Online images amplify gender bias.

Each year, people spend less time reading and more time viewing images1, which are proliferating online2-4. Images from platforms such as Google and Wikipedia are downloaded by millions every day2,5,6, and millions more are interacting through social media, such as Instagram and TikTok, that primarily consist of exchanging visual content. In parallel, news agencies and digital advertisers are increasingly capturing attention online through the use of images7,8, which people process more quickly, implicitly and memorably than text9-12. Here we show that the rise of images online significantly exacerbates gender bias, both in its statistical prevalence and its psychological impact. We examine the gender associations of 3,495 social categories (such as 'nurse' or 'banker') in more than one million images from Google, Wikipedia and Internet Movie Database (IMDb), and in billions of words from these platforms. We find that gender bias is consistently more prevalent in images than text for both female- and male-typed categories. We also show that the documented underrepresentation of women online13-18 is substantially worse in images than in text, public opinion and US census data. Finally, we conducted a nationally representative, preregistered experiment that shows that googling for images rather than textual descriptions of occupations amplifies gender bias in participants' beliefs. Addressing the societal effect of this large-scale shift towards visual communication will be essential for developing a fair and inclusive future for the internet.

Experimental evidence for structured information-sharing networks reducing medical errors.

Errors in clinical decision-making are disturbingly common. Recent studies have found that 10 to 15% of all clinical decisions regarding diagnoses and treatment are inaccurate. Here, we experimentally study the ability of structured information-sharing networks among clinicians to improve clinicians' diagnostic accuracy and treatment decisions. We use a pool of 2,941 practicing clinicians recruited from around the United States to conduct 84 independent group-level trials, ranging across seven different clinical vignettes for topics known to exhibit high rates of diagnostic or treatment error (e.g., acute cardiac events, geriatric care, low back pain, and diabetes-related cardiovascular illness prevention). We compare collective performance in structured information-sharing networks to collective performance in independent control groups, and find that networks significantly reduce clinical errors, and improve treatment recommendations, as compared to control groups of independent clinicians engaged in isolated reflection. Our results show that these improvements are not a result of simple regression to the group mean. Instead, we find that within structured information-sharing networks, the worst clinicians improved significantly while the best clinicians did not decrease in quality. These findings offer implications for the use of social network technologies to reduce errors among clinicians.

Social learning and partisan bias in the interpretation of climate trends.

Vital scientific communications are frequently misinterpreted by the lay public as a result of motivated reasoning, where people misconstrue data to fit their political and psychological biases. In the case of climate change, some people have been found to systematically misinterpret climate data in ways that conflict with the intended message of climate scientists. While prior studies have attempted to reduce motivated reasoning through bipartisan communication networks, these networks have also been found to exacerbate bias. Popular theories hold that bipartisan networks amplify bias by exposing people to opposing beliefs. These theories are in tension with collective intelligence research, which shows that exchanging beliefs in social networks can facilitate social learning, thereby improving individual and group judgments. However, prior experiments in collective intelligence have relied almost exclusively on neutral questions that do not engage motivated reasoning. Using Amazon's Mechanical Turk, we conducted an online experiment to test how bipartisan social networks can influence subjects' interpretation of climate communications from NASA. Here, we show that exposure to opposing beliefs in structured bipartisan social networks substantially improved the accuracy of judgments among both conservatives and liberals, eliminating belief polarization. However, we also find that social learning can be reduced, and belief polarization maintained, as a result of partisan priming. We find that increasing the salience of partisanship during communication, both through exposure to the logos of political parties and through exposure to the political identities of network peers, can significantly reduce social learning.

Divergences in color perception between deep neural networks and humans.

