Main

Much of human sociality depends on people’s ability to perceive and make inferences about what others think and do1. Human social sensing has been studied in separate strands of research within psychology and sociology. However, its potential to advance social science has not been fully realized. In this Perspective, we show that human social sensors—individuals who are strategically selected and asked about subjective representations of their immediate social environments—can help to describe and predict political and health-related societal trends2,3,4,5,6,7,8,9,10 as well as to advance theoretical and practical understanding of human sociality11,12,13.

Psychologists have long studied how individuals represent and are influenced by their social environments14. However, they have rarely used people’s knowledge about their social environments as a measurement device to learn more about beliefs and behaviours in society. A primary reason for this reluctance is the notion that human social cognition is fraught with biases15. Examples are false consensus, when individuals who support a particular view believe that this view is more common than non-supporters believe16, and self-enhancement, in which people overestimate their performance relative to others’17. By contrast, sociologists have developed sociometric techniques to collect people’s subjective reports about the structure of their social networks18 and have used them to learn more about society through better sampling of different populations19,20, deriving better estimates of their size21, and investigating how networks affect the spread of beliefs and behaviours22,23. However, sociologists have typically not collected people’s subjective representations of what their social contacts believe or do, or data on the cognitive processes that underlie social influence (but see ref. 24).

Two key developments pioneered by computational social science can bridge this gap between psychology and sociology. The first is the continuously increasing amount of data about human social networks, derived from the multitude of digital and other traces of people’s connections on social media, by phone, or in person25,26,27,28. Such data enable a better understanding of how human social cognition interacts with social environments and how apparent cognitive biases can be a product of an unbiased mind accurately perceiving biased samples from its social environment29,30,31. The second is the development of computational models of human social dynamics that aim to recreate different patterns of belief and behavioural change and to better understand social systems32,33,34,35,36,37. These models produce quantitative predictions of societal trends and provide important insights about events and interventions that could steer social systems in different directions38,39,40. These models can be fruitfully combined with information from human social sensors to enable better understanding of the cognitive mechanisms that underlie social dynamics11,41,42 (Fig. 1).

Fig. 1: Human social sensing as a resource for computational social science.
figure 1

a, Human social sensors can provide useful information when asked directly about beliefs and behaviours in their immediate social circles. They can provide information about parts of the society that can be otherwise difficult to reach (question marks). b, When human social sensors are selected to be representative of a population of interest, their subjective reports about their immediate social environments (social circles) can help to describe and predict real-world social phenomena. Top, the 2018 and 2020 US elections were predicted better by social-circle expectations than by traditional polling questions about own intentions3. Bottom, flu vaccination behaviour is related to perceived vaccination in social circles as reported in the year before (blue bars) even after accounting for own vaccination behaviour reported in the year before and sociodemographic characteristics (grey bars show the estimated marginal means)5. Error bars are standard errors. c, Reports from human social sensors and other empirical data can inform models of social dynamics, which can be used to make quantitative predictions. For example, computational models inspired by statistical physics can be used to predict belief change12,13. As shown on the right, social beliefs reported by human social sensors can be represented as networks. Consistent networks (here shown as networks in which all nodes agree) are associated with lower levels of cognitive dissonance (or energy in statistical physics’ terms) and therefore have a higher probability than less consistent networks, especially when their overall uncertainty (temperature) is low. By testing and comparing different models, we can generate theoretical insights about social systems and derive new, empirically testable research questions.

Full size image

As we describe next, these developments in computational social science open doors to the use of human social sensing for better description and prediction, as well as to models of social dynamics that are empirically grounded in human social cognition.

Describing and predicting social dynamics

Even though social scientists have never had access to as much data as today, many social phenomena are still hard to understand, including voting behaviour, civil unrest, vaccine hesitancy and epidemic spread. How can we capture early signals of emerging trends within standard research budgets and time frames? And how can we collect data while respecting the privacy of individuals who may not want to reveal their own beliefs? Social phenomena can be difficult to anticipate not only in light of the inherent limits of predicting complex societal systems, but also because important parts of the society are hard to reach or are intentionally hidden43,44 (Fig. 1a). Some important social phenomena happen so fast and unexpectedly that researchers do not have enough time to collect data from sufficiently large samples. Indirect measures of social worlds, such as traces of people’s activity on various social media platforms, are valuable45 but cannot fully compensate for these information deficits. The relevant traces are often unavailable to researchers or are prohibitively costly46. Many people do not use these technologies, and those who do might intentionally modify their digital traces for fear of social costs47.

