Reasoning beyond available data is a ubiquitous feature of human cognition. But while the availability of first-hand data typically diminishes as the concepts we reason about become more complex, our ability to draw inferences seems not to. We may offset the sparsity of direct evidence by observing the statements of others, but such social meta-inference comes with challenges of its own. The strength of socially-provided evidence depends on multiple factors which themselves must be inferred, like the knowledge, social goals, and independence of the people providing the data. Here, we present the results of an experiment aimed at examining how people draw conclusions from information provided by others in the context of social media posts. By systematically varying the degree of consensus along with the number of people and distinct arguments involved we are able to assess how much each factor affects the conclusions reasoners draw. Across a range of topics we find that while people are influenced by the number of people on each side of an argument, the number of posts is the dominant factor driving belief revision. In contrast to well established findings in simpler domains, we find that people are largely insensitive to the diversity of the arguments made.

Sampling assumptions — the assumptions people make about how an example of a category or concept has been chosen — help us learn from examples efficiently. One context where sampling assumptions are particularly important is in social contexts, where a learner needs to infer the knowledge and intentions of the information provider and vice-versa. The pedagogical sampling assumptions model describes a Bayesian account of how learners and providers should behave given different assumptions they have about the other (e.g., is the provider trying to deceive or help me? Does the learner trust me?). In this study, we tested how well this model could describe learning behaviour in the rectangle game, where a fictional information provider revealed clues about the structure of a rectangle that the learner (a participant) needed to guess. Participants received clues from either a helpful information provider, a provider who was randomly sampling clues, or one of two kinds of unhelpful providers (who could mislead but could not lie). We found that people learned efficiently and in line with model predictions when the provider was helpful and that this was the case even when no cover story was provided. However, although participants could identify that unhelpful providers were not being helpful, they struggled to learn the strategy those providers were using.

Making inferences about claims we do not have direct experience with is a common feature of everyday life. In these situations, it makes sense to consult others: an apparent consensus may be a useful cue to the truth of a claim. This strategy is not without its challenges. The utility of a consensus should depend in part on the sources of evidence that underlie it. If each person based their conclusion on independent data then the fact that they agree is informative. If, instead, everyone relied on the same primary source, the consensus is less meaningful. However, the extent to which people are actually sensitive to this kind of source independence is still unclear. Here, we present the results of three experiments that examine this issue in a social media setting, by varying the sources of primary data cited via retweets. In each experiment, participants rated their agreement with 12 different claims before and after reading four tweets that were retweeted on the basis of either the same or different primary data. We found that people were sensitive to source independence only when it was clear that the tweeters had relied on the primary data to reach their conclusion. Implications for existing research are discussed.

In social contexts, learners need to infer the knowledge and intentions of the information provider and vice-versa. In this study, we tested how well participants could infer the intentions of different information providers in the rectangle game, where a fictional information provider revealed clues about the structure of a rectangle that the learner (a participant) needed to guess. Participants received clues from either a helpful information provider, a provider who was randomly sampling clues, or one of two kinds of unhelpful providers (who could mislead but could not lie). We found that people learned efficiently and in line with the predictions of a Bayesian pedagogical model when the provider was helpful. However, although participants could identify that unhelpful providers were not being helpful, they struggled to learn the strategy those providers were using, even when they had the opportunity to practise being a deceptive information provider.

In many situations where people communicate (e.g., Twitter, Facebook etc), people self-organise into ‘echo chambers’ of like-minded individuals, with different echo chambers espousing very different beliefs. Why does this occur? Previous work has demonstrated that such belief polarisation can emerge even when all agents are completely rational, as long as their initial beliefs are heterogeneous and they do not automatically know who to trust. In this work, we used agent-based simulations to further investigate the mechanisms for belief polarisation. Our work extended previous work by using a more realistic scenario. In this scenario, we found that previously proposed methods for reducing belief polarisation did not work but we were able to find a new method that did. However, this same method could be reversed by adversarial entities to increase belief polarisation. We discuss how this danger can be best mitigated and what theoretical conclusions be drawn from our findings.

