Author Correction: Tracking historical changes in perceived trustworthiness in Western Europe using machine learning analyses of facial cues in paintings

,


Introduction
The original version of this Article contained an error in the fourth sentence of the second paragraph of the Introduction in which the region of the sample of portraits was omitted. The correct sentence states 'More precisely, we apply recent machine-learning methods to extract quantitative information about the evolution of social cues contained in Western European portraits.' instead of 'More precisely, we apply recent machine-learning methods to extract quantitative information about the evolution of social cues contained in portraits. ' We also added two clarifying sentences from the second paragraph of the introduction which read 'Crucially, this algorithm does not provide information on a person's face but rather on the way this face is likely to be perceived by others, based on a specific image. Indeed, first impressions from faces are highly sensitive to factors such as variations in lighting and pose.' We added a clarifying sentence at the end of the Introduction which reads 'In this article, all occurrences of the words 'trustworthiness' and 'dominance' refer to subjective perceptions of trustworthiness and dominance from faces and not to individuals' actual level of trustworthiness or dominance.' The original version of this Article contained an error in the fifth sentence of the second paragraph of the introduction which omitted to clarify how the model was built, and incorrectly referred to 'trustworthiness ratings on portraits' rather than 'ratings of perceived trustworthiness'. The correct sentence states 'The algorithm is built on models of human perception of faces to generate automatic human-like ratings of perceived trustworthiness based on the muscle contractions (facial action units) detected in facial displays in portraits using the open software OpenFace.' instead of 'The algorithm generates automatic human-like trustworthiness ratings on portraits based on the muscle contractions (facial action units) detected in facial displays using the open software OpenFace'.
The original version of this Article contained an error in the seventh sentence of the second paragraph of the abstract which incorrectly described the avatars used for training the algorithm as being 'controlled for trustworthiness' rather than 'generated to display varying levels of perceived trustworthiness'. The correct sentence states 'This algorithm was trained on avatars generated to display varying levels of perceived trustworthiness and optimized using a random forest procedure (see Supplementary Methods for more details).' Instead of 'This algorithm was trained on avatars controlled for trustworthiness and optimized using a random forest procedure (see Supplementary Methods for more details).'

Results
The fifth sentence of the first paragraph of the results section omitted to note the small size in the increase of perceived trustworthiness. The correct sentence states 'Although the increase of perceived trustworthiness is small, these results are consistent with more qualitative works documenting a so-called 'Smile Revolution' 27 and a rise of prosocial displays in paintings and in novels 28 .' instead of 'Overall, these results are consistent with more qualitative works documenting a so-called 'Smile Revolution' 25 and a rise of prosocial displays in paintings and in novels 26 .' We now include in the second section of the Results a clarifying paragraph to discuss effect sizes and cite related work. The following text has been added along with references 46-51 to the Reference list.
'These results provide evidence in favor of the association between economic wealth and social trust at the society level. However, due to the small effect sizes and the limitations of the historical economic indicators 46,47 , as well as to the fact that GDP per capita is only a partial measure of wealth (which does not account, for example, for inequalities in wealth distribution 48 ), we replicated our analyses with an alternative variable known to be associated with countries' wealth: the number of book titles per capita. Indeed, although the number of book titles per capita is thought to be linked to human development variables, it has also been shown to be associated with national income 48-51 . Supporting the analyses conducted with GDP per capita, we found a significant positive association between the number of book titles per capita and the level of perceived trustworthiness in the portraits of the National Portrait Gallery (affluence only model: b = 0.35 ± 0.06, z = 6.15, p < 0.001; model controlling for time: b = 0.21 ± .06, z = 3.45, p = 0.001) and of the Web Gallery of Art, although not robust to the inclusion of time in this latter case (affluence only model: b = 0.29 ± 0.10, z = 2.77, p = 0.006; model controlling for time: b = 0.14 ± 0.11, z = 1.26, p = 0.208).' The full details of the added references are as follows.

Discussion
The original version of the Article omitted clarifying information from the discussion. The added section reads as follows: 'The algorithm was built to estimate how human raters would rate the perceived trustworthiness of faces. It can be used in scientific research for this purpose. The algorithm does not quantify the actual trustworthiness of an individual, and was not intended for this purpose.' The original version of the Article omitted a section to discuss limitations of the study in the Discussion. The added section reads as follows: 'At this point, it is important to note the small correlation between the perceived trustworthiness ratings provided by human raters and those retrieved by our algorithm. However, this small effect size is to be expected. First, the avatars on which the algorithm was trained did not represent the texture of the faces, even though this information may influence human raters' evaluations. Similarly, the avatars are bold and our algorithm is thus blind to haircut, even though these cues are known to influence first impressions from faces (see e.g., 21 ). Finally, our algorithm was trained to generate ratings of perceived trustworthiness based on the facial features that represent the shared component of first impressions from faces. Indeed, individuals rely on both shared and idiosyncratic features when forming a first impression on a new face, and our algorithm was designed to produce scores only based on the former.
Finally, several limitations are to be noted. First, one cannot assume that the evolution of perceived trustworthiness depicted in this study extends to the larger population of the period. The phenomenon described in this article might, for instance, be limited to the relatively elite, wealthy population represented in the portraits. In line with this possibility, there is evidence that social attitudes can vary with socioeconomic status 55-58 . Second, our study is based on the assumption that facial cues that are used as cues to assess perceived trustworthiness are shared across time. Although recent evidence 59-61 points towards such a stability, further work is needed to fully test this assumption. Third, times series of GDP per capita and living standards are only estimates, and their precision may fluctuate throughout the studied time period and fail to fully capture the evolution of living standards and inequalities 46-48 .'

Methods
The fourth section of the Methods section omitted a clarifying sentence. The added sentence reads 'We limited our analysis to paintings, excluding other medium types at the National Portrait Gallery, such as drawings, sculptures and photographs. In addition, only portraits for which the image was available on the website of the National Portrait Gallery were analyzed (3152 over 3161 paintings).' The ninth sentence of the first paragraph of the fourth section of the Methods section contained an error in which a clarifying phrase was omitted 'however we did not control for the provider of the portraits (e.g., purchased, transferred from another museum or given by a private donator).' The correct sentence reads 'Importantly, in order to ensure that the portraits accurately reflected the level of trust at the time the portrait was painted and to avoid re-interpretation of past historical figures, only portraits painted during the sitter's lifetime were analyzed (number of analyzed portraits: N = 1962), however we did not control for the provider of the portraits (e.g., purchased, transferred from another museum or given by a private donator).' instead of 'Importantly, in order to ensure that the portraits accurately reflected the level of trust at the time the portrait was painted and to avoid re-interpretation of past historical figures, only portraits painted during the sitter's lifetime were analyzed (number of analyzed portraits: N = 1962).'

Figure legends
The Figure 1 legend contained an error in part B in which 'displays of 'trustworthiness' was used instead of 'perceived trustworthiness', dominance was used instead of 'perceived dominance' and omitted a clarifying phrase 'for representation purposes, in this Figure, evaluation of perceived trustworthiness value was fitted by a local polynomial regression with a span of 0.75 and'.
The correct sentence reads 'B. Evolution of ratings of perceived trustworthiness in the National Portrait Gallery (modeled for representation purposes, in this Figure, evaluation of perceived trustworthiness value was fitted by a local polynomial regression with a span of 0.75 and adjusted for perceived dominance) and GDP per capita in England. (log-transformed for representation purposes).' instead of 'Evolution of displays of trustworthiness in the National Portrait Gallery (modeled trustworthiness value adjusted for dominance) and GDP per capita in England.'