The Future of Decisions From Experience: Connecting Real-World Decision Problems to Cognitive Processes

Attention

In real-world decision problems, the decider usually faces a multitude of possible options and experiences. Under these circumstances, attentional processes are likely to mediate which options will be considered and which experiences will be heeded. In contrast, experiments in DfE usually restrict choices to only two options and implicitly assume that participants pay full attention to all available experiences. Yet even under these simplified circumstances, experimental manipulations, eye tracking, and cognitive modeling have shown that attention processes can affect preferential choices in the sampling protocol (Ashby & Rakow, 2016; Glickman et al., 2018; Glöckner et al., 2016; Vanunu et al., 2021; Zilker & Pachur, 2021). In particular, the saliency of an experience in terms of its significance or visual features can capture the attention of the decider and thereby influence behavior. Experiments and choice modeling have shown that relatively high numbers in a stream of experiences receive more weight in subsequent decisions (Spitzer et al., 2017; Vanunu et al., 2020). This in turn can lead to choice tendencies for high-variance options (Tsetsos et al., 2012). Moreover, when increasing the visual saliency of numbers by using a dot or a different color, participants’ choices were consistent with overweighting these salient numbers (Glickman et al., 2018; Kunar et al., 2017).

Few studies have explored instances of structural complexity, for example, by increasing the number of available choice options in experiments. In these situations, participants choose risky options more frequently (Hills et al., 2013; Noguchi & Hills, 2016). This effect can be explained by participants paying more attention to the highest outcomes, which come from the high-variance options and makes them thus appear more favorably. Beyond this effect, more research in DfE should explore how the number of available options or experiences affect decision and search strategies. This expansion could extend results from DfD by finding that people can switch strategies—that is,changing which attributes they heed, how many of them to take into account, and which order to employ in examining them (Fiedler & Glöckner, 2012; Krajbich et al., 2010; Meissner et al., 2020; Mullett & Stewart, 2016; Payne et al., 1988).

Furthermore, attention can also be influenced by additional cues in the environment. When investors check their portfolios to learn about the performance of their assets, they look more often when the market goes up, which in turn leads to more trading (Quispe-Torreblanca et al., 2022). Relatedly, prices of certain assets have been shown to be overvalued because these assets have received a lot of attention in social media (Bali et al., 2021). These findings have the potential to relate observational real-world data to experimental research in DfE and hence to extend research questions by examining factors that make people pay attention to experiences more broadly.

Numeric cognition

Many real-world decision problems, most notably in finance, require decision makers to deal with substantial numeric feedback—for example, in terms of returns on investment. Often, researchers assume that decision makers have perfect perception and representation of this numeric information. Consequently, in DfE experiments behavior is usually assessed with respect to the characteristics of the theoretical underlying distribution of the samples or with characteristics of the actual samples (e.g., the sample mean). People, however, perceive and represent numeric information imprecisely and in a biased fashion, which then can influence preferential choices (Khaw et al., 2021; Schley & Peters, 2014; Woodford, 2020). In numeric cognition, numerosity representation is often described as a compressed mental number line with a logarithmic function (Dehaene, 2003; Dehaene et al., 2008; Longo & Lourenco, 2007). Though this line of research deals with processes relevant to choosing and learning from numeric experience, it is not usually connected with DfE (e.g., Cheyette & Piantadosi, 2019; Whalen et al., 1999).

Numerosity information can be provided symbolically (e.g., Arabic numerals) or nonsymbolically (e.g., bars or dots representing quantity). Large amounts of numerosity information must often be perceived, weighted, and integrated in DfE. Thus, even when numeric information is available symbolically, the representation and integration of these numeric values may become imprecise and compressed because of capacity limitations (Khaw et al., 2021). In domains such as healthy food choices, however, experiences usually consist of nonsymbolic magnitudes, such as the amount of food one is served in a restaurant. These representations also provide a link to the animal-cognition literature, in which nonsymbolic magnitudes are the default presentation format (Pisklak et al., 2019; Shafir et al., 2008). Nonsymbolic numerosity information could lead to less precise representations than symbolic numerosity information. Thus, modeling number perception explicitly and understanding how numeric cognition might differ between symbolic and nonsymbolic presentations (Duffy et al., 2021; Garcia et al., 2022; Zeigenfuse et al., 2014) can help us to better understand preferential behavior from experiences.

