The Elephant Looks Different in Every Country: What 2.6 Billion Sketches Reveal About Hidden Cultural Diversity in Human Thought
The Elephant in the Room Looks Different Depending on Where You're From
Walk into any schoolroom on Earth and ask children to draw an elephant. In Nigeria, you might get a trunk pointing skyward, tusks curving toward the earth. In Japan, you might see a smaller animal, compact and cute, with big ears that dwarf the rest of the body. In rural Wisconsin, you might get a gray blob with a trunk—and ears so large they nearly fill the page. None of these drawings are wrong. But they are not the same.
This is not a trivial observation. It is, according to a landmark new study, a window into something profound about how human beings think—and how much of that thinking our words systematically hide from us.
Researchers have just analyzed 2.6 billion human-drawn sketches—perhaps the largest cognitive dataset ever assembled—spanning 236 countries and territories across every inhabited continent. What they found upends a longstanding assumption in psychology and linguistics: that human concepts are fundamentally universal, shared, stable across cultures. The team, led by Arianna Pera, Mauro Martino, and colleagues at the MIT Media Lab, discovered that when you ask people to picture a concept rather than name it, the universality evaporates. Concepts that looked identical across languages through the lens of words turn out to have rich, textured variation when you trace them through visual imagination.
The finding is not merely academic. It suggests that everything from international diplomacy to AI design to cross-cultural psychology may have been operating on an incomplete map of human thought—one that ignored the silent, visual dimension where many of our deepest cultural differences actually live.
"Words are effective as communication devices because they compress rich experiential variation into shared conventions," the researchers write. "This potentially obscures hidden individual and cultural differences in how concepts are mentally represented." In other words: language is a remarkably efficient tool for agreeing on labels. It is a terrible tool for capturing what those labels actually mean inside different heads.
The study's most striking number is this: when they compared how well different methods predicted the actual cultural distances between countries—how different two cultures really are from each other—sketch-based representations outperformed text-based ones by 45 percent. Nearly half again as accurate. Imagine you were trying to map the cultural landscape of the world. You could use the words people use, which has been the gold standard for decades. Or you could use the pictures people draw. The pictures, it turns out, tell a richer, more honest story.
The Science
To understand what Pera and her colleagues actually did, you need to appreciate the dataset they were working with—and the extraordinary ambition of the project.
The researchers drew on data from Quick, Draw!, a game developed by Google that challenges players to sketch objects as quickly as possible while an AI tries to guess what they're drawing. The game was launched in 2016 and quickly became a global phenomenon. By the time the researchers accessed the data, it contained over 2.6 billion sketches spanning more than 50 million unique drawing sessions. The geographical footprint was remarkable: 236 countries and territories contributed drawings, making it arguably the most culturally diverse cognitive dataset ever compiled.
This is not a curated sample of carefully selected participants. It is the messy, glorious, global cacophony of human visual imagination, captured at scale.
The team then applied a suite of computational techniques to extract meaning from this vast corpus. Each sketch was converted into a high-dimensional vector—essentially, a mathematical fingerprint that captures the essential features of the drawing. Sketches of the same concept that looked similar would produce similar vectors; sketches that looked different would produce vectors far apart in this mathematical space. This technique, borrowed from natural language processing, is called embedding—and it allows researchers to measure conceptual distance computationally.
For any given concept (say, "airplane"), the researchers could now ask: how many distinct visual "types" of airplane drawings exist in the data? Do people in Japan draw airplanes the same way people in Brazil do? Do certain kinds of concepts show more visual variation than others?
They compared their sketch-based embeddings with word embedding models—sophisticated statistical tools that capture the meaning of words based on how they are used in text across millions of documents. If human concepts are truly universal, these two different ways of measuring them should converge. The same elephant should occupy the same mental space whether you find it through words or pictures.
They did not converge.
To measure cross-cultural similarity—how similar are the conceptual structures of two countries—the researchers used a technique called Procrustes analysis, which asks: if you rotate and scale two different datasets to match each other as closely as possible, how much overlap remains? This gives a mathematical measure of similarity between cultural concept spaces. They then compared this to established measures of cultural distance derived from the literature—geographic distance, trade relationships, migration patterns, historical colonial ties. If sketch-based measures captured real cultural differences, they should correlate with these external benchmarks. And they did, dramatically so.
