of any research directly examining this proposed ability in the context of narratives, evidence about other domains suggests that this is possible; for example, people can fluidly reclassify objects dependent on goals (Barsalou, Reference Barsalou1983) and can manipulate their mental images dependent on queries (Kosslyn, Reference Kosslyn1975).

8.3. Emotions are used to manage decisions extended over time

Hamlet's uncertainty paralyzed him for three acts; by the fifth act, it was too late. Hamlet learned the hard way that strong, conflicting arguments produce ambivalence that can stop action in its tracks (Armitage & Conner, Reference Armitage and Conner2000; Festinger, Reference Festinger1962; Rosner et al., Reference Rosner, Basieva, Barque-Duran, Glöckner, von Helversen, Khrennikov and Pothos2022; Rucker, Tormala, Pety, & Briñol, Reference Rucker, Tormala, Pety and Briñol2014; Smelser, Reference Smelser1998). Many a couple and many an investor have talked themselves in circles while romantic and financial opportunities slipped away; committing to one distinct course of action often yields better fortunes, even if one can never be certain a choice is truly “right.” Moreover, high-stakes decisions often are extended through time, requiring commitment. Failure in this respect leads many novice investors to overtrade and defray their gains through transaction costs (Barber & Odean, Reference Barber and Odean2000). Conviction bears dividends.

Yet conviction also bears risks. Blindly following a plan, while ignoring new information, is equally a recipe for disaster as Hamletian paralysis. Complex problems such as the COVID pandemic, Putin's Ukraine invasion, or climate change require different approaches as events or our knowledge evolve. Emotions are instrumental in the inter-related processes of conviction management – gaining conviction (acting in the face of ambivalence), maintaining conviction (committing to a sustained course of action), and moderating conviction (taking account of new evidence and potentially changing course as the situation changes). To manage conviction is to manage our emotional attachments to a person, object, or course of action. A lack of emotional attachment against the temporary vicissitudes of fortune yields indecision, yet an inability to reappraise a truly changing situation can yield calamity.

Cognition and affect are intertwined in generating conviction. Conviction-generating narratives integrate information about the past and expectations about the future to emotionally support a course of action. Experiments have probed this process. In a purely cognitive model, confidence in a decision is proportional to the strength of the arguments in its favor; in a purely affective model, confidence in a decision is proportional to its propensity to trigger approach emotions. Instead, cognition and emotion work together (Bilovich, Johnson, & Tuckett, Reference Bilovich, Johnson and Tuckett2020). When situations trigger approach emotions, investors find favorable arguments more relevant, with the converse for avoidance emotions. Perceived relevance in turn influences investors’ choices. Although emotions profoundly affect choice, they do so by influencing intermediate cognitive processing.

This interplay between cognition and affect is also illustrated by the conviction-generating strategies cited by the investment managers in Tuckett's (Reference Tuckett2011) interview study (Chong & Tuckett, Reference Chong and Tuckett2015). Numerous respondents (90%) referred to one or more “attractor” narratives, producing excitement over an investment due to an exceptional opportunity for gain. Attractor narratives typically cite either the investment's intrinsic properties (e.g., exceptional products) or the investor's own special skills (e.g., exceptional insight). A similar proportion of respondents (88%) referred to one or more “doubt-repelling” narratives, reducing anxiety over an investment. Doubt-repelling narratives typically raise and then counter a potential concern, placing bounds on either uncertainty (e.g., competent managers) or downside surprise (e.g., solid fundamentals).