A limitation of examining choices in experimental DfE research is that observed choices can only reveal ordinal relative judgments; they do not provide information about the absolute values participants assign to individual options. In a financial-investment context, choice problems can be framed differently, for example, when deciding at what price to buy or sell a given asset. This problem requires a more precise estimate of the value of the asset (e.g., Golan & Ert, 2015; Pachur & Scheibehenne, 2017). Preceding the renewed interest in DfE, a long tradition in psychology examined numerosity judgments about the mean of number sequences. These studies usually concluded that people were accurate, intuitive statisticians (Peterson & Beach, 1967). More recently, however, mean or sum estimates from experienced samples have been found to be slightly downward biased (Brezis et al., 2015; McGowan et al., 2022; Scheibehenne, 2019). This difference can be explained by the fact that earlier studies had smaller sample sizes and focused more on noise rather than on systematic deviations from the true sample mean. Directly comparing mean estimates and preferential valuations of number streams suggests that answers deviating from the sample mean could be partly due to a compressed mental number line in the perception of numbers, rather than to risk or skewness preferences (Olschewski, Newell, et al., 2021).

Finally, the context in which people perceive numerosity information matters in DfD: For example, the range and the distribution of numeric values shape choice consistency and risk taking (Frydman & Jin, 2022; Stewart et al., 2015). Some recent results point toward the impact of context when experiencing numeric feedback (Bavard et al., 2021; Hayes & Wedell, 2023; Madan et al., 2021; Palminteri & Lebreton, 2021; Prat-Carrabin & Woodford, 2022). Hence, taking the context into account—such as the range of expected or experienced outcomes—can lead to a better understanding of individual and societal choice patterns (e.g., selling decisions regarding assets from idiosyncratic portfolios; Antoniou et al., 2021).

When experiences are not directly consequential, such as when observing other people, an important question is when to stop sampling information and make a final decision (Glickman et al., 2022). This problem is experimentally examined in the free-sampling version of the sampling protocol (Hills & Hertwig, 2010). This protocol has also been used as a simulation tool to educate retail investors (Bradbury et al., 2015; Kaufmann et al., 2013). Several studies concluded that in this protocol, participants sample too little information when sampling bears no monetary costs (Hertwig et al., 2004; Wulff et al., 2018). These analyses, however, usually did not take the costs of information acquisition (in terms of opportunity costs or cognitive effort) into account (Fiedler et al., 2021; Vul et al., 2014). Furthermore, they also did not measure the subjective value of additional information, which could be low if experiences are perceived imprecisely (Olschewski & Scheibehenne, 2023). Thus, future research could improve understanding of search processes by measuring additional variables such as confidence, precision of number representation, or cognitive effort (Xiang et al., 2021).

One way to measure the subjective value of additional information is to manipulate the number of sampled experiences exogenously. Studies with this design showed that participants in choice tasks became more accurate in picking the higher expected-value option with a larger sample size (Tsetsos et al., 2012). In another study, estimation accuracy decreased from four to eight samples, but increased again from eight to sixteen samples (Brezis et al., 2015). The latter finding was explained by a dual-systems approach, in which a few samples were integrated analytically but more samples were processed according to the imprecise number system. Thus, additional information might not always be valuable with respect to a more veridical representation of a choice problem. Cognitive processes (such as the representation of numerosity) and conscious strategies to integrate experiences can affect the value of information and should be taken into account to better understand search behavior.

