Billions of Doodles Show Concepts Aren't as Universal as Language Suggests

Researchers challenged the long-standing assumption that human concepts are broadly universal, an idea usually tested by comparing words across languages. Their argument: language is a compression scheme. Shared vocabulary smooths over the messy, personal experience behind a concept, so measuring universality through words may hide real differences in how people actually picture things in their heads. To get underneath language, the team analyzed 2.6 billion sketches of everyday concepts drawn by people in 236 countries and territories, treating drawings as a direct window into visual imagination.

The sketches told a different story than words do. A single concept typically fractured into several distinct visual forms rather than one shared template, and the divergence was largest for concepts tied to hands-on, tactile experience — suggesting that mental imagery tracks embodied, physical interaction, not just dictionary definitions. When the researchers built embedding models from the sketches and compared their geometry to word-embedding models, the two diverged: the visual representations retained semantic and cultural nuance that the language models flattened out.

The practical upshot is a measurement problem. Cultural similarities inferred from sketches matched established cross-cultural distance measures 45% better than text-based estimates did, implying that conclusions about what humans share depend heavily on the modality used to probe them. For anyone building or evaluating AI systems on text embeddings alone, the finding is a caution: language-only representations may systematically erase cultural and experiential variation that other modalities preserve.