Talking with strangers is surprisingly informative

In a typical experiment, each participant had a 10-min conversation with a randomly matched, unacquainted participant. Before the conversation, participants read about general features of the conversation (e.g., how long it would be and that their conversation partner was recruited in the same way they were) and then reported how much they expected to learn (e.g., “How much do you think you’ll learn from the conversation?” and “How much do you think you’ll learn about the other person?”; see Method and Materials). After the conversation, participants reported how much they had actually learned (e.g., “How much did you learn from the conversation?” and “How much did you learn about the other person?”). In most experiments, participants answered additional questions about hedonic and social aspects of the conversation to replicate existing findings (e.g., how much they enjoyed the conversation, how much they liked their conversation partner; SI Appendix has more information). We discuss these for experiment 7 and in Discussion to differentiate learning from other conversation outcomes.

In experiment 1, we recruited visitors to public parks in Chicago as participants. As a robustness check, each randomly matched pair received either conversation prompts perceived to be relatively mundane in pretesting (e.g., “Tell me about what you do for a living”) or to be relatively interesting (e.g., “Tell me about an interesting person you know or whom you’ve met”). As shown in Fig. 1, participants learned more than they expected from the conversation across the two learning measures [in general, b = 1.14, t(206) = 8.20, P < 0.001; other person, b = 1.22, t(206) = 9.30, P < 0.001], regardless of the prompts they received (|b|s < 0.25, |t|s < 0.87, Ps > 0.38).

Experiment 2 suggests that underestimating learning is robust, occurring whether people are explicitly trying to learn in their conversation or not. In this experiment, we randomly instructed participants to try to learn something from their conversation partner (learning goal) or to simply have a pleasant conversation (control goal). The conversations were not directed by any prompts, allowing us to also test the robustness of experiment 1 in unguided conversations. As shown in Fig. 1, participants again learned more from their conversations than they expected [in general (M_pre = 4.66, SD = 1.75 vs. M_post = 6.12, SD = 2.09), b = 1.45, t(198) = 10.68, P < 0.001; other person (M_pre = 5.36, SD = 1.78 vs. M_post = 6.71, SD = 1.71), b = 1.35, t(198) = 9.77, P < 0.001], regardless of whether they were explicitly instructed to try to learn or not (|b|s < 0.29, |t|s < 0.91, Ps > 0.36) (additional analyses conducted at the level of the pair yielded similar results; SI Appendix).

Experiment 3 used additional measures of learning and tested one explanation for our results. We measured learning in three different ways. First, in addition to asking about learning in general and learning about the conversation partner, we also asked about learning useful information. As shown in Fig. 2, participants underestimated how much they would learn in all three categories [other person, b = 1.07, t(205) = 8.47, P < 0.001; in general, b = 0.43, t(205) = 2.65, P = 0.009; useful information, b = 1.28, t(205) = 8.61, P < 0.001]. Underestimating learning was not restricted to a single type of information and included what participants themselves reported to be useful information.

Second, we asked participants to write what they expected to learn and then what they actually learned, which two independent coders then rated on abstractness/concreteness. If people genuinely learn information in conversation, rather than simply reporting learning information, then their text descriptions should be more concrete and detailed after the conversation than before the conversation. Indeed, participants’ descriptions of what they reported actually learning were more concrete (M = 4.23, SD = 1.41) than what they expected to learn [M = 2.55, SD = 1.02; b = −1.68, t(204) = −17.3, P < 0.001].

Third, we recruited a separate set of third-person raters whom we randomly assigned to read participants’ descriptions and rate how much they believed the participants learned from their conversation. We altered participants’ expectations text from the future tense to the past tense to match the experiences text and then asked raters to predict how much a participant learned based on each text description. Consistent with genuine learning from conversation, the raters identified more learning in descriptions of what participants actually learned (M = 5.67, SD = 1.25) than of what they expected to learn [M = 4.39, SD = 1.33; b = 1.28, t(2,869) = 23.28, P < 0.001]. Taken together, experiment 3 indicates that participants themselves reported learning more useful information in conversation than expected, they wrote more concrete information after conversations than before conversations, and separate raters thought people learned more when reading their reports from after the conversation than from before it.

Finally, experiment 3 also tested a plausible explanation for why people underestimate learning. Because people have direct access to what they know but not to what others know, people may think they have relatively little to learn from others compared to what others might have to learn from them. To test this possibility, participants in experiment 3 also reported how much they expected others to learn from the conversation. Inconsistent with this explanation, participants actually expected that they would learn more general information from their partner than their partner would learn from them [b = 0.28, t(205) = 2.39, P = 0.018]. Participants did not expect significant differences in how much they would learn about their partner compared to how much their partner would learn about them [b = 0.15, t(205) = 1.43, P = 0.15] (Fig. 2). We also observed no significant differences in text descriptions of what participants expected to learn from their partner compared to what their partner would learn from them in terms of the number of words generated [M_self = 14.6, SD = 11.1 vs. M_other = 15.1, SD = 12.0; b = −0.54, t(205) = −0.73, P = 0.46], how easy or difficult it was to generate the content [M_self = 3.57, SD = 2.88 vs. M_other = 3.55, SD = 2.93; b = 0.02, t(205) = 0.12, P = 0.91], or how confident participants were about the content they generated [M_self = 8.40, SD = 2.41 vs. M_other = 8.38, SD = 2.37; b = 0.02, t(205) = 0.14, P = 0.89]. Underestimating learning in conversation does not seem to stem from believing that there is relatively little to learn from strangers.