As we describe next, human social sensors can provide useful reports about the beliefs and behaviours of other people around them, despite apparent cognitive biases in human social cognition and social network biases. Studies that rely on human social sensors can resolve some ethical concerns with collecting social data, and can be conducted economically using standard research tools such as surveys.

In one line of studies with human social sensors, people have been asked to report about properties of their immediate social environments—their social circles. For example, participants in election polls were asked what percentage of their family, friends, and acquaintances would vote for different candidates. These social-circle questions improved predictions compared to traditional polling questions about participants’ own voting intentions in three recent US elections2,3 (Fig. 1b) as well as in three recent elections in European countries with larger numbers of political options (France2, the Netherlands and Sweden4). Social-circle questions were also useful in predicting participant’s own behaviours. For example, people’s reports about the flu vaccination behaviour of their social contacts predicted their own vaccination likelihood a year later, beyond their own past vaccination behaviour and sociodemographics5 (Fig. 1b). These results suggest that human social sensors can help to derive more accurate descriptions and predictions of current and emerging societal trends. One likely reason for this gain in accuracy is that reports about people’s friends indirectly improve the representativeness of the survey sample, allowing researchers to gain more information about the overall population3,48. People also might be more willing to report socially undesirable characteristics of their social contacts than of themselves49. In addition, people’s estimates of their social circles today can provide hints about how they themselves will change in the future due to social contagion, which improves predictions2,5,6.

Another line of studies has shown that people can also provide useful reports about broader populations50,51,52,53. In election polls, these judgements can anticipate election results7,54,55, and prediction markets have been successful in predicting future events across a variety of fields, including business56, medicine57, politics8, and sports58. People’s reports about broader populations might be largely based on what they know about their immediate social environments3,9, but they also include information from other sources such as the media, experts, and general education. Human social sensing of both immediate and broader populations can therefore be usefully combined. The theoretical challenge here is to determine how much weight to give to each of these types of social sensing data. A recent method called Bayesian bootstrapping3 offers a theoretical solution to the integration problem, and has been used for US election polls in 2018 and 2020 to produce forecasts that combine participants’ own voting intentions, their social-circle reports and their predictions of the overall election outcome. The accuracy of these election forecasts surpassed those based on any one type of question alone.

These demonstrations of the usefulness of human social sensors are in line with studies showing the well-developed human capacities for social sensing (Box 1), but appear to contradict decades of research in social psychology that produced a long list of cognitive biases in social judgment15,59. However, these seemingly contradictory findings can be reconciled by considering the statistical properties of social environments in which human cognition operates60,61,62,63,64,65,66,67, which are often ignored in the studies that show cognitive biases. For instance, the degree to which people are surrounded by similar others (homophily)68, together with basic memory processes, can explain30,69 whether people’s judgements of broader populations show false consensus16 or its opposite, a well-documented false uniqueness bias70. And, depending on the shape of the true frequency distribution of a particular belief or behaviour, people’s estimates of the overall population will appear to be biased towards self-enhancement17 (when the true distribution is skewed left so most people perform well), towards the opposite bias of self-depreciation71 (when the true distribution is skewed right so most people perform poorly), or in both directions (when the true distribution is symmetrical)69. A parsimonious explanation for these effects is that the apparent biases result from an unbiased mind that accurately perceives a biased social world. By contrast, relying on the assumption of a biased mind while ignoring the social world requires a different explanation for each observed bias. For example, self-enhancement bias has been explained by inadequate metacognitive skills but self-depreciation bias observed in the same studies by false consensus bias72, although both biases can be explained by the same basic memory process operating in a social environment characterized by some homophily and a symmetric distribution of the target characteristic69. The fact that apparent cognitive biases might stem from biased samples rather than faulty social cognition means that people’s reports about their social environments can contain useful information about the social reality.

When using human social sensing methods, one still needs to be aware of some persistent social network biases that are likely to occur even in the absence of cognitive biases. One such social network bias is homophily bias—people tend to have social circles that are similar to them68, so their reports about the beliefs and behaviours of their social contacts will often resemble their own. This does not mean that these reports do not contain useful information beyond individuals’ own beliefs and behaviours. People frequently report that many of their social contacts are not like them69, and their reports include useful information about diverse segments of the population3. The homophily bias, however, has an important implication for studies with human social sensors: if the goal is to use their reports to describe a broader population, then sensors need to be a representative sample from that population, allowing homophily biases in social-circle reports to cancel out on average. That said, the homophily bias can also be useful to researchers interested in specific (sub)populations: some sampling techniques19,20 rely on homophily to reach samples of small, geographically dispersed, and/or stigmatized populations, such as unhoused people or undocumented immigrants73.

