State of the ArtA novel lens on welfare and development through artwork
Throughout history, painters have produced visual representations that reflect the social, political, and economic conditions of their time, offering interpretations of ritual and daily life, power and protest, and poverty and prosperity—often in contexts where other forms of data are sparse or absent. This website leverages new algorithmic methods to extract sentiment and material features from a corpus of 627,369 paintings by 32,670 artists.
For the best experience, we recommend exploring the website on a computer rather than on a phone.
To extract contextual information embedded in historical paintings, we exploit open-access repositories from Google Arts and Culture (138,396 paintings), WikiArt (263,125 paintings), and Wikidata (225,848 paintings).
Notes: Panel (a) shows the distribution of production years in Google Arts and Culture, Wiki-Art, and Wiki-Data. Panel (b) shows the relative coverage over time for the following "harmonized" countries and country groups: Austria (AT, red), Belgium (BE, gold), Denmark (DK, light red), France (FR, navy blue), Germany (DE, black), Italy (IT, green), Japan (JP, pink), the Netherlands (ND, orange), Russia (RU, light blue), Spain (SP, yellow), the United Kingdom (UK, red), the United States (US, blue), other major European countries with more than 1,000 artworks (O1, e.g., Portugal, Sweden), other major Asian countries (O2, e.g., China, India), other major producers from other continents (O3, e.g., Argentina, Australia, Brazil, Mexico, South Africa), and smaller producers (O4, all countries with fewer than 1,000 artworks).
Both our emotion and material representation classifications rely on a state-of-the-art Convolutional Neural Network (CNN) architecture.
The sentiment classifier exploits the empirical distribution of about 20 emotions labels per artwork, {qe}e∈ℰ—fear, sadness, anger, disgust, awe, contentment, amusement, excitement, and a residual "other" category, complemented with textual justifications—from Achlioptas et al. (2021) and Mohamed et al. (2022) to best fit a predicted probability distribution, {pe}e∈ℰ. The model minimizes the Kullback-Leibler divergence metric between the target probability distribution, {qe}e∈ℰ, and the predicted probability distribution, {pe}e∈ℰ,
The extraction of visual features relevant to material living standards relies on a two-stage procedure. The calibration stage exploits a large language model to generate (a) a wealth index on a 0-10 scale, (b) a relevance index on a 0-10 scale, and (c) a structured list of visual elements conveying information on living standards—such as the quality and abundance of objects, the physical appearance of characters, and contextual cues about social positioning. The prediction stage uses measures (a) and (b) as targets to train an in-house algorithm that jointly predicts the wealth and relevance indices. As in the sentiment pipeline—where annotators' justifications inform our understanding of their choices—the structured artwork descriptions (c) provide insight into the mechanisms underlying the inference of living standards.
Landscape Mode
This section is available only in landscape mode. Please rotate your device.
Dynamics of emotions at the country level
Select countries and time periods to explore how emotional expression in art has evolved over time. The chart shows the likelihood that paintings are associated with specific emotions across different periods. Use the legend to highlight or filter emotions, and observe how artistic sentiment shifts across eras and regions.
Our algorithms produce (i) a probabilistic vector of emotions, pj, and (ii) living standard indicators (and informativeness measures, zj), wj, for each painting j ∈ 𝒥. We can exploit these painting-specific characteristics to examine emotional and material welfare through socioeconomic transformations.
Panel (a) shows that a 169% increase in GDP per capita—equivalent to sustained annual growth of 2% over 50 years—is associated with a 0.50 standard deviation decline in sadness, alongside moderate increases in amusement, excitement, and awe. These sizable effects do not, however, capture the full emotional response to changes in living standards: conveyed emotions also react strongly to other dimensions of the economic environment, notably uncertainty and inequality.
Panel (b) shows that economic uncertainty triggers a very distinct emotional pattern. In uncertain times, artists depict higher levels of unstable emotions (fear and amusement) and lower levels of stable emotions such as contentment.
Finally, Panel (c) shows that wealth inequality primarily influences the dispersion of emotions. For instance, the share of total wealth held by the top 1% in Europe declined by roughly 40 percentage points between the late nineteenth century and 1980, before rising again by about 10 points in recent decades. Such a 10-percentage-point increase in wealth concentration would be associated with a very substantial 0.21 standard deviation increase in the disagreement index.
Notes: These figures display the estimates of regressions relating the predicted emotions of a given painting, {pe}e∈ℰ, to indicators of economic development and inequality: (a) log GDP per capita; (b) the volatility of GDP per capita calculated over a window of [-2,+4] years; and (c) the concentration of wealth among the top-1% wealthiest individuals, mostly available from the mid-nineteenth century onward. The coefficients represent separate regressions with the following left-hand side variables: the predicted probability of expressing fear (dark-blue, plain circle), sadness (light-blue, plain square), disgust (teal, plain diamond), anger (black, plain triangle), other (gray, cross), awe (gold, hollow square), contentment (salmon, hollow triangle), amusement (pink, hollow diamond), and excitement (orange, hollow circle); a positivity index in light blue; and a measure of disagreement. All left-hand side variables are standardized, so the reported estimates can be interpreted in terms of standard deviations across country/year observations. The bands represent 10%, 5%, and 1% confidence intervals.
Landscape Mode
This section is available only in landscape mode. Please rotate your device.
Dynamics of emotions at the artist level
Search for painters and explore patterns in their emotional expression using the interactive visualization below. The radar chart compares how different emotions are represented in their work relative to the global average. The emotion timeline traces how emotional expression evolves over the course of their career—click the emotion buttons to add or remove them from the plot. The darker underlying curve indicates periods of peak productivity, showing the volume of work produced at each age.
*This selection is discretionary and will be expanded over time.
Exploration
Explore how material living standards and emotional sentiment are depicted in artworks across countries, time periods, and artists. Create overlaid histograms to identify patterns and divergences in the representation of prosperity and optimism across painters, cultures, and centuries
Filters
Positivity distribution (-1 to 1)
Wealth distribution (0 to 10)
Plots
Methodology
Authors
Website design
Contact us
Fill out the form below if you want to contact us (e.g., to report an issue, suggest additional visualizations, or mention a relevant reference).