Billions of Sketches Reveal Hidden Cultural Variation in Human Concepts
This lightning talk explores how over 2.6 billion crowd-sourced sketches from 236 countries expose fundamental gaps in our understanding of human concepts. By comparing visual representations directly against language-based embeddings and cultural similarity data, the research reveals that embodied, visual concept structure diverges sharply from what language captures—and aligns far better with real cultural variation. The findings challenge text-only AI models and reframe long-standing debates about conceptual universality in cognitive science.Script
Can you tell which country someone is from by how they draw a cat? Over 2.6 billion sketches from 236 countries suggest you can, and that language has been hiding this diversity all along.
Most concepts don't have a single visual form. The researchers found that the median concept has two distinct clusters of visual exemplars, with some like crow splitting into 21 different attractors. Concepts we touch and manipulate with our hands, those high in haptic and sensorimotor grounding, cluster into the most stable, consistent forms across people and cultures.
When the authors compared visual sketch embeddings to multilingual word embeddings, they found the two spaces barely agree. The rank correlation between image-based and language-based concept similarity is just 0.098. A pizza slice and a whole pizza, visually distinct clusters, are not semantic neighbors in language space. Language compresses what vision reveals.
Visual concept networks recover intuitive cultural communities. English-speaking countries, South America, Europe, and East Asia emerge as coherent clusters when similarity is defined by shared sketch patterns. The same analysis using language embeddings yields weak, incoherent structure. Image-based conceptual proximity predicts cultural similarity, measured via the World Values Survey, 45 percent better than language does.
The dataset is enormous but not perfectly balanced. It skews toward internet-enabled, anglophone populations, with limited demographic control and potential noise from users gaming the platform. Still, the scale affords robust detection of stable, population-level patterns that language alone would miss.
If language-only models can't capture the conceptual diversity present in how people draw, they can't fully model human thought. This work calls for multimodal architectures trained on sketches, gestures, and embodied traces, not just text. To dive deeper into this research and create your own video summaries, visit EmergentMind.com.