Uncertainty can undermine conviction, but so can excessive certainty in dynamic situations. At a single time-point, people are more confident in investments described as having a specific, predictable return (8%) rather than falling within a range (3–13%) (Batteux, Bilovich, Johnson, & Tuckett, Reference Batteux, Khon, Bilovich, Johnson and Tuckettin press; Du & Budescu, Reference Du and Budescu2005). However, building conviction by masking uncertainty is not sustainable: Once point forecasts are shown to be unreliable – an inevitable event under uncertainty – trust is reduced in the forecaster (Batteux, Bilovich, Johnson, & Tuckett, Reference Batteux, Bilovich, Johnson and Tuckett2021a). This has implications for risk communication: When uncertainty is communicated in vaccine announcements, trust in vaccines is buffered against subsequent negative outcomes (Batteux, Bilovich, Johnson, & Tuckett, Reference Batteux, Bilovich, Johnson and Tuckett2021b).

Conviction is not good or bad in itself. It is needed to overcome ambivalence and sustain commitment, but is only adaptive if it does not preclude learning. When emotions are regulated well (Gross, Reference Gross1998), conviction buffers against the vicissitudes of the unfolding situation, but can be moderated. In such an integrated state, we can be sensitive to new information and adjust decisions in an orderly way; one adopts a particular narrative but acknowledges the possibility of error and stays attuned to evidence for competing narratives. In contrast, in a divided state, ambivalence is hidden by attentional neglect of information inconsistent with the preferred narrative. Whereas new information in integrated states can trigger curiosity and evidence integration, incongruent information is rejected in a divided state (Tuckett, Reference Tuckett2011; Tuckett & Taffler, Reference Tuckett and Taffler2008). This is why ambivalence has been linked both to maladaptive responses, such as confirmation bias and behavioral paralysis, and adaptive responses, such as broader attention and willingness to consider multiple perspectives (Rothman, Pratt, Rees, & Vogus, Reference Rothman, Pratt, Rees and Vogus2017). Decision-makers who experience balanced emotions are likelier to rely on “wise reasoning” strategies, such as epistemic humility and integrating diverse perspectives (Grossmann, Oakes, & Santos, Reference Grossmann, Oakes and Santos2019). Integrated conviction management is adaptive because it permits decision-makers to accept a narrative as provisionally true and act accordingly – a crucial characteristic when there is a cost to changing course – while accumulating evidence in the background, changing course when clearly merited.

Integrated conviction management is closely linked with one way that feedback loops help decisions to become adaptive over time. Although under radical uncertainty we often have no choice but to make some decision without a clear sense of whether it is the best option, we can accumulate evidence about what does and does not work. Thus, acting on one narrative can yield information that is then used to reappraise that same narrative, potentially leading to a shift in narrative and decision in an iterative manner. Extended decisions can often be treated as a series of experiments, providing information about what does and does not work (Fenton O'Creevy & Tuckett, Reference Fenton O'Creevy and Tuckett2022).

9.1. Shared narratives facilitate social coordination

Decisions are socially embedded in part through their shared consequences – when decisions are taken collectively, as in political decision-making, or when one individual's decision affects others in their social group. Socially coordinated decisions can generate more value than the sum of their individual components (Chwe, Reference Chwe2001). However, coordination is challenging, due to both divergent interests and divergent information. Shared narratives help to coordinate both interests and information.

Reputation-tracking is key to aligning individual incentives with collective interest (Rand & Nowak, Reference Rand and Nowak2013; Tennie, Frith, & Frith, Reference Tennie, Frith and Frith2010). People are motivated to evaluate others’ reputation based on their actions, even actions only affecting third parties. For instance, people evaluate others’ moral character based on prosocial actions such as donation (Glazer & Konrad, Reference Glazer and Konrad1996; Johnson, Reference Johnson2020), volunteering (Johnson & Park, Reference Johnson and Park2021), and eco-friendly actions (Griskevicius, Tybur, & van den Bergh, Reference Griskevicius, Tybur and van den Bergh2010), generating incentives for apparent altruism. Conversely, bad reputations are costly: People sacrifice resources to punish free-riders (Jordan, Hoffman, Bloom, & Rand, Reference Jordan, Hoffman, Bloom and Rand2016). Because we are aware that others are tracking our reputations, we are motivated to take actions bringing reputational benefits and avoid actions bringing reputational harm.