Nonnumeric perception

There are many instances in people’s daily lives in which feedback is nonnumeric—for example, when forming impressions about other people (Prager et al., 2018). Other examples include the side effects of medications, food choices, and climate-change consequences, which are mostly experienced through nonnumeric feedback. So far, there has been little research about nonnumeric feedback in DfE experiments. There is an adjacent line of research, however, that has dealt with nonnumeric stimuli, namely ensemble perception (Whitney & Yamanashi Leib, 2018). This line of research examines how participants develop summary evaluations of stimuli, such as line orientation, color hues, or facial expressions. This research has also demonstrated that higher-level concepts such as average economic value can be estimated quite precisely from sets of images of consumer products (Yamanashi Leib et al., 2020).

These nonnumeric experiences are processed differently than numeric ones. For example, when asked to provide summary judgments separately for clusters of visual and numeric stimuli, participants’ accuracy decreased as a function of cluster size for numeric, but not for visual, stimuli (Rosenbaum et al., 2021). Importantly, research in ensemble perception usually uses different experimental designs than DfE (Whitney & Yamanashi Leib, 2018). First, in ensemble perception research stimuli are presented simultaneously and not sequentially, as in DfE. Second, the choice in ensemble-perception research is usually perceptual with an objectively correct answer, whereas DfE research examines preferential decisions. Including stimuli sets and theories on ensemble perceptions into experimental DfE could be a promising avenue for future study of nonnumerical experiences.

Finally, when examining information search, experimental DfE research usually studies only numeric experiences. These experiments present only a limited number of search options and thus constrain information search. In real-world decision problems, however, there are many experiences to consider, including nonnumeric ones—for example, on social-media sites on the Internet. An understudied aspect of the information-search issue concerns when people actively search for this nonnumeric information (perhaps, for example, with respect to other people’s experiences with medical side effects or climate change), to overcome potential misinformation (see Pennycook & Rand, 2021), and when they are satisfied with the status quo of their own experiences and look no further. This decision could also depend on people’s subjective confidence in the current state of knowledge, which in turn could be determined by the underlying cognitive processing of these nonnumeric experiences. Hence, by broadening the experimental paradigms to include nonnumeric and context-rich stimuli, research in DfE can help to tackle the pervasive problem of fake news and misinformation.

Content of memory

In standard DfE experiments, memory processes are usually not the focus of analysis. Implicitly, it is assumed that participants only take the information sampled immediately prior to the decision into account (e.g., Kopsacheilis, 2018). Decisions, however, can also be affected by episodic memory (Duncan & Shohamy, 2016; Madan et al., 2014), such as when people prefer options they recall better (Weilbächer et al., 2020). One unresolved topic is how these episodic memories are represented in DfE. For example, do people recall specific past episodes (e.g., when milk cost £1.12) to inform their choices or do they learn less concrete properties of the environment (e.g., when milk costs between £0.80 and £1.50)? This distinction between different types of memory representations is also reflected in computational models of memory and decision-making. Instance-based episodic models assume that people have precise representations of individual experiences (e.g., Gonzalez et al., 2003; Hotaling et al., 2019; Lengyel & Dayan, 2007), whereas gist models assume that people have highly abstracted representations of past experiences (Brainerd et al., 1999; Steyvers & Griffiths, 2008). These differences in representation are important when designing practical intervention strategies. For example, if the goal is to reduce problematic behavior such as speeding or gambling, reminders of concrete past experiences would be more effective when people precisely recall specific past experiences compared with when they base their decisions on a general gist representation.

To better understand the memory representations that influence choice, we need to update the paradigms used in DfE research. To date, many experimental paradigms in DfE have used a limited set of outcomes per decision problem, often only three (see Wulff et al., 2018). This reduces the memory demands of the task and makes it difficult to assess people’s memory for individual samples or episodes. Thus, an emerging area of research is how memory supports decision-making when continuous outcomes are used (e.g., Mason et al., 2022; Olschewski, Newell, et al., 2021; Spektor et al., 2019). Continuous outcomes can also build bridges to the fields of retrospective evaluations and numeric cognition (discussed above). Using continuous outcomes as well as nonnumeric stimuli in DfE tasks will allow researchers to better understand the types of memory representations used and to distinguish between instance-based or gist-based samples.