Deep neural networks (DNNs) are increasingly proposed as models of human vision, bolstered by their impressive performance on image classification and object recognition tasks. Yet, the extent to which DNNs capture fundamental aspects of human vision such as color perception remains unclear. Here, we develop novel experiments for evaluating the perceptual coherence of color embeddings in DNNs, and we assess how well these algorithms predict human color similarity judgments collected via an online survey. We find that state-of-the-art DNN architectures - including convolutional neural networks and vision transformers - provide color similarity judgments that strikingly diverge from human color judgments of (i) images with controlled color properties, (ii) images generated from online searches, and (iii) real-world images from the canonical CIFAR-10 dataset. We compare DNN performance against an interpretable and cognitively plausible model of color perception based on wavelet decomposition, inspired by foundational theories in computational neuroscience. While one deep learning model - a convolutional DNN trained on a style transfer task - captures some aspects of human color perception, our wavelet algorithm provides more coherent color embeddings that better predict human color judgments compared to all DNNs we examine. These results hold when altering the high-level visual task used to train similar DNN architectures (e.g., image classification versus image segmentation), as well as when examining the color embeddings of different layers in a given DNN architecture. These findings break new ground in the effort to analyze the perceptual representations of machine learning algorithms and to improve their ability to serve as cognitively plausible models of human vision. Implications for machine learning, human perception, and embodied cognition are discussed.

An Online experiment during the 2020 US-Iran crisis shows that exposure to common enemies can increase political polarization.

A longstanding theory indicates that the threat of a common enemy can mitigate conflict between members of rival groups. We tested this hypothesis in a pre-registered experiment where 1670 Republicans and Democrats in the United States were asked to complete an online social learning task with a bot that was labeled as a member of the opposing party. Prior to this task, we exposed respondents to primes about (a) a common enemy (involving Iran and Russia); (b) a patriotic event; or (c) a neutral, apolitical prime. Though we observed no significant differences in the behavior of Democrats as a result of priming, we found that Republicans-and particularly those with very strong conservative views-were significantly less likely to learn from Democrats when primed about a common enemy. Because our study was in the field during the 2020 Iran Crisis, we were able to further evaluate this finding via a natural experiment-Republicans who participated in our study after the crisis were even less influenced by the beliefs of Democrats than those Republicans who participated before this event. These findings indicate common enemies may not reduce inter-group conflict in highly polarized societies, and contribute to a growing number of studies that find evidence of asymmetric political polarization in the United States. We conclude by discussing the implications of these findings for research in social psychology, political conflict, and the rapidly expanding field of computational social science.

Topological measures for identifying and predicting the spread of complex contagions.

The standard measure of distance in social networks - average shortest path length - assumes a model of "simple" contagion, in which people only need exposure to influence from one peer to adopt the contagion. However, many social phenomena are "complex" contagions, for which people need exposure to multiple peers before they adopt. Here, we show that the classical measure of path length fails to define network connectedness and node centrality for complex contagions. Centrality measures and seeding strategies based on the classical definition of path length frequently misidentify the network features that are most effective for spreading complex contagions. To address these issues, we derive measures of complex path length and complex centrality, which significantly improve the capacity to identify the network structures and central individuals best suited for spreading complex contagions. We validate our theory using empirical data on the spread of a microfinance program in 43 rural Indian villages.

Experimental evidence for scale-induced category convergence across populations.

Individuals vary widely in how they categorize novel and ambiguous phenomena. This individual variation has led influential theories in cognitive and social science to suggest that communication in large social groups introduces path dependence in category formation, which is expected to lead separate populations toward divergent cultural trajectories. Yet, anthropological data indicates that large, independent societies consistently arrive at highly similar category systems across a range of topics. How is it possible for diverse populations, consisting of individuals with significant variation in how they categorize the world, to independently construct similar category systems? Here, we investigate this puzzle experimentally by creating an online "Grouping Game" in which we observe how people in small and large populations collaboratively construct category systems for a continuum of ambiguous stimuli. We find that solitary individuals and small groups produce highly divergent category systems; however, across independent trials with unique participants, large populations consistently converge on highly similar category systems. A formal model of critical mass dynamics in social networks accurately predicts this process of scale-induced category convergence. Our findings show how large communication networks can filter lexical diversity among individuals to produce replicable society-level patterns, yielding unexpected implications for cultural evolution.

Probabilistic social learning improves the public's judgments of news veracity.

