Is Grammar Built In?

The poverty of the stimulus argument runs like this. Children arrive at a grammar that goes far beyond what they have heard. Several grammars would be compatible with the input any particular child receives, and yet every child converges on the same one. Something must be narrowing the options, and since it is not in the input it must be in the child.

Chomsky's name for what is supplied is universal grammar: a set of structural constraints shared by all human languages and provided by the brain rather than by experience. The idea reorganised linguistics in the 1960s and is still the most influential proposal in the field, and it has had an unusual career.

What belongs inside it has shrunk enormously. Early versions proposed an elaborate apparatus of principles with parameters set by exposure: a switchboard the child flips according to which language is being heard. The later minimalist programme proposes almost nothing beyond recursion and the interfaces to sound and meaning. A theory whose core keeps contracting looks different from one being progressively confirmed. Its defenders describe the contraction as refinement, and distinguishing a research programme that is converging from one that is retreating is famously difficult. Lakatos wrote a whole account of science around the problem and applying it here has not produced agreement either.

The main alternative is usage-based, and Tomasello's version is the most developed. It rests on two claims. The input is much richer than the poverty argument allows, especially once child-directed speech and the surrounding social context are counted properly. And humans have general capacities (statistical pattern-finding, analogy, and above all an unusual ability to work out what another person intends) which were not available as explanations when the argument was first made. It does not deny that humans are special; it relocates the specialness away from a dedicated grammar module.

Both positions have vastly better evidence than in 1960. Corpus studies have quantified what children actually hear. Cross-linguistic work has found far more structural variation than early universalist claims predicted. Statistical learning experiments have shown infants extracting surprisingly complex regularities from brief exposure. And the two camps have moved towards each other without either conceding: the innate component proposed has shrunk, the learning mechanisms proposed have grown more powerful.

Which brings the language models in, and it is worth being careful about what they show. The existence proof matters: the strongest version of the poverty argument held that no general learner could reach adult syntax from the available data, and that is now hard to maintain. The quantity matters equally, in the other direction: a child hears perhaps tens of millions of words and achieves full competence, while these systems require several orders of magnitude more and still make errors no child makes.

So the honest summary is that the evidence has improved in every direction and nobody has been forced to concede. That is unusual and it is not a failure. Some questions turn out to be genuinely difficult, and this is one of the few places in the cognitive sciences where a sixty-year-old argument is still worth having rather than merely still being had.