Similarly, stochastic processes in experimental DfE tasks are usually stationary, rather than nonstationary, which is likely to impact the types of memory representations used to guide choices. In contrast, nonstationary or dynamic task environments would better reflect the real-world contexts in which people learn about investment returns or social interactions, where both the rate of return and the behavior of people, respectively, can change over time. In such nonstationary environments, episodic learning models that incorporate individual samples, rather than simply running averages, can explain choices better (Bornstein et al., 2017; Gershman & Daw, 2017). Therefore, more work examining nonstationary environments can help researchers to better understand how memory representations affect DfE (see also Konstantinidis et al., 2022).

From a theoretical perspective, an important future advancement is to integrate models of memory representations of real-world experiences with models of choice in DfE. As an example, decision by sampling (DbS; Stewart et al., 2006) specifies what people store in long-term memory when making decisions. By doing so, the theory provides a process account of how the subjective value of a choice option is constructed by comparing the rank position of a target item within a small subset of samples retrieved from long-term memory. The distribution of items in long-term memory is assumed to reflect real-world frequencies. In principle, if DbS were combined with a process detailing how that store of information is formed and updated with new experiences, the resulting theory could be applied to DfE to explain the influence of long-term memory processes on DfE. A challenge to this approach, however, is that exactly how people represent the distribution of past experiences in memory is not known (Szollosi et al., 2022; Tran et al., 2017).

Time course

In the standard DfE experimental paradigm, feedback from a given action is usually provided immediately and is based on a single action. This feedback structure contrasts with real-world problems, such as climate change or healthy eating, in which outcomes are the result of a series of decisions and feedback is often delayed. One approach to study these environments has been microworld experiments, in which participants are presented with a simulated world where they can make decisions over multiple in-experiment years or seasons, getting repeated feedback in a manner similar to DfE tasks (Dörner & Güss, 2022; Gonzalez et al., 2005; Kumar & Dutt, 2018; Liang et al., 2019; Meyer, 2012; Newell et al., 2016). Whereas these studies capture some of the complexity of real-world decisions, introducing long time delays between actions and outcomes, which require long-term memory processes, was beyond the scope of these studies (but see Lejarraga, 2010). Thus, a future direction could be to examine learning in DfE over longer periods of time, perhaps using multiweek studies in which participants are invited to submit decisions and receive feedback each week through online or smartphone applications; this method has been fruitfully used to study spatial cognition or mental health (see Coutrot et al., 2018; Gillan & Rutledge, 2021). This extended timeline could allow researchers to study how decision makers consolidate memories of experiences over longer time periods and how this affects decisions and attitudes in DfE.

If experiences from across a longer time scale affect behavior in DfE, an important open question is how these experiences are integrated with more immediate feedback. This issue is an extension of the general problem of integrating prior information with new incoming data. In context-rich domains, such as health- or environment-related decisions, semantic memories, which represent general knowledge about the world, can be conceptualized as prior information. For example, when trying to evaluate the risk associated with nuclear power plants, people may be more likely to recall general information about concepts such as nuclear technology, chemistry, or atoms as opposed to direct (episodic) experiences. Recent research in risk perception could help pave the way to combining semantic information with DfE. For example, researchers have examined how the structure of word distributions in the natural language, extracted from Internet databases, influences individuals’ risk perception of events such as nuclear war (Bhatia et al., 2019). Relatedly, aspect listing of risk associations has been shown to affect overall self-reported risk preferences (Steiner et al., 2021). Moreover, research in probability judgment has used Bayesian inference models to combine prior information sampled from memory with information sampled directly from the immediate environment (Zhu et al., 2020). A similar approach could be adopted in DfE research to probe how semantic memory representations influence behavior when the interpretation of feedback depends on semantic networks. Furthermore, taking prior beliefs more thoroughly into account in DfE could improve understanding in settings where ideological beliefs might lead to individual differences in the semantic networks of deciders. This in turn might lead to different behaviors, even when deciders encounter the same experiences in the immediate environment (e.g., Hahnel et al., 2020).