The methodology section of the paper is meticulous, the kind of scholarly rigor that comes from knowing the findings will be scrutinized. The researchers controlled for drawing quality (more skilled drawers might introduce noise), for GDP per capita (wealthier countries might have more experienced digital users), for internet penetration rates, for the number of participants per country. They ran their analyses on subsets of the data to ensure robustness. They compared their results against multiple alternative explanations. The signal held.
What They Found
The first major finding is almost poetic in its simplicity: when you ask people to draw a concept, the drawings are not uniform. They are not even close to uniform.
Take the concept "car." You might imagine that a car is a car is a car—that people around the world share a stable visual prototype. But the data tells a different story. Single concepts unfolded into what the researchers call "multiple distinct visual exemplars," clusters of drawings that share visual features but differ meaningfully from one another. Some clusters represent side views; others show front views. Some depict sedans; others capture trucks or compact city cars. And critically, these clusters are not distributed randomly across cultures. They are systematically structured by geography, by cultural history, by the material environments in which people live.
This visual diversity was not uniform across all concepts. The researchers categorized concepts along several dimensions—abstract versus concrete, tangible versus intangible—and found that certain types of concepts showed dramatically more visual variation than others.
Visual Variation by Concept Type
Cross-cultural variation in visual representation by concept type, measured as the normalized variance in sketch embeddings across countries.
| Label | Value |
|---|---|
| High (tools, objects) | 0.85 |
| Medium (animals) | 0.62 |
| Low (abstract) | 0.31 |
The strongest finding concerned haptic interaction: concepts that involve touch, manipulation, physical handling showed the most pronounced cross-cultural variation in visual representation. A hammer, a toothbrush, a pair of scissors—these objects sit in your hand. They have weight and texture and a specific grip. And across cultures, people interact with them differently. They hold them differently, use them in different contexts, encounter them in different material environments.
This finding connects to a deep tradition in cognitive science called embodied cognition—the idea that concepts are not abstract symbols floating in some Platonic mental space but are grounded in the physical, sensory experiences of having a body. When you think of a hammer, the researchers argue, you recruit motor programs associated with hammering. When you imagine a toothbrush, you simulate the proprioceptive experience of brushing. These embodied traces vary across cultures because cultures shape bodies differently—different tools, different gestures, different physical worlds to navigate.
"Visual imagery reflects variation in embodied experience as much as conventional definitions," the researchers write. This is a quiet revolutionary claim. It suggests that the body—not just the mind—is a carrier of cultural difference, and that visual representation accesses these bodily differences in a way that language cannot.
Alignment with Cultural Distance Measures
Correlation with established cultural distance measures derived from cross-cultural psychology, economic geography, and sociological research.
| Label | Value |
|---|---|
| Sketch-based measures | 0.82 |
| Text-based measures | 0.56 |
The divergence between sketch-based and text-based conceptual representations is the study's second headline finding. When the researchers compared the geometry of sketch embeddings—how concepts relate to each other in visual space—with the geometry of word embeddings across 58 languages, they found systematic divergence. The same concept occupied different positions relative to other concepts depending on whether it was measured through drawings or through words.
This is not a minor technical disagreement. It is a fundamental difference in the structure of knowledge.
Consider the concept "snow." In English, "snow" is a single word. It relates to winter, cold, weather. But for the Inuit, multiple distinct words for different types of snow have been documented—granular snow, packed snow, slushy snow. The language captures experiential distinctions that a single English word misses. Now imagine this principle extended to hundreds of concepts across dozens of cultures. The sketch data suggests that visual representation captures rich semantic and cultural structure that language models systematically compress.
When language encodes a concept, it forces convergence. The word "house" must be understood by both a Norwegian and a Kenyan, which means it can only carry the semantic features they share. But the drawing of a house—the specific shape of the roof, the materials visible in the sketch, the presence or absence of a chimney—can carry cultural information that the word intentionally discards.
This brings us to the study's most striking quantitative result: the 45 percent improvement in alignment with established cultural distances.
Global Distribution of Sketch Data
Distribution of sketches across countries by population size, showing both total sketch counts and unique concept types represented.
| Label | Value |
|---|---|
| 0-50M | 12 |
| 50M-200M | 28 |
| 200M-500M | 45 |
| >500M | 15 |
The researchers measured how well three different approaches predicted the actual cultural distances between countries—distances derived from decades of cross-cultural psychology research, from economic geography, from sociological studies. A purely linguistic approach, based on word embeddings across languages, performed respectably but missed much. A sketch-based approach, using the visual exemplars people produced, performed dramatically better.