Another social network bias that is important to consider when using human social sensing is the friendship paradox—the phenomenon that one’s friends, on average, always have more friends than oneself74 (Fig. 2). This occurs because people with more friends will be more likely to occur in one’s friend group, and consequently people’s reports about their social circles will inevitably over-represent individuals with many friends29. Depending on the correlation of the characteristic of interest with the number of friends one has, as well as the number of friends one’s friends have, asking people about their social circles can yield biased population estimates. Characteristics that have a positive correlation with the degree in the social network are likely to be overestimated, while those that have a negative correlation with the degree are likely to be underestimated.

Fig. 2: Friendship paradox.
figure 2

a, This example from Feld74 shows a friendship network among a group of high-school students. The number of friends and the number of friends of friends for each student are shown in parentheses. The eight students in the figure have a total of 20 friends. Adding up each student’s friends yields 60 friends of friends. This means that the students have an average of 2.5 friends each and that their 20 friends have an average of 3.0 friends each. This occurs because the students with many friends (for example, Sue and Alice) are the friends of many of the other students. That people’s friends on average have more friends than people themselves have is inevitable in any network where there is any variation in the number of friends (their degree). b, This network bias can be useful for describing and predicting social phenomena. For example, surveying the friends of initially selected participants helped to identify an early outbreak of an epidemic. Reproduced with permission from ref. 6.

Full size image

However, the friendship paradox is not necessarily detrimental to the usefulness of human social sensors. For many societal trends, from the spread of disease to the spread of misinformation, individuals with more social connections might be a more useful early indicator than an average individual in the population. They might be driving the trends in beliefs, fashions or voting intentions and can therefore serve as ‘early warning signals’ for later trends. Furthermore, better connected people are often more likely to be ‘infected’ with a characteristic of interest, such as a contagious disease or misinformation. Identifying those people can help with managing the spread of epidemics and choosing the best-positioned human social sensors6,48,75,76, as well as in implementing public health interventions77.

Using human social sensors to describe and predict social phenomena can alleviate some ethical concerns related to asking about sensitive issues. Human social sensors know that they are participating in a research study, unlike in some applications of ‘mechanical social sensors’ where people might not be aware that their data are used for research purposes78. In addition, human social sensors do not need to reveal the identity of their specific social contacts. They can provide useful information about the relative frequency of a particular characteristic in their social circles (for example, defined as adults they were in contact with within a specified time period) without revealing information about any particular individual. This gives researchers the opportunity to obtain some indication of the characteristics of hidden populations, while protecting their privacy.

Empirically grounded models

To understand why in certain societies new beliefs—such as opinions on climate change or vaccines—spread more quickly or polarization rather than consensus emerges, and to turn these insights into predictions, researchers have developed a number of analytic and computational models of social dynamics, in particular of belief dynamics and collective behaviour11,30,32,33,34,35,36,37,38,39,40,41,42,79,80,81,82,83,84,85.

Most models assume that these processes occur on a social network, reflecting the fact that much of human thought and action happens in the context of interaction with other people86. Computational social science has contributed a large amount of data on the structure of social environments25,26,27,28, but most models of social dynamics still do not incorporate plausible psychological components, such as how people actually experience and interact with their social networks. This can reduce the ability of these models to help with understanding and predicting real-world patterns of belief dynamics and collective behaviour87,88. It has long been recognized that human ability to adapt to an ever-changing social world is shaped by both the structure of social environments and cognitive processes89. What eventually matters for explaining how social environments affect beliefs and behaviours is how these environments are subjectively represented in individual minds. The nature of subjective representation has been a subject of great interest to sociologists from classical works90,91 to recent times92,93.

Subjective representations of social networks, as reported by human social sensors, can be used to enhance the descriptive validity and predictive power of models of social dynamics. This grounding in knowledge about human social cognition is important because these representations, unlike indirect ‘objective’ measures of social environments94, depend on what people attend to at the moment, their past experiences, and their overall social context95. Different people can experience the same social network structures differently, depending on how much they like their social contacts96, how interdependent they are97, and whether they perceive others as members of their group or as outsiders98.