An important means for reputation management is how we justify choices to others (Lerner & Tetlock, Reference Lerner and Tetlock1999; Mercier & Sperber, Reference Mercier and Sperber2017). Narratives often play this justificatory role, maintaining reputation in the face of disagreement and coordinating group activity when other stakeholders must adopt the same decision. For instance, stories shared by the Agta, hunter–gatherers in the Philippines, express messages promoting cooperation, allowing skilled storytellers to achieve greater cooperation (Smith et al., Reference Smith, Schlaepfer, Major, Dyble, Page, Thompson and Migliano2017). Closer to home, scientists often debate what “story” they will sell to readers and reviewers. Why are narratives so effective for reputation maintenance?

First, narratives contain causal structure that can generate reasons justifying a position. For example, when deciding which parent should be awarded custody of a child, people favor the parent with both more extreme positive (above-average income) and negative attributes (work-related travel) over one with more neutral attributes (typical income and working hours), since the former gives more positive reasons favoring custody. But when asked instead who should be denied custody, people again choose the more extreme parent because there are also more negative reasons against custody (Shafir, Simonson, & Tversky, Reference Shafir, Simonson and Tversky1993). More extreme attributes support more causally potent explanations, generating both supporting and opposing narratives.

Second, narratives can not only justify decisions after the fact but to persuade others to adopt our perspective (Krause & Rucker, Reference Krause and Rucker2020). Arguments communicated with a narrative are often more persuasive than those communicated with facts alone (Chang, Reference Chang2008; De Wit, Das, & Vet, Reference De Wit, Das and Vet2008; Shen, Ahern, & Baker, Reference Shen, Ahern and Baker2014), in part because narratives are readily understood (Section 9.2). Narratives induce emotional engagement, mental imagery, and attention, creating “narrative transportation” that can lead reasoners to believe elements of the story (Adaval & Wyer, Reference Adaval and Wyer1998; Escalas, Reference Escalas2004; Green & Brock, Reference Green and Brock2000; Hamby, Brinberg, & Daniloski, Reference Hamby, Brinberg and Daniloski2017; Van Laer, de Ruyter, Visconti, & Wetzels, Reference Van Laer, de Ruyter, Visconti and Wetzels2014). Moreover, the broader narratives espoused by a communicator, such as moral and political worldviews, can lend additional credence to their claims (Johnson, Rodrigues, & Tuckett, Reference Johnson, Rodrigues and Tuckett2021; Marks, Copland, Loh, Sunstein, & Sharto, Reference Marks, Copland, Loh, Sunstein and Sharto2019). Persuasion is crucial in coordination because it allows a group to have the same narrative in their heads, making narratives a part of our collective or transactive memory (Boyd, Reference Boyd2009; Chwe, Reference Chwe2001; Hirst, Yamashiro, & Coman, Reference Hirst, Yamashiro and Coman2018; Wegner, Reference Wegner, Mullen and Goethals1987) and providing a shared plan for coordinated action.

9.2. Shared narratives shape social learning and evolve

Decision-making is also socially embedded through the informational environment. Human knowledge arises largely through our cumulative cultural heritage (Boyd, Richerson, & Henrich, Reference Boyd, Richerson and Henrich2011; Henrich, Reference Henrich2018). Indeed, because it is so often the ability to access external knowledge when needed that is crucial for decision-making rather than our internal knowledge itself, we often confuse knowledge inside and outside our heads (Sloman & Fernbach, Reference Sloman and Fernbach2017).

