From the AI Research Compass wiki

Innatism: Holds that language depends on innate, specialized structure, so statistical learning from data alone cannot explain or produce real linguistic competence.

Beliefs

Innatism is the oldest school in the linguistics department and it has not changed its mind. Children acquire rich grammatical knowledge quickly, from limited and noisy input, and converge on rules they were never explicitly shown. That poverty of the stimulus, the argument goes, implies an innate language faculty with built-in constraints on possible grammars. Large language models, trained on more text than any human could read in many lifetimes, do not refute this; if anything, their enormous data appetite is evidence that they learn differently from children, and that fluent output is not the same as competence.

It sits in the centre of the Language–World axis and in the lower half, toward Harness: its concern is theory rather than system design, but it expects minds to be built from innate, structured components rather than one blank-slate network. Its paradigm level is beyond revolutionary, the most extreme on the chart: it predates neural networks and regards much of the current enterprise as engineering that says little about the human mind.

Arguments

  • Children reach grammatical competence with orders of magnitude less linguistic input than large language models receive, which suggests strong inductive biases.
  • Human languages share structural properties, such as structure-dependent rules, that appear to constrain what children consider possible.
  • Language models can learn patterns no human language uses as readily as attested ones, which calls into question whether they model human language specifically.

Critiques

  • Distributional and scale-first camps argue that language models acquiring sophisticated syntax from raw text weakens the poverty-of-the-stimulus argument.
  • The proposed innate structures have been hard to specify precisely, and the theory has shifted over the decades.
  • Engineers note that, whatever its merits as cognitive science, the doctrine has produced few working language systems.

Representative works

Relationships

How to get it

Land on this cell in the full or pro test: x ≈ −0.26, y ≈ −0.89, paradigm ≈ +0.94.

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