Integrating feedback

In many real-life scenarios, such as in finance, making choices requires considering different types of information from different sources, including one’s own previous experience (as in DfE) and descriptive summaries of the choice options (e.g., outcomes and probabilities, as in DfD). Recently, experimental studies have investigated how introducing descriptions in a typical DfE task affects choice behavior (e.g., Barron et al., 2008; Hertwig et al., 2018; Jessup et al., 2008). An early study found that descriptions are neglected when available in a DfE task (Lejarraga & Gonzalez, 2011). Further research has shown that descriptive information can have an impact on choice, but mostly at the early stages of the task, when there is little experience and descriptive summaries offer valuable information about the choice options; as more experience is accumulated, this influence of description diminishes (Weiss-Cohen et al., 2016). Another factor is a high number of available choice options, which makes learning from experiences harder (Ashby et al., 2017; Frey et al., 2015; Konstantinidis et al., 2015). Thus, when choice options increase, descriptive information can help participants to maximize rewards, but an even higher number of choice options can make participants neglect descriptive information because of information overload (Jacoby et al., 1974; Weiss-Cohen et al., 2018).

Though there seems to be a general tendency for weighting experienced information higher than described information in experimental paradigms, the reason for this is not clear. One possibility is that experiences are more salient. Another is that sequential experiences can be more easily integrated into existing valuations than described outcomes and associated probabilities can (see Busemeyer & Townsend, 1993; Erev et al., 2017; Glöckner et al., 2012). In a contrary finding, however, participants seemed to prefer descriptive over experienced information (Lejarraga, 2010) when they can choose between them, and participants were more confident about decisions based on described information than experienced information (Lejarraga & Lejarraga, 2020).

Further, the influence of descriptive information may depend on other characteristics of the choice environment. For example, the type of experience, whether nonconsequential (as in the sampling) or consequential (as in the repeated-choice protocol), may allow for different degrees of influence of descriptive information. In the studies conducted thus far, experience was usually consequential (e.g., Weiss-Cohen et al., 2016, 2018). A possible hypothesis is that nonconsequential experiences might have less influence on choices than consequential ones, because of lower emotional involvement. Furthermore, descriptive information can have different epistemic content. Descriptions can offer information about the underlying stochastic processes (as in casino games) or summaries of empirical observations (as in the stock market or sport bets). In principle, descriptions about stochastic processes are more informative than descriptions of empirical observations, which could be subject to sampling error or temporal trends—but how this distinction influences choice behavior is not known.

Another important aspect of real-world decision-making is the reliability and trustworthiness of information. In other domains, such as sensorimotor learning (e.g., Körding & Wolpert, 2004) or spatial cognition (e.g., Chen et al., 2017), people approximate the Bayesian ideal when integrating multiple information sources. In DfE research, however, explicitly manipulating these factors is rare, and little is known about how people integrate information varying in reliability. One possibility is that descriptive summaries may be perceived as unreliable because they are usually provided by someone else. Personal and situational factors can determine how trustworthy descriptive information is perceived to be, and this evaluation can affect its impact on choice. For example, advertisements provide summary information only to increase product sales, which suggests that this information is highly unreliable. In contrast, experiences are often directly observed as a consequence of one’s actions. Vicarious experiences, however, can be unreliable. For example, observing the efficacy of a diet or medical treatment in someone else’s life cannot be considered reliable information concerning the potential impact of those approaches on the observer, because the impact of those treatments may depend on unobserved situational and physical heterogeneity. Moreover, when experiences are communicated, these experiences can be biased: Extreme events, for example, are more likely to be passed on between people (Plonsky & Teodorescu, 2020a; see also Fiedler, 2000). Understanding how people deal with perceptions of the reliability of feedback can have applications into policy implementations. For example, warning labels often convey descriptive information about gambling, health, or security risks in order to improve decision-making. Because people combine information from different sources to make judgments and decisions, effective design and development of descriptive communication should also take into account people’s personal experience with a situation (e.g., Weiss-Cohen et al., 2021).