Communication of narratives is a crucial way we learn beyond immediate experience (Boyd, Reference Boyd2018), with some even suggesting that the adaptive advantage of sharing narratives is the main reason that language evolved (Donald, Reference Donald1991). However, we do not suggest that narratives in the full form described in Section 5 migrate wholesale from one mind to another. The knowledge that is transferred from one mind and stored in another is relatively skeletal, or even a placeholder (Rozenblit & Keil, Reference Rozenblit and Keil2002) as when we store the source of a piece of information rather than the information itself (Sloman & Fernbach, Reference Sloman and Fernbach2017; Sparrow, Liu, & Wegner, Reference Sparrow, Liu and Wegner2011). Instead of full narrative representations, we assume instead that primitive elements – narrative fragments such as basic causal schemas, memorable analogies, and emotional color – are the key narrative elements that are shared and shape social learning. These narrative fragments, communicated consistently within a social group, give rise to a set of elements that are common among the narratives represented within those group members’ individual minds – what we are calling a shared narrative.

Several approaches seek to model how ideas spread and evolve (Boyd & Richerson, Reference Boyd and Richerson1985; Dawkins, Reference Dawkins1976; Sperber, Reference Sperber1996). A common insight is that ideas spread when they pass through two sets of cognitive filters: Constraints on encoding (attention, memory, and trust) and constraints on communication (motivation and ability to share). But for a culturally transmitted idea to not only spread but be acted on, that idea must pass through a third filter – constraints on action. The idea must be perceived as actionable and produce motivation to act. Narratives can pass through all three filters – encoding, communication, and action – and therefore are likely to socially propagate.

First, narratives are easy to remember. People often represent information as scripts, or stereotyped sequences of events (Schank & Abelson, Reference Schank and Abelson1977). Because people so naturally represent information using causal–temporal structure, people are far better at remembering information organized as stories (Bartlett, Reference Bartlett1932; Kintsch, Mandel, & Kozminsky, Reference Kintsch, Mandel and Kozminsky1977; Thorndyke, Reference Thorndyke1977). Humans’ remarkable ability to remember information encoded as stories is perhaps most impressively attested by oral traditions, such as the transmission of the Greek epics and Hindu and Buddhist historical texts over centuries purely through word-of-mouth (Rubin, Reference Rubin1995).

Second, people like to talk about narratives, making them susceptible to spreading through word-of-mouth (Berger, Reference Berger2013). This may be because narratives are well-suited to balancing novelty against comprehensibility (Berlyne, Reference Berlyne1960; Silvia, Reference Silvia2008), with the most contagious narratives including a small number of novel concepts against a larger background of familiar concepts (Norenzayan, Atran, Faulkner, & Schaller, Reference Norenzayan, Atran, Faulkner and Schaller2006). Narratives can convey new information in digestible form because they match our default causal–temporal representations of events (Schank & Abelson, Reference Schank and Abelson1977) and focus on the behavior of human actors, commandeering our natural tendency toward gossip (Dunbar, Reference Dunbar1996). Moreover, as discussed above, narratives are highly persuasive, and therefore commonly used when trying to convince others.

Finally, narratives lend themselves to action. Because narratives have a causal–temporal organization, and are often organized around the actions of individuals (Mandler & Johnson, Reference Mandler and Johnson1977), they provide a ready template for intervening on the world. Indeed, causal knowledge is crucial precisely because it can be used to bring about desired outcomes (Woodward, Reference Woodward2003). Culturally acquired knowledge of physical causation is embedded in physical tools, which can manipulate the physical world. Analogously, culturally acquired knowledge of social causation is embedded in narratives, which serve as templates for manipulating the social world. We suggest that narratives which lead to effective actions are particularly likely to survive this filter, just as effective institutions are likelier to survive social evolution (Hayek, Reference Hayek1958). Therefore, narratives that have survived this crucible of cultural evolution potentially lead to adaptive decision-making even under radical uncertainty.

Because narratives live and die by cultural evolution (Boyd & Richerson, Reference Boyd and Richerson1985; Henrich, Reference Henrich2018), their propagation depends on their social, economic, and physical environment. Narratives surrounding masks at the start of the COVID-19 differed profoundly across countries (Hahn & Bhadun, Reference Hahn and Bhadun2021), due both to social norms and differing experience with infectious disease. Henrich's (Reference Henrich2020) account of how the West became prosperous suggests that narratives originated by the Catholic Church altered norms around cousin marriage, generating new family structures and patterns of cooperation that led to markets and science.

