Linguistic Relativism

From the AI Research Compass wiki

Linguistic Relativism: A model can only think what its language and training corpus let it say, so the choice of language shapes the intelligence.

Beliefs

Linguistic Relativism holds that a model’s world is exactly as large as its language. Whatever never made it into the corpus, the vocabulary or the grammar of the training data is, for the model, not merely unknown but unthinkable. The doctrine takes this seriously in both directions: a model trained mostly on one language inherits that language’s categories, idioms and blind spots, and a model can gain new concepts simply by being given new words for them.

It sits near the Language edge and a little above the middle of the Monolith–Harness axis: it does not care whether the model is one giant network or a system, only that its medium of thought is linguistic. Its Mainstream shade marks an established view, the respectable middle-aged cousin of the more absolutist language camps above it. Where Logos Absolutism says language is everything, Linguistic Relativism says something subtler and more anxious: language is everything you get, so choose carefully.

Arguments

  • Cross-lingual evaluations routinely show that capability, calibration and cultural knowledge vary with the language of the prompt, even inside one multilingual model.
  • Introducing a named concept (a new term, a defined variable, a labelled category) measurably changes what a model can reason about in-context, which suggests vocabulary is doing real cognitive work.
  • Low-resource languages show that missing text is not a cosmetic gap: whole domains of knowledge simply never reach the model.

Critiques

  • Grounding camps argue that language is a lossy summary of experience, and that a model’s limits are the limits of its data, not of language as such.
  • Multimodal models acquire visual and spatial skills that no caption ever spelled out, which is awkward for a strict reading of the doctrine.
  • The strong Sapir–Whorf version is contested even for humans; importing it wholesale into ML risks dressing up a data-coverage problem as philosophy.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −0.84, y ≈ −0.20, paradigm ≈ −0.61.

A first draft. Corrections and better arguments are welcome through a pull request. Edit this entry on GitHub Edit its data