Billions of Sketches Reveal Hidden Cultural Variation in Human Concepts

For decades, cognitive scientists have debated the extent to which human concepts are universal. Much of this debate has centered on linguistics. If people in different corners of the globe use similar words for "sun," does that imply they share a fundamental mental architecture? While large-scale textual analyses have suggested broad similarities, recent studies have yielded conflicting results. Some point to striking cultural variations in how we define emotions or food.

The central problem is that language is a "lossy" medium. Words act as powerful tools for communication. They do this by compressing vast, messy streams of sensory experience into shared, manageable abstractions. When we say "dog," we discard the specific texture of fur or the unique gait of an animal. We do this to reach a functional consensus. Consequently, studying only text may obscure the hidden differences in how cultures actually perceive the world.

The Compression Trap of Linguistic Universality

Current approaches to measuring conceptual universality rely on the statistical associations of words in massive online corpora. Researchers use word embedding models to map this. These are mathematical representations where words with similar meanings sit close together in a high-dimensional space. By doing this, they can map the "geometry" of human thought across languages. However, this method prioritizes shared linguistic conventions over the richness of experience.

As noted by the authors, if a concept is studied only through its linguistic label, we risk missing the "embodied" dimension of cognition. Embodied cognition is the theory that mental representations are grounded in sensory and motor experiences. For example, thinking about a "hammer" involves implicit mental simulations of grip and impact. Because language smooths over these sensory nuances, it may mask the cultural diversities researchers seek to uncover. Until now, the field lacked a dataset large enough to probe the visual imagination at a global scale.

Mapping the Latent Geometry of Sketches

To bypass the limitations of text, the researchers analyzed 2.6 billion sketches from the QuickDraw dataset. This data spans 236 countries and territories. Unlike words, which serve as compressed symbols, sketches provide a high-resolution probe of how people visually construe concepts. The authors used a multi-stage pipeline to decode this information:

1. Embedding Generation: Each sketch was processed through DINOv2, a self-supervised vision transformer (a neural network that learns visual features without human labels). This transformed raw pixels into 384-dimensional feature representations. These capture structural properties like shape and orientation.
2. Dimensionality Reduction: To make these complex shapes comparable, the researchers used PCA (Principal Component Analysis) and UMAP (Uniform Manifold Approximation and Projection). These techniques project high-dimensional data into a navigable 2D space.
3. Cluster Identification: Using DBSCAN (a density-based clustering algorithm that finds groups of points separated by low-density areas), the authors looked for "visual attractors." These are stable, recurrent forms used to represent a single concept.

The researchers found that concepts do not typically collapse into a single, universal prototype. Instead, they "unfold" into multiple distinct visual exemplars. As shown in, a concept like "phone" might split into several clusters representing landlines and smartphones.

A "fish" might separate into left-facing and right-facing orientations. The median number of these visual clusters per concept is 2. However, they range from a single form (like a "donut") to as many as 21 (for a "crow").

Sensory Experience and Cultural Divergence

The study reveals that these visual clusters are systematically linked to how we interact with the world. The authors measured "clusterability"—the proportion of sketches that fall into well-defined visual groups rather than being scattered as "noise." They compared this against various sensorimotor properties. They found that concepts involving haptic interaction (sensory experiences mediated by touch) tend to form significantly more coherent visual clusters .

The study also shows that visual geometry diverges sharply from linguistic geometry. The authors report a macro average rank correlation of only 0.098 between visual and semantic similarity rankings. This low number means that if two concepts look similar (like a pizza slice and a triangle), they are rarely semantically close in a word-based model.

When shifting the lens to entire nations, the divergence grows. The researchers built networks of countries based on how similarly they depict concepts. The sketch-based network revealed clear regional communities, such as South America, Europe, and Asia .

These were much harder to discern in the word-based network. Finally, when comparing these networks to the World Values Survey (a benchmark for global cultural distances), the authors found a significant result. Sketch-based similarity aligned 45% more closely with established cultural patterns than text-based measures did .

Limitations of the Digital Canvas

While the scale of 2.6 billion sketches is unprecedented, the findings have limits. The dataset is characterized by a heavy predominance of Anglophone users. The United States accounts for 41.3% of the sketches. This creates a potential bias. The visual representations of certain cultures may be over-represented. This could skew the perceived "global" norms.

Furthermore, the researchers note that participant location was inferred from IP addresses. This introduces uncertainty regarding true geographic origin. There is also a socio-economic dimension. Participation in an online sketching game requires digital literacy and device access. This likely biases the sample toward more privileged cohorts. Lastly, the 20-second time limit in QuickDraw may constrain sketch complexity. This might force users toward simplified, "iconic" representations.

The Verdict: A Multimodal Mandate

The evidence suggests that the quest for "universal concepts" depends on the measurement modality. If we look only at words, the world appears remarkably unified. If we look at images, the world appears more diverse. The findings indicate that language acts as a filter. It suppresses the cultural textures we seek to understand.

For practitioners in artificial intelligence, the takeaway is clear. If we want Large Language Models (LLMs) to grasp the nuance of human culture, text alone is insufficient. To model human cognition accurately, we must integrate multimodal data. This means incorporating visual and sensory-related imagery into training. The next frontier of intelligence lies in the interplay between what we say and what we see.

Figures from the paper