9.3. Shared narratives propagate through social networks

If shared narratives facilitate coordination and learning, then as they shift over time and propagate through social networks, they should be tied to large-scale outcomes. Since economic actors’ decisions are driven by narratives (Section 5.2.2), changes in the emotional content of socially available narratives may shift attitudes toward risk. If so, then measures of approach and avoidance emotions in economic narratives should predict the direction of economic aggregates – output, employment, GDP growth – that depend on investment. Nyman, Kapadia, and Tuckett (Reference Nyman, Kapadia and Tuckett2021) studied this claim using text-mining techniques on internal Bank of England commentary (2000–2010), broker research reports (2010–2013), and Reuters news articles (1996–2014).

First, relative sentiment is a leading indicator of economic volatility and consumer sentiment. This reflects the idea that a preponderance of approach over avoidance emotions is needed to produce conviction to invest (Keynes, Reference Keynes1936). For each document, the proportion of words signaling approach (e.g., “excited,” “ideal”) and avoidance words (“threatening,” “eroding”) was calculated, with the difference between these indices constituting relative sentiment. Shocks to relative sentiment in the UK had negative effects on industrial production, employment, and the stock market, with these impacts lasting for nearly 20 months (Nyman et al., Reference Nyman, Kapadia and Tuckett2021). Tuckett and Nyman (Reference Tuckett and Nyman2018) also found that relative sentiment also predicted changes in investment and employment in the UK, US, and Canada more than 12 months out (Tuckett, Reference Tuckett2017). For example, a plot of relative sentiment against major events in the lead-up to the global financial crisis shows a precipitous decline in relative sentiment in the year leading up to the failure of Bear Sterns in March 2008. A similar analysis using 1920s data from the Wall Street Journal found that sentiment shocks, beyond economic fundamentals, impacted production and stock values leading up to the Great Depression (Kabiri, James, Landon-Lane, Tuckett, & Nyman, Reference Kabiri, James, Landon-Lane, Tuckett and Nyman2023); sentiment likewise appears to account for the slow recovery from the 2008 recession (Carlin & Soskice, Reference Carlin and Soskice2018).

Second, excessive homogeneity around narratives portends trouble. Nyman et al. (Reference Nyman, Kapadia and Tuckett2021) used topic modeling to assign each article in the text database to a particular narrative and compute the degree of narrative topic consensus at each time point. Prior to the crisis, homogeneity increased around narratives high in approach emotions (excitement) and lacking avoidance emotions (anxiety), which could have been a potential warning sign of impending financial system distress (Nyman et al., Reference Nyman, Kapadia and Tuckett2021). This supports the idea of groupfeel or emotional conformity as a driver of booms and busts and the notion that integrated emotional states that manage ambivalence are better-suited to stable decision-making, compared to divided states that ignore discordant information (Tuckett & Taffler, Reference Tuckett and Taffler2008).

Macroeconomic crises are necessarily rare and atypical, so no method will reveal definitive answers about their causes. But these techniques can also be used at a more micro-scale. For example, Tuckett, Smith, and Nyman (Reference Tuckett, Smith and Nyman2014) studied relative sentiment in news articles about Fannie Mae. From 2005 to mid-2007, sentiment became increasingly exuberant, along with Fannie Mae's share price, and unmoored from economic realities reflected in the Case-Shiller Housing Price index. Such states can result from the fetishization of some “phantastic object” (Tuckett & Taffler, Reference Tuckett and Taffler2008) – in this case, mortgage securitization. A similar analysis of Enron's internal emails in 2000–2002 revealed comparable emotional–narrative dynamics of build-up and collapse surrounding the deregulation of the California energy market and Enron's impending (ill-fated) entry into broadband. Overall, the confluence of macro- and micro-level analyses converges to suggest that emotional–narrative sentiment spreads through social networks and may causally influence economic outcomes.