The preceding experiments indicate that people underestimate how much they will learn in conversation and that this gap stems not from people being unmotivated to learn or thinking they have relatively little to learn from strangers. Instead, we hypothesize that people underestimate learning in conversation because of the inherent uncertainty of conversation. A conversation is dynamic, involving two people who can discuss a wide range of material, with no single person in complete control of the conversation’s content. This makes it very difficult to bring to mind any certain details of how a conversation will unfold, and hence difficult to bring to mind any information about what one might learn. Because people tend to base their expectations on the information that comes most readily to mind (16), the conversation people are likely to imagine is going to contain systematically less information than their actual experience of a conversation. This could lead people to systematically underestimate how much they are likely to learn in an actual conversation to the extent that its contents are uncertain beforehand. We therefore predict that expectations about learning will be better calibrated when people are better able to anticipate the content of the conversation.

We tested this uncertainty explanation in the subsequent three experiments. In experiment 4, we compared expected vs. actual learning from conversation against learning in a context that involves more certainty: surfing the web. Like conversation, surfing the web offers tremendous opportunity to learn. Unlike conversation, it is easier to imagine what might be learned surfing the web because the user chooses the content and hence can anticipate more concretely what might be seen. We assigned participants to either spend 10 min talking with another randomly assigned participant or 10 min surfing the web, asking participants in both cases to use their time to learn interesting things. As shown in Fig. 3, participants’ learning expectations were more miscalibrated in the conversation condition than in the web surfing condition [b = 0.65, t(200) = 2.30, P = 0.023]. Specifically, participants in the conversation condition significantly underestimated how much they would learn [b = 0.54, t(101) = 2.70, P = 0.008], whereas participants in the web surfing condition did not [b = −0.10, t(99) = 0.52, P = 0.61]. These results indicate that people do not always underestimate how much they are likely to learn, but rather are likely to do so when it is especially hard to know the details of what one might learn beforehand, such as in conversation.

Not all conversations, however, have the same degree of uncertainty about what might be learned. In experiment 5, we manipulated certainty about the conversation’s content by varying whether people spoke with a stranger or with someone they knew relatively well. To do so, we recruited multiple pairs of closely acquainted dyads (friends, spouses, and family members) to each session and asked each participant to have two conversations: one with the acquainted person and one with a stranger (someone else’s acquainted person). Participants first reported what they expected to learn from each conversation using the same items as in experiment 4 and then reported how much they actually learned after each conversation. As shown in Fig. 3, participants significantly underestimated how much they would learn from a conversation with a stranger [in general, b = 0.86, t(103) = 4.27, P < 0.001; other person, b = 1.36, t(103) = 7.13, P < 0.001; useful information, b = 0.99, t(103) = 4.11, P < 0.001] but did not significantly underestimate learning from their acquainted conversation partner (|t|s < 1.66, Ps > 0.10; interactions, |t|s > 1.8, Ps < 0.07).

Experiment 5 indicates that people are especially likely to underestimate learning when the conversation’s content is especially uncertain, revealing information that cannot be anticipated before actually having the conversation. We test this role of uncertainty most directly in experiment 6. We recruited participants in unacquainted dyads to have a conversation over a videoconferencing platform and report how much they learned (SI Appendix has full details), and then we showed these conversations to observers, whom we asked to estimate how much the participants reported learning in the conversation. We manipulated whether the observers made these estimates before watching the video, when its content was uncertain, or after watching the video, when its content was completely known. Observers underestimated how much participants in the actual conversation reported learning to a greater extent when they predicted learning before watching the conversation than after watching the conversation (SI Appendix, Fig. S6). This difference in calibration was statistically significant for learning about the other person [b = −0.49, t(418) = −3.24, P = 0.001], was in a similar pattern but marginal for learning useful information [b = −0.35, t(398) = −1.85, P = 0.06], and was nonsignificant for learning general information [b = −0.22, t(402) = −1.12, P = 0.26]. Similarly, observers’ estimates on some measures were less correlated with participants’ ratings before watching the conversation than after watching the conversation [other person, r_before = 0.04 vs. r_after = 0.19; in general, r_before = 0.02 vs. r_after = 0.23; useful information: r_before = −0.05 vs. r_after = 0.03].