"Cross-cultural similarities derived from sketches align 45% more closely with established cultural distances than do text-based measures," the researchers report. "This suggests that patterns of human conceptual universality may depend critically on the modality through which concepts are measured."
Forty-five percent is not a rounding error. It is a substantial gap that reveals something fundamental about the limitations of language-based approaches to understanding human conceptual diversity.
The finding held across geographic regions, across development levels, across the spectrum of cultural variation. Sketch-derived similarity matrices looked more like the true cultural map of the world than any text-derived alternative. This is not because sketches are somehow more "correct"—it is because they access a different dimension of cultural information, one that linguistic analysis has been blind to.
Why This Changes Things
To appreciate the significance of this study, you need to understand how profoundly the assumption of conceptual universality has shaped modern psychology, linguistics, and AI development.
For decades, the dominant view in cognitive science has been that human concepts are fundamentally universal—that the way we categorize the world, the way we group objects and events into meaningful categories, reflects innate cognitive architecture shared across all humans. This view has been bolstered by studies using word associations, translation equivalents, and linguistic typology. When you ask people in different cultures to name objects, to describe scenes, to translate words into their own language, the patterns look remarkably similar. The evidence seemed clear: concepts are universal.
But this evidence, the new study suggests, was always compromised by its methodology. Words are social objects. They require agreement to function. When you translate a word from one language to another, you are not mapping mental representations—you are mapping communicative conventions. And communicative conventions optimize for agreement, for shared reference, for the elimination of ambiguity. They compress precisely the variation that makes thought interesting.
Visual representation, by contrast, is not a communicative tool. It is a cognitive one. When someone sketches an elephant, they are not trying to be understood by an audience. They are accessing their own mental imagery, their own embodied simulation of what an elephant looks like. And this imagery carries the fingerprints of their cultural environment—the elephants they've seen in zoos, in books, in cartoons, in documentaries, in the specific material and cultural contexts that have shaped their cognitive development.
The implications for artificial intelligence are profound.
Large language models—the systems behind chatbots and AI assistants—are trained on text. Billions of words scraped from the internet, digitized, processed. These models have achieved remarkable fluency, but critics have long worried that they are learning patterns of word co-occurrence without grounded understanding, producing text that mimics human language without accessing the embodied, perceptual, cultural dimensions of concepts.
This study provides the most direct evidence yet that such concerns are well-founded. When language models learn concepts from text, they learn compressed, decontextualized abstractions. They learn that "house" relates to "home" and "building" and "residence"—but they lose the specific visual signatures that distinguish a Norwegian chalet from a Nigerian compound from a Japanese machiya. They miss the haptic knowledge of how a house feels to build, to live in, to inherit.
The researchers are careful not to overclaim. They are not saying that text-based approaches are useless, or that visual data should replace linguistic data. They are saying that different modalities capture different dimensions of conceptual structure, and that a complete picture of human thought requires both.
But the implication for AI is clear: if we want systems that truly understand human concepts—that can navigate the rich texture of cross-cultural variation—we may need to move beyond text. We may need to incorporate visual, embodied, tactile data. We may need AI that can sketch as well as speak.
For psychology and neuroscience, the study challenges a methodological consensus. For decades, cross-cultural psychology has used translated questionnaires, linguistic matching tasks, and word-association studies as its primary tools. These studies have built an impressive edifice of knowledge about human universals and differences. But that edifice may rest on a methodological foundation that systematically overlooks the dimensions of thought that matter most.
The haptic finding is particularly important. The researchers found that concepts involving physical interaction—objects you hold, manipulate, touch—show the most pronounced cross-cultural variation. This suggests that embodied experience is a major source of conceptual diversity, and that embodied measures are necessary to access it. Asking someone to name a hammer tells you about their linguistic knowledge. Asking them to draw one tells you about their sensorimotor experience of hammers, their tactile history, their physical encounters with the object.
This aligns with a growing body of research in cognitive neuroscience showing that concepts are grounded in sensorimotor systems—that when you think about a hammer, you recruit the same brain regions you would use to actually hammer. If this grounding is culturally variable, then the concepts themselves must be culturally variable in ways that purely linguistic studies cannot detect.
The findings also have implications for how we think about globalization and cultural homogenization.
