Constructing Meaning from Language: Visual Knowledge in People Born Blind and in Large Language Models

A key function of language is to enable concept construction, but empirically disentangling the contribution of language from the contribution of other (e.g., sensory) experiences is challenging. Comparing visual knowledge across sighted people, people born blind, and artificial intelligence (AI) trained exclusively on text provides rare insight. Blind people acquire rich visual knowledge, including normative meanings of light emission and visual perception verbs; similarity of colors; and size, shape, and texture of distal objects. Evidence from text-trained AI models suggests that such visual knowledge can be derived from language. Going beyond meanings of single words, blind people also construct causal intuitive theories of color, light, and visual perception, enabling generative inferences about visual phenomena (e.g., inferring the likelihood that a sighted character will see an object at a distance and the likely number of colors for a given artifact). Language enables concept construction from the ground up, without sensory evidence, and is intimately linked to intuitive theories.