Structural priming is a standard technique in psycholinguistics. If you hear a passive sentence (“The ball was kicked by the girl”), you are more likely to produce a passive sentence next, even about an unrelated topic. The priming effect demonstrates that the abstract structure — passive voice, independent of the specific words — has psychological reality. People don't just process word sequences; they process structural templates, and those templates persist between sentences.
For decades, this finding was cited as evidence for hierarchical grammar. The structures that prime are constituents — grammatical units like noun phrases (“the dog”) and verb phrases (“chased the cat”) that combine according to hierarchical rules. If constituents prime, the reasoning went, then the mental representation of language must be organized around constituents. The method confirmed the theory.
Nielsen and Christiansen, publishing in Nature Human Behaviour in 2026, tested what happens when you prime non-constituents — word sequences that cross grammatical boundaries and have no status in hierarchical grammar. Sequences like “added to a” (verb + preposition + determiner) span two different constituent boundaries. According to hierarchical grammar, these strings shouldn't function as units. They shouldn't prime.
They prime. Across four preregistered experiments with 497 participants, non-constituent sequences showed reliable priming effects. People who processed “added to a” were faster to process “defined by the” — a string sharing the same part-of-speech pattern (verb + preposition + determiner) but no words, no meaning, and no grammatical constituents in common. Two additional corpus analyses — eye-tracked reading and naturalistic conversation — confirmed the effect outside the laboratory.
The general pattern: a method that serves as evidence for a theory when tested on the theory's own examples can serve as evidence against it when tested on examples the theory excludes. Structural priming confirmed hierarchical grammar because it was tested on constituents — the units the theory posits. The method was not designed to detect non-constituent structure, so it was never asked to. When finally asked, it detected it immediately. The method was never partial to hierarchical grammar. It was agnostic — capable of detecting whatever structure exists. The partiality lived in the input selection: which sequences the experimenters chose to test. For decades, they chose constituents, because constituents were what the theory said mattered. The method dutifully confirmed. The confirmation was real, but it was incomplete, and its incompleteness was invisible because the theory determined which questions to ask.