Mental models and pattern recognition

In addition to learning from direct experience, the processes of information integration can be influenced by people’s assumptions about the underlying stochastic process that generates rewards. These assumptions are particularly relevant in context-rich domains, such as climate change, where people have additional theories about how the world works. The impact of these additional assumptions on learning and behavior relates to the distinction in reinforcement learning between model-free learning, which is based on trial and error, and model-based learning, which incorporates a model of how actions and outcomes are connected in an environment (Daw et al., 2011). Most analyses in DfE have assumed that participants do not incorporate such higher-order representations about the stochastic process. For example, when the underlying odds of winning or losing in a DfE task are stationary, participants have been assumed to also treat the task as stationary. Yet recent evidence suggests that participants behave as if the environment is dynamic, and they develop higher-order mental models and concepts about upcoming outcomes, resulting in identifiable search and choice patterns (Barron & Leider, 2010; Clotfelter & Cook, 1993; Cohen & Teodorescu, 2021; Oskarsson et al., 2009; Szollosi et al., 2019; but see the discussion in Ashby et al., 2017; Plonsky & Teodorescu, 2020b; Yechiam et al., 2020). Expectations about upcoming rewards can originate from identifying certain regularities in the environment, such as that larger rewards are less likely to be received than smaller rewards (the risk-reward regularity; see Pleskac & Hertwig, 2014). There is evidence of the use of such heuristics: Leuker et al. (2019) found that when lotteries are atypical (i.e., do not follow this risk-reward regularity), participants take longer to make choices and show higher levels of attention (measured by eye tracking) in a DfD task. Similarly, probability judgments and search patterns in DfE have shown that people assume a negative relation between probability and reward (Hoffart et al., 2019).

If participants construct such mental models in controlled experiments, these mental models are likely to also affect behavior in, for example, debates about the right policies to fight climate change, a topic to which people bring their own lay theories. Studying and integrating these expectations into learning models can thus improve explanations of behavior (see also Leuker et al., 2018). Moreover, real-world decision problems often contain complex interdependencies, such as when climate-change policies affect important outcome measures or restrict future choice options. Future studies could incorporate these interdependencies into experimental DfE research by including choices between multiple actions with complex causal relations or allowing for sequential decisions that can change the composition of a future decision environment (see Brehmer, 1992; Liang et al., 2019; Meyer, 2012). Also, little research has examined whether (and how) people learn and represent correlations between choice options from experience (Kareev, 2000; Laudenbach et al., 2022; Olschewski, Diao, & Rieskamp, 2021). In this context, it matters whether feedback from options that are not chosen is available or not (i.e., full vs. partial feedback; Ben Zion et al., 2010; Grosskopf et al., 2006). When feedback of forgone options is available in a large set of uncorrelated options, people sometimes chase risky options that provided a huge reward most recently. Finally, it is important to understand whether people can transfer experiences in one environment to other environments when these environments share critical, model-based features in an adaptive way (Juechems et al., 2022). These learning transfers, for example, could depend on similarity functions between experiences, options, or environments, and more research could explore under what circumstances such similarity is used to guide learning (Spektor et al., 2019).

Another behavioral observation that may originate from participants trying to recognize and exploit patterns in sequentially occurring experiences is recency. Recency means that recent events have a higher influence on choice behavior than normatively warranted, assuming a stationary generating stochastic process (see, e.g., Hogarth & Einhorn, 1992; Murdock, 1962). The effect of recency on choice, however, could depend on whether experiences are consequential or not, as recency has more consistently been reported with consequential experiences (e.g., Madan et al., 2014). In contrast, in the sampling protocol, recency effects are stronger when participants self-determine their number of samples but are less consistently found when the number of samples is externally determined (Wulff et al., 2018). In the former case, recency might not be due to pattern recognition but rather could follow from strategic information search (i.e., people stop sampling when they see a good outcome from the preferred option or a bad outcome from the nonpreferred option).