Popular discourse often assumes that globalization leads to cultural convergence—that as people around the world watch the same movies, use the same smartphones, wear the same brands, their concepts become more similar. The linguistic data might seem to support this—everyone knows the word "smartphone," everyone has the same basic categories. But the sketch data suggests that visual imagination may be more resistant to homogenization than linguistic convention. Even as people share more words, they may still picture the world differently. The icons on a smartphone interface may be standardized, but the mental images of a phone, a house, a car, a family—these retain cultural specificity.
This is not necessarily good or bad. It is simply a reminder that cultural convergence is not the same as cognitive convergence, and that the aspects of culture most resistant to global flattening may be the least accessible to linguistic analysis.
What's Next
The study opens as many questions as it answers—perhaps more.
First, what exactly drives the visual variation? The researchers show that haptic interaction concepts vary most, but they do not fully explain why. Is it differences in physical tools across cultures? Differences in embodied experience? Differences in how concepts are taught and learned? The data points toward embodied experience as a key factor, but the precise causal mechanisms remain unclear.
Second, how stable is visual representation over time? The Quick, Draw! data was collected in a specific window—roughly 2016 to the present. It would be valuable to know whether visual concepts change as cultures change, whether they are more or less stable than linguistic concepts, whether they show generational differences within cultures. This is a dataset that could be revisited in a decade to ask how globalization has actually reshaped visual imagination.
Third, what happens with concepts that are more abstract—democracy, justice, happiness, love? The study focuses primarily on concrete objects (cars, elephants, houses), but its theoretical implications extend to abstract concepts. If even concrete concepts show this much visual variation, abstract concepts might show even more. The researchers have begun to explore this, but the analysis is preliminary.
There are also methodological questions worth pursuing. The sketch-based approach captures what people visualize, but visual imagination is only one non-linguistic modality. What would we find if we asked people to gesture concepts, to physically act them out, to describe the sounds or smells associated with them? The embodied cognition tradition suggests that concept representation is multimodal—that words are just one channel through which conceptual structure manifests. This study provides compelling evidence that we have been missing much by focusing on that single channel.
For AI researchers, the practical implication is clear: multimodal models—systems that learn from images, audio, sensorimotor data, as well as text—may be necessary to capture the full richness of human conceptual structure. Recent advances in vision-language models suggest this direction is technically feasible. The harder question is whether it is conceptually coherent—whether the different modalities can be meaningfully integrated or whether they will reveal irreducible incommensurability.
The researchers acknowledge several limitations. The Quick, Draw! data is not a representative sample of any population—it skews toward younger, more digitally literate users in countries with high internet penetration. This means the cross-cultural patterns they find might be confounded by demographic differences between countries. They do their best to control for this, but the limitation is real. Future studies might use more stratified sampling or target specific populations for detailed comparison.
They also acknowledge that drawing is not a perfectly transparent window into mental imagery. Drawing requires motor execution, which introduces noise. Some people are better artists than others. Some cultures may have different drawing conventions that affect how concepts are visually represented. The paper includes analyses meant to address these concerns, but they cannot be fully eliminated.
Most fundamentally, the study raises a question about the nature of concepts themselves.
If the same word maps to different visual representations across cultures—if "house" means something different in visual imagination in Norway versus Kenya—then what does it mean for the concept to be "the same"? The researchers are careful to note that they are not claiming concepts are incommensurable across cultures. People can communicate, can translate, can understand each other despite these differences. But they are claiming that linguistic translation may obscure more than it reveals, and that true cross-cultural understanding requires grappling with the experiential, embodied, visual dimensions of thought that language inevitably compresses.
This is not a new idea—anthropologists and cognitive scientists have argued for decades that concepts are culturally situated, that abstract universalism ignores embodied particularity. But this study provides the most direct, large-scale, quantitative evidence for that claim. It transforms a philosophical intuition into an empirical result.
The 45 percent improvement in cultural distance prediction is the number to carry forward. It tells us that we have been systematically underestimating the extent of cross-cultural variation in human thought—not because the variation is new, but because our tools have been blind to it. For decades, we have been mapping human conceptual structure with instruments that discard exactly the variation we most need to understand.
The sketches reveal what words hide. And what they reveal is this: that even in a world of global communication, shared vocabulary, and increasing cultural exchange, the visual imagination remains profoundly local, shaped by the specific bodies, hands, tools, and physical environments through which each culture encounters the world. The universality of language may be an illusion—a useful, necessary illusion, but an illusion nonetheless. The diversity of visual thought is real, and it is waiting to be understood.
What we picture when we picture the world is not a universal prototype. It is a cultural artifact. And that may be the most important thing this study has to teach us.