The impressive recent performance of large language models such as GPT-3 has led many to wonder to what extent they can serve as models of general intelligence or are similar to human cognition. We address this issue by applying GPT-3 to a classic problem in human inductive reasoning known as property induction. Our results suggest that while GPT-3 can qualitatively mimic human performance for some inductive phenomena (especially those that depend primarily on similarity relationships), it reasons in a qualitatively distinct way on phenomena that require more theoretical understanding. We propose that this emerges due to the reasoning abilities of GPT-3 rather than its underlying representations, and suggest that increasing its scale is unlikely to change this pattern.

When assessing the quality of health information encountered online, reasoners may rely on the wisdom of others and the degree of consensus apparent. However, it is unclear whether reasoners weigh the opinions of others evenly or make assumptions about the amount of evidence that each has seen. We investigated this question in an online experiment where people were asked to rate their belief in a series of health claims both before and after reading responses from other users. The degree of consensus among these users and their level of expertise (non-experts vs. expert organisations) was manipulated within-subjects. While we found belief change increased monotonically with the degree of consensus for both experts and non-experts, our results indicate qualitatively different patterns of increase between the two groups. Our study suggests that people reason from consensus using nuanced assumptions about the evidence underlying other people's opinions.

When reasoning about a claim, it makes sense to be more persuaded if lots of other people agree. But, there are many factors that make weighing the evidence behind a consensus complicated. For example, a consensus might be more or less informative depending on the type of claim, or whether each consensus member formed their opinions independently. These factors might also influence people differently depending on their own assumptions or preferences. In this study we used a mock social media paradigm to assess how persuaded people were by two factors: the presence of consensus (no consensus vs. consensus), and source independence (a consensus based on independent information sources vs. a consensus formed off shared, dependent sources). We varied these factors at both the group and individual level. At the group level, we assessed a third factor: whether people were influenced by the type of claim being reasoned about (we assessed 60 different claims divided into 4 categories). Almost everyone was more persuaded by consensus trials compared to no consensus trials. However, the strength of this effect was credibly stronger if the claim was likely to have a ground truth. We found that around one third of participants were sensitive to source independence. Of these, three quarters were more persuaded by a consensus based on independent sources, but the quarter who were more persuaded by dependent sources were persuaded just as strongly.

Many people turn to social media for public health information, but such platforms often contain conflicting and inaccurate medical advice. To assess complex health claims online, people may consider the prevailing consensus; however, previous work suggests that people may not be very sensitive to important cues to consensus “quality”. To explore further, across two experiments we tested people’s sensitivity to the consensus-quality cues of source diversity and source expertise. Via a mock Twitter platform, participants rated their belief in a series of health claims both before and after reading various kinds of tweets about the claims. Experiment 1 showed that experts (both individual medical experts and health organisations) were more persuasive than non-experts. Additionally, stances that were supported by a diverse set of sources were more persuasive. Experiment 2 showed that participants continue to favour experts even when outnumbered in tweet quantity by non-experts. When experts were not present, however, participants favoured high tweet quantity. Both experiments suggest that cues to consensus quality (namely, expertise and source diversity) and consensus quantity (tweet quantity) are salient cues in belief revision. These findings are important in understanding how socially acquired health information (and misinformation) shifts opinion, and the role that experts can play.

There is increasing pressure on social media companies to reduce the spread of misinformation on their platforms. However, they would prefer not to be the arbiters of truth as the truth can be subjective or otherwise hard to determine. Instead, they would prefer that social media users themselves show better discernment when deciding which information to share. Here we show that allowing people to share only those social media posts that they have indicated are true significantly improves sharing discernment, as measured by the difference in the probability of sharing true information versus the probability of sharing false information. Because it doesn’t require social media companies to be the arbiters of truth, this self-censorship intervention can be employed in situations where social media companies suspect that individuals are propagating misinformation but are not sufficiently confident in their suspicions to directly censor the individuals involved. As such, self-censorship can usefully supplement externally imposed (i.e. traditional) censorship in reducing the propagation of false information on social media platforms.