Finally, experiment 7 examined whether miscalibrated expectations of learning are distinct from miscalibrated expectations about the overall positive experience of conversation, and provided one more test of learning with a more objective measure. We manipulated the topic of conversation for US-based participants by either allowing them to have an unconstrained conversation about anything they liked (open-ended condition) or instructing them to have a more constrained conversation about the United States (US-focused condition). As a more objective measure of learning, we asked participants to list each piece of information they actually learned in the conversation. To measure expectations of learning, we asked participants before their conversation to report how many pieces of information they expected to learn. As shown in Fig. 4, participants significantly underestimated how many pieces of information they would learn in both the US-focused and the open-ended conditions, but did so significantly less in the US-focused condition (b = 0.02, t = 2.07, P = 0.038)*. The self-reported learning measures showed a similar interaction pattern [b = 0.92, t(203) = 3.42, P < 0.001], with participants significantly underestimating learning in the open-ended condition [b = −0.92, t(103) = −5.16, P < 0.001] but not in the US-focused condition [b = 0, t(100) = 0.00, P = 1]. In contrast, participants’ tendency to underestimate how much they would enjoy the conversation and like their partner did not differ between conditions (|ts| < 1.4), indicating that learning is a distinct outcome of conversation (SI Appendix, Fig. S7). In fact, participants in the constrained conversation reported learning less than participants in the open-ended condition [b = −0.90, t(201) = −3.41, P < 0.001], but they reported enjoying the conversation marginally more [b = 0.45, t(201) = 1.88, P = 0.06] and liking their partners significantly more [b = 0.48, t(201) = 2.45, P = 0.015] (SI Appendix, Fig. S7). Although existing research suggests that people’s subjective sense of learning can be disconnected from actual learning [e.g., (17)], the results of experiment 7 suggest that they are aligned when learning from strangers in conversation (SI Appendix has an expanded discussion).

Aristotle argued that “All [people] desire by nature to know,” (4) but our experiments suggest that people may not fully understand what they can come to know from talking with other people. Decisions about whom to approach or avoid in daily life are based at least in part on the presumed value of interacting with another person, including how much will be learned in the interaction. Our experiments indicate that people consistently underestimate how much they would learn from strangers in conversation. Undervaluing learning does not stem from a belief that there is little to learn from strangers. In fact, our participants thought they would learn more from others than others would learn from them. Undervaluing learning is also not an invariant bias in judgment, as people were relatively well calibrated about learning from some activities (e.g., surfing the web) and with some conversation partners (e.g., closely acquainted people). Instead, undervaluing learning appears to stem from the uncertainty of conversation, such that people underestimated learning most when having open-ended conversations with strangers. Ironically, not knowing what could be learned in conversation may keep people from having the very experiences that would show them how much they can learn in conversation.

Learning is obviously not the only outcome of conversation. Existing research indicates that people also underestimate how positive the hedonic and social outcomes of conversations will be (5, 7, 12). Indeed, participants in our experiments also tended to underestimate how much they would enjoy their conversations and how much they would like their conversation partners. Although the underestimation of the instrumental value of learning is correlated with the underestimation of enjoyment (r = 0.37; range: 0.22 to 0.54) and liking (r = 0.28; range: 0.14 to 0.46), as might be expected, experiment 7 indicated that underestimating learning is not simply part of a more general tendency to underestimate all positive outcomes in conversation. We believe the hedonic value of conversation comes from the social dynamics of conversation that create a sense of connection with another person (such as responsiveness, disclosure, and reciprocity), whereas the value of learning in conversation comes from the content of the conversation itself.

One unanswered question is how a conversation’s length affects expected and actual learning. In contrast to the consistent results we observed across experiments, London-area commuters in a field experiment on trains did not show a significant tendency to underestimate learning when asked to have a conversation with a stranger during their commute (18). These conversations, however, tended to be very short, with many being only a few minutes long. Indeed, the extent to which people underestimated learning was positively correlated with the reported duration of the conversation [b = 0.04, t(94) = 3.80, P < 0.001], suggesting that underestimating learning from conversation is dependent on conversation length. A supplemental experiment we conducted manipulating the duration of conversation found that when people knew how long the conversation would be, they underestimated learning from both a 5-min and a 20-min conversation, suggesting that expectations of conversation length may matter as well (SI Appendix). A full understanding of how the duration of conversation is, or is not, incorporated into expectations and experiences of conversation is an important topic for future research.

A second unanswered question is why people’s expectations remain miscalibrated despite presumably having had many conversations throughout their lives and hence many opportunities to learn. One barrier to calibrated expectations may come from the sheer diversity of knowledge that could be learned, which makes generalizing from one conversation to another especially challenging. Another barrier is that social beliefs can be self-fulfilling, such that people find out what they can learn only from the conversations they actually have but don’t find out how much they could have learned from the conversations they avoided (19). If people primarily engage in conversations with the people they expect to learn a lot from, then they will fail to learn how much they could have learned from someone they deemed uninteresting or uninformative. We would predict that more experience in open conversations with randomly selected people, rather than self-selected people, would lead to better calibrated expectations of learning in conversation. Without such experience, miscalibrated expectations about how much can be learned from other people may keep people from learning more in everyday life, frustrating their desire to know by keeping them from approaching a surprisingly informative source of knowledge.

All procedures of these experiments were approved by the institutional review board of the University of Chicago.