In turn, researchers can use these empirically grounded computational models to understand how the same subjective representations of social networks can influence people’s beliefs and behaviours in different ways, depending on the strategies people use to integrate their social considerations. Models that combine human social sensing with behavioural data from experiments and the real world can implement and compare different plausible integration strategies99,100,101,102,103,104, from heuristics such as random choice and averaging88,105,106, majority rule107,108, and non-compensatory strategies such as following an ‘expert’109,110, to more general normative mechanisms that provide explanations at a different level111,112, such as Bayesian learning113,114,115 and logic116. Such computational implementations and comparison of plausible integration strategies would enable researchers to compare different models and establish minimal models that can still reproduce main empirical patterns.

One useful framework for developing and comparing computational models that can incorporate plausible psychological mechanisms is statistical physics11,36,41,79,87,88,110,117,118,119,120,121,122,123. Statistical physics models of social dynamics can reproduce a variety of empirically observed patterns of belief spread using only a few components, showing that models of complex systems do not by themselves need to be complicated. Traditionally, the field of statistical physics studies the collective behaviour of interacting building blocks of a system, typically atoms or their components, by introducing a function that maps a microstate of the system (that is, a description of the state of every microscopic constituent) onto a single number (for example, the energy in a physical system). That is, statistical physics analyses directly connect microscopic and macroscopic descriptions of the same system by using a function that maps many microscopic variables onto a single macroscopic variable that characterizes an important (and in physics, directly measurable) feature of the current state of the system. Using this collective variable, the statistical physics analyses then assign probabilities to each of these microstates that can be updated according to dynamical rules in an evolving system.

While people are far more complex, dynamically evolving social behaviours in large populations tend to obey statistical patterns that are amenable to mathematical modelling of their micro- and macrostates. With advances in computational social science and human social sensing, models of social dynamics inspired by statistical physics can be empirically grounded in people’s social cognition and networks as measured in surveys and experiments, through analysis of social media, or from other digital footprints124. While the details of any social system certainly do not show one-to-one mapping to analogous physical systems110, the main premise of minimizing energy in physics corresponds to the idea prevalent in social sciences that many choices can be modelled as minimization of dissonance or maximization of utility125. Statistical physics models can incorporate psychological concepts such as cognitive dissonance126 (energy), uncertainty or lack of attention (temperature), subjective representations of networks (linkages), and belief integration strategies (updating rules)110,121,127. The statistical physics framework can be used to implement and compare many different models that have so far been studied independently and without empirical testing11.

In a recent example, van der Does et al.12 used a computational model developed within a statistical physics framework and grounded in human social sensor data to investigate how different educational interventions affect the cognitive dissonance that stems from inconsistent social and moral beliefs. This analysis helped to uncover how people integrate these considerations, as well as conditions under which changes in dissonance due to interventions lead to belief change. In another example, Dalege and van der Does13 used a network theory of individual attitudes inspired by statistical physics121 (Fig. 1c) to investigate how dissonance in subjective social belief networks can predict changes in science-related beliefs. They asked participants in a longitudinal national survey to estimate the percentage of groups such as their family and friends, scientists, or medical doctors that believe in the relative safety of products such as genetically modified food and childhood vaccines. These quantitative reports from several survey waves were used to reconstruct participants’ changing subjective social representations as networks85. Changes in these networks were shown to depend on their inherent dissonance.

The statistical physics framework is just one of many possible analogies that can be useful when attempting to understand and model complex social systems. Beyond the Ising and Potts models that gave rise to the examples mentioned above119, other analogies from physics have been used to understand complex social systems, including percolation128, diffusion101, Monte Carlo methods120, and quantum physics129. Analogies from other disciplines, including epidemiology130,131 and evolution105,132, have been used as well. As one of the basic tools of human thought, analogies can be very useful when trying to better understand a complex phenomenon133. However, it remains important to recognize the worldview, assumptions, and methodologies that are transferred along with the analogy from one system to the other, and to avoid introducing unnecessary baggage by overusing any particular analogy134. Empirical grounding of models based on analogies with other systems, using data from human social sensors among others, provides an important way to constrain these models and check their usefulness for describing complex social systems.

