Social trust is a cornerstone of prosperous and stable societies, contributing to economic growth, reduced crime, and more inclusive institutions. However, understanding the origins and evolution of social trust is challenging due to the difficulty of documenting it historically. While historical observations suggest a steady rise in social trust in Europe since the early modern period, evidenced by increased religious tolerance, the decline of witch hunts and honor killings, and the growth of intellectual freedom, quantitative evidence remains limited. Traditional methods, like examining etiquette manuals or legal changes, offer only partial insights. This study offers a novel approach by leveraging the rich visual data embedded within historical portraits. Recent research demonstrates that specific facial cues, such as smiling or wider eyes, consistently signal perceived trustworthiness across cultures and individuals. Building on this, the researchers designed an algorithm using machine learning techniques to quantify perceived trustworthiness from facial features depicted in European portraits. This automated approach provides a quantitative proxy for social trust throughout history, allowing for a large-scale analysis of this important social phenomenon.

The study draws upon extensive research in social cognition demonstrating the consistent use of facial features as cues for assessing trustworthiness. Experimental work has established that specific facial action units (muscle contractions associated with expressions) reliably predict perceived trustworthiness judgments. This research establishes a foundation for the development of an algorithm that can automatically assess perceived trustworthiness from images of faces. The authors also cite work showing how cultural artifacts, like paintings and literature, reflect societal preferences and underlying mentalities. Studies have shown that painters tend to exaggerate features associated with friendly faces, and that viewers prefer portraits with certain gaze patterns. This body of work suggests that cultural artifacts can serve as valuable indicators of shifts in social attitudes and preferences.

The researchers developed a machine-learning algorithm to estimate perceived trustworthiness from facial cues in portraits. The algorithm was trained on a dataset of avatars generated to exhibit varying levels of perceived trustworthiness and dominance. A random forest procedure was used for optimization. To validate the algorithm, the researchers tested its performance on four databases of natural faces, comparing its output to human ratings. The algorithm's validity was further assessed by its ability to reproduce known biases in human perception, such as the tendency to rate younger, feminine, and happy faces as more trustworthy. The algorithm was then applied to two large datasets of historical portraits: the National Portrait Gallery (NPG) database (1962 English portraits from 1505 to 2016) and the Web Gallery of Art (WGA) database (4106 portraits from 19 Western European countries between 1360 and 1918). To control for potential confounding factors, perceived dominance was also estimated and controlled for in all analyses. Furthermore, the algorithm was applied to a database of selfies (SelfieCity, N=2277) to assess whether trustworthiness displayed in portraits correlated with actual levels of social trust in individuals' environments. Analyses were conducted using linear models controlling for sitter's gender, age, and perceived dominance. GDP per capita and a democracy index (Polity2) were used as potential predictors of perceived trustworthiness fluctuations. Time-lag analyses examined the temporal dynamics between trustworthiness ratings and GDP per capita or institutional changes.

The analysis of both the NPG and WGA portrait datasets revealed a significant increase in perceived trustworthiness ratings over time (NPG: b=0.14, z=7.49, p<0.001; WGA: b=0.07, z=5.33, p<0.001). This increase was consistent with historical accounts of a “Smile Revolution” and a rise in prosocial displays. The analysis of selfies from the SelfieCity database showed a positive correlation between perceived trustworthiness in selfies and levels of interpersonal trust and cooperation in the individuals' respective cities. Further analysis revealed a strong association between higher GDP per capita and higher levels of perceived trustworthiness in portraits in both datasets (NPG: b=0.03, z=7.13, p<0.001; WGA: b=0.09, z=3.16, p=0.002). The association remained significant even after adjusting for the effect of time. GDP per capita proved to be a better predictor of trustworthiness changes than the Polity2 democratization index. Time-lag analyses indicated that changes in GDP per capita predicted future changes in perceived trustworthiness in portraits approximately two decades later, but not vice-versa, suggesting that economic changes might precede shifts in perceived trustworthiness. Similar analyses using the number of book titles per capita as a proxy for affluence also supported the relationship between affluence and trustworthiness.

The findings suggest a significant link between economic affluence and perceived trustworthiness in historical portraits. The consistent rise in perceived trustworthiness across two independent datasets supports the idea of a genuine societal shift towards higher social trust. This shift is more strongly associated with economic growth (GDP per capita) than with institutional changes (democratization), implying that economic prosperity may play a critical role in shaping social trust. The results align with previous research highlighting the impact of economic factors on social attitudes and behaviors. The study's innovative methodology, employing machine learning to analyze facial expressions in historical portraits, opens new avenues for quantitative research on the history of mentalities and social preferences. The findings complement existing qualitative historical accounts and enhance our understanding of the intricate relationship between economic development and social trust.

This study demonstrates, using a novel machine learning approach, a significant increase in perceived trustworthiness in Western European portraits between 1500 and 2000. This rise correlates strongly with increased affluence, as measured by GDP per capita, rather than solely with institutional changes. This research provides quantitative support for qualitative historical observations regarding the growth of social trust and suggests that future research should focus on exploring the causal mechanisms underlying this relationship and examining potential regional or social class variations in this trend. The methodology developed offers a powerful tool for analyzing historical data in the social sciences.

The study acknowledges several limitations. First, the portrait datasets may not fully represent the broader population, potentially over-representing the wealthy elite. Second, the assumption that facial cues related to trustworthiness remained constant across the study period needs further investigation. Third, historical economic indicators, such as GDP per capita, may not fully capture the nuances of wealth distribution and living standards. These limitations need to be considered when interpreting the results.