The digital spread of misinformation is one of the leading threats to democracy, public health, and the global economy. Popular strategies for mitigating misinformation include crowdsourcing, machine learning, and media literacy programs that require social media users to classify news in binary terms as either true or false. However, research on peer influence suggests that framing decisions in binary terms can amplify judgment errors and limit social learning, whereas framing decisions in probabilistic terms can reliably improve judgments. In this preregistered experiment, we compare online peer networks that collaboratively evaluated the veracity of news by communicating either binary or probabilistic judgments. Exchanging probabilistic estimates of news veracity substantially improved individual and group judgments, with the effect of eliminating polarization in news evaluation. By contrast, exchanging binary classifications reduced social learning and maintained polarization. The benefits of probabilistic social learning are robust to participants' education, gender, race, income, religion, and partisanship.

The reduction of race and gender bias in clinical treatment recommendations using clinician peer networks in an experimental setting.

Bias in clinical practice, in particular in relation to race and gender, is a persistent cause of healthcare disparities. We investigated the potential of a peer-network approach to reduce bias in medical treatment decisions within an experimental setting. We created "egalitarian" information exchange networks among practicing clinicians who provided recommendations for the clinical management of patient scenarios, presented via standardized patient videos of actors portraying patients with cardiac chest pain. The videos, which were standardized for relevant clinical factors, presented either a white male actor or Black female actor of similar age, wearing the same attire and in the same clinical setting, portraying a patient with clinically significant chest pain symptoms. We found significant disparities in the treatment recommendations given to the white male patient-actor and Black female patient-actor, which when translated into real clinical scenarios would result in the Black female patient being significantly more likely to receive unsafe undertreatment, rather than the guideline-recommended treatment. In the experimental control group, clinicians who were asked to independently reflect on the standardized patient videos did not show any significant reduction in bias. However, clinicians who exchanged real-time information in structured peer networks significantly improved their clinical accuracy and showed no bias in their final recommendations. The findings indicate that clinician network interventions might be used in healthcare settings to reduce significant disparities in patient treatment.

Networked collective intelligence improves dissemination of scientific information regarding smoking risks.

Despite substantial investments in public health campaigns, misunderstanding of health-related scientific information is pervasive. This is especially true in the case of tobacco use, where smokers have been found to systematically misperceive scientific information about the negative health effects of smoking, in some cases leading smokers to increase their pro-smoking bias. Here, we extend recent work on 'networked collective intelligence' by testing the hypothesis that allowing smokers and nonsmokers to collaboratively evaluate anti-smoking advertisements in online social networks can improve their ability to accurately assess the negative health effects of tobacco use. Using Amazon's Mechanical Turk, we conducted an online experiment where smokers and nonsmokers (N = 1600) were exposed to anti-smoking advertisements and asked to estimate the negative health effects of tobacco use, either on their own or in the presence of peer influence in a social network. Contrary to popular predictions, we find that both smokers and nonsmokers were surprisingly inaccurate at interpreting anti-smoking messages, and their errors persisted if they continued to interpret these messages on their own. However, smokers and nonsmokers significantly improved in their ability to accurately interpret anti-smoking messages by sharing their opinions in structured online social networks. Specifically, subjects in social networks reduced the error of their risk estimates by over 10 times more than subjects who revised solely based on individual reflection (p < 0.001, 10 experimental trials in total). These results suggest that social media networks may be used to activate social learning that improves the public's ability to accurately interpret vital public health information.

Color associations in abstract semantic domains.

The embodied cognition paradigm has stimulated ongoing debate about whether sensory data - including color - contributes to the semantic structure of abstract concepts. Recent uses of linguistic data in the study of embodied cognition have been focused on textual corpora, which largely precludes the direct analysis of sensory information. Here, we develop an automated approach to multimodal content analysis that detects associations between words based on the color distributions of their Google Image search results. Crucially, we measure color using a transformation of colorspace that closely resembles human color perception. We find that words in the abstract domains of academic disciplines, emotions, and music genres, cluster in a statistically significant fashion according to their color distributions. Furthermore, we use the lexical ontology WordNet and crowdsourced human judgments to show that this clustering reflects non-arbitrary semantic structure, consistent with metaphor-based accounts of embodied cognition. In particular, we find that images corresponding to more abstract words exhibit higher variability in colorspace, and semantically similar words have more similar color distributions. Strikingly, we show that color associations often reflect shared affective dimensions between abstract domains, thus revealing patterns of aesthetic coherence in everyday language. We argue that these findings provide a novel way to synthesize metaphor-based and affect-based accounts of embodied semantics.