In contrast to this positive recency, a wavy-recency effect has been reported when rare events are present (Plonsky et al., 2015, Plonsky & Erev, 2017; Szollosi et al., 2019). The wavy-recency effect emerges after a rare, extreme outcome occurs amid a sequence of smaller, more common outcomes. In this case, people initially respond to this large deviation with negative recency and are less likely (after a rare win) or more likely (after a rare loss) to select the option that yielded that rare event. Within a few trials, however, people quickly reverse themselves and go back to selecting the option that yielded the large win and avoiding the option that yielded the large loss (showing positive recency)—even more than baseline levels.

These sequential effects of experiences can depend on subtle differences in mental accounting. Imas (2016) showed that experienced losses can increase risk taking when losses are not realized but can decrease risk taking when losses are realized. This distinction could be connected to consequential and nonconsequential experiences in DfE and has the potential to identify interventions that can change the influence of recent experiences. For example, in gambling, players could be nudged to repeatedly realize their gains and losses during a visit to a casino. Similarly, nudging a patient to change doctors during a medical treatment might make patients conclude a series of experiences and “realize” the overall gain or loss of a past treatment. As a consequence, the experience of treatment success or failure could affect subsequent risk taking in health-related decisions differently depending on whether the patient changed doctors or not play a larger role in real-world behavior.

In general, recency can be helpful for adapting to changing environments (Konstantinidis et al., 2022). In important societal problems, such as vaccine uptake to prevent the spread of a disease, however, conspiracy theories or ideological reasons may prevent some people from taking recent experiences into account (Jennings et al., 2021; Pertwee et al., 2022). Consequently, future research in DfE should also examine circumstances in which recent information is ignored, either because people have a strong prior conviction or because motivated cognition biases people’s learning processes to underweight certain experiences.

Social learning

Learning often happens in social environments, interactions, and contexts; an important question is how learning in social contexts differs from learning in nonsocial contexts. Social decision-making has been extensively studied with strategic games in which a payoff matrix of possible actions and rewards for all players is explicitly described, and participants receive information about the actions of other players. Such action-feedback schemes closely resemble DfE, and substantial research has documented that feedback about other players’ actions affects individual behavior and group outcomes of cooperation (e.g., Bereby-Meyer & Roth, 2006; Fischbacher & Gächter, 2010). Much less is known, however, about how people learn in social interactions when descriptions of the consequences of their actions are not available.

In the prisoner’s dilemma, where cooperation leads to higher social welfare but is dominated by defecting under pure self-interest, participants can learn to cooperate when the full payoff matrix is described (e.g., Andreoni & Miller, 1993). If no description is provided and people only experience outcomes from their interaction with other players, cooperation diminishes (Martin et al., 2014). This decline shows how experiencing outcomes (as examined in DfE) may not be enough to promote cooperation in situations where the social interactions are complex. This finding dovetails with the idea that in complex or partly ambiguous situations, such as those in which people learn from individual experiences, people can interpret information self-servingly and behave less prosocially (Dana et al., 2007; Olschewski et al., 2019) compared with decisions under certainty (Engel, 2011) or in DfD (Polman & Wu, 2020). Relatedly, participants in the prisoner’s dilemma are less responsive to information about the cooperation rates of their fellow participants when that information is presented through experiences rather than through descriptions (Isler et al., 2022).

Another question about social processes is how people learn from experience when uncertainty is generated by the behavior of other humans, rather than stochastic processes. This uncertainty can be studied with the ultimatum game (Güth et al., 1982), which gives one participant (the proposer) the possibility to distribute a certain amount of money while the other participant (the responder) can decide whether to accept the proposal or reject it; rejecting the proposal results in both participants receiving nothing. In this scenario, the proposer can reduce uncertainty about what the responder will do by referring to social norms. One such social norm is that splitting an outcome equally between two players is always acceptable. Consequently, social norms can replace learning through feedback and in that way reduce information search and impact behavior (Fleischhut et al., 2021). Experiments in DfE can be extended to take social interactions into account to study how beliefs about social norms and fairness interact with personal experiences to shape cooperation and trust, or the opposite.