Outlook

Further research on human social sensing could investigate ways to increase its informational value, explicitly include it in models of social dynamics, and combine it with machine learning. One important research direction is further development of theoretically grounded strategies for sampling human social sensors135, for example by taking advantage of the friendship paradox10,48,76,136. Research is also needed on statistical methods for deriving proper population inferences based on human social sensing data, by, for instance, building on the methodology for deriving point estimates and associations137 from samples recruited via methods such as respondent-driven sampling138, successive sample size estimation139 and methods based on recruitment timing140,141, as well as on the methodology for correction of underreporting biases142. It is also important to better understand to what extent the reports of human social sensors are affected by different measurement errors143, to explore psychometric models to estimate their accuracy144,145, and to study their reliability and performance bounds45.

The accuracy of human social sensing can be further increased by providing truth-telling incentives with algorithms such as the Bayesian truth serum146, peer-prediction147, Bayesian markets148 or choice-matching149. These game-theoretic algorithms do not assume that honesty can be independently checked. Instead, they leverage the fact that people with different characteristics should make different predictions of the prevalence of these characteristics in the population. For example, Bayesian truth serum incentives have been used to estimate the prevalence of questionable research practices by academic psychologists150. Apart from increasing informational value at the individual level, truth serum scores can also be used to assess which sensors are more accurate detectors of true population distributions151.

Data from human social sensors can be usefully combined with existing administrative records and various digital trace data. Governments across the world are opening up administrative data sources for social science research152,153. The value of such data sources can be enhanced by asking human social sensors to provide the necessary context for people not covered by the records or to provide evaluations of relevant social interactions at a specific point and location in time154. Human social sensing can also be used to overcome frequent problems related to digital trace data, such as unknown populations of inference, missing covariates, and overall unclear measurement properties155,156,157,158.

Computational models of belief dynamics can be further improved by explicitly incorporating social sensing processes. Rather than assuming that everyone has the same representation of the social world, different people might be more or less likely to detect others’ beliefs, and beliefs about different issues might be more or less socially visible3,159,160. Without taking social sensing processes into account, one might inaccurately conclude that differences in the spread of beliefs stem from differences in objective social network structures, whereas in fact the differences might stem from what people perceive subjectively. While we show that social sensing is generally accurate, research on individual differences in this accuracy and differences in accuracy between topics is an important avenue for future research.

Human social sensing provides exciting opportunities for interactive machine learning, where it could inform the training data, predictions, and algorithms161,162,163. Human social sensors can detect early language signals of emerging societal trends, contained in words and phrases the meaning of which can be understood only by some parts of social networks. For example, while many algorithms exist to detect overt hate speech online164, more subtle covert signals (including metaphors, humour and memes) can be difficult to detect. Human social sensors who can interpret these covert signals can be selected using theory-driven predictions47. These subgroups of human social sensors could also help to illuminate the motivations of groups using hate speech, and help to detect, understand, and counter such ‘dangerous speech’165,166.

The ability of human social sensors to pick up on societal trends could be used in hybrid human–machine forecasting systems167,168,169,170 that combine machine forecasts with forecasts based on human social sensor data. Human social sensors can also contribute to the development of better machines and algorithms that affect all aspects of our lives171,172. Examples include news ranking algorithms, self-driving vehicles, algorithmic trading and pricing, online dating, and crime assessment173.

Summary

Developments within computational social science can help to integrate research in psychology and sociology in order to enable wider use of human social sensing across all social sciences. In turn, human social sensors can help social scientists to come closer to the next frontier in the study of human social systems174,175 and to achieve more rigorous theoretical understanding and predictions of real-world complex social phenomena.

As described above, human social sensors are likely to be particularly useful when asked about beliefs and behaviours in their immediate social environments, and when they are sampled to represent the population of interest and provide early indicators of emerging trends. While we have focused on reports collected in surveys, researchers can also use social media, apps, or wearables to prompt participants to report about their social environments at particular time points and places. Such data can serve as an important bridge between data on offline networks traditionally collected in sociometric surveys and online networks studied through data available from the social media.

Apart from its scientific interest, data from human social sensors can also be useful for policy interventions that are most likely to succeed given a current state of public needs and opinions. Especially after sudden incidents that require quick interventions (for example, mass shootings, environmental catastrophes, or public unrest), a smaller number of well selected human social sensors could quickly provide estimates about many other members of the same population. Human social sensors can be engaged through citizen science platforms, which were so far almost exclusively targeted towards the natural sciences176. Developing these platforms to collect data from human social sensors (for example, to report symptoms of contagious disease in their social circles177), possibly combined with machine learning algorithms, could be an effective way of forecasting societal trends and engaging the public in social science research.