Accessing the unsaidThe role of scalar alternatives in children’s pragmatic inference

Picture a four-year-old in a lab. She's looking at a picture of Cookie Monster holding an apple, a banana, and an orange — all three fruits in his hands. The experimenter asks, "Is Cookie Monster holding some of the fruits?" She says yes. The researcher marks it wrong. But here's the thing — she's not confused, she's not guessing, and she's not being careless. She's doing something systematic. Figuring out exactly what she's doing, and why, reveals something deep about how children learn language. The puzzle is called scalar implicature. When adults hear the sentence "Cookie Monster is holding some of the fruits" in a scene where he's holding all of them, they reject it. Not because it's literally false — "some" is logically compatible with "all," since holding all three fruits technically means holding some of them. Adults reject it because they run an inference: a cooperative speaker who could have said "all" would have said "all." The fact that they said "some" signals that "all" isn't true. That's a scalar implicature — the inference that a weaker word, like "some," excludes the stronger alternative, "all." Children systematically fail to run that inference. Papafragou and Musolino showed this in 2003, and the finding has been replicated many times since. But why? Two camps formed. One said children know about stronger alternatives like "all" but lack the processing resources — the working memory, the inferential bandwidth — to compute the implicature reliably during conversation. The other camp proposed something more specific and interesting: maybe children fail because they don't spontaneously access the relevant alternatives in the first place. If you don't know that "all" belongs on the same scale as "some," the inferential step never gets started. David Barner, Neon Brooks, and Alan Bale set out to pit those two explanations against each other. Their key theoretical move is to draw a distinction between two types of alternatives. Context-independent alternatives are pre-packaged lexical scales. The scale of "some," "many," "most," and "all" is an example. These are stored in the language system, and accessing them requires knowing which words form a contrast set with which other words. That's learned knowledge. Contextual alternatives, by contrast, are supplied by the scene itself. If the picture contains a drum, a ball, and a trumpet, the objects themselves make the alternatives available — no stored scale is required. Barner and colleagues reasoned that if children's failures are really about processing limits, then both types of alternatives should be equally hard, because the inferential computation is the same either way. But if the failures are about scale-specific knowledge, children should succeed when alternatives are contextually provided and fail when they have to retrieve a lexical scale. The experiment they designed is elegant. Sixty four-year-olds, with a mean age of just under 54 months, looked at picture cards showing three objects and heard questions about them. Half the conditions used context-independent quantifier language, asking "Is Cookie Monster holding some of the fruits?" The other half used contextual alternatives, naming the specific objects in the scene: "Is Cookie Monster holding the banana, the apple, and the orange?" The critical test was scenes where all three objects fit the description but the question only asked about a subset. Would children strengthen the meaning and say no, or would they accept the partial description as good enough? There was one more manipulation. In half of the conditions, the experimenters inserted the word "only" into the test sentences. "Is Cookie Monster holding only some of the fruits?" The word "only" is a focus particle — it grammatically forces what linguists call exhaustification, meaning it makes the exclusive reading part of the sentence's literal content rather than something you have to infer. Adding "only" was a way of giving children a boost: if their failures are really about not having enough cognitive resources to compute the inference, then making the exclusive reading grammatically explicit should help them. If their failures are about not knowing the lexical alternatives, then "only" should help when contextual alternatives are available but do nothing for the some/all scale — because "only" can't supply the alternative "all" if the child doesn't already have it in mind. The results were striking, and they fell exactly along the predicted line. When alternatives were provided by the context, when the experimenter named the specific objects, children behaved like adults. In scenes where all three objects fit the description, children accepted the unexhaustified subset description ninety-two point nine percent of the time, which makes sense: "the apple and the banana" is literally true even when all three fruits are present. But when "only" was added, acceptance dropped to just fourteen percent. Children used "only" to derive the strengthened meaning and correctly rejected the partial description. They could exhaustify. The machinery was there. However, on the context-independent some/all conditions, the story was completely different. Children accepted two-item-true trials eighty percent of the time and correctly rejected false trials eighty-seven point two percent of the time, so their basic comprehension was fine. The problem appeared on the critical test: when all three items fit the description but the question asked about a subset using "some," children did not strengthen the meaning. They accepted the partial description. And adding "only" did nothing — there was no significant difference between conditions with and without "only" on those trials. The three-way interaction between trial type, alternative type, and grammatical exhaustification was highly significant, with an F-statistic of thirteen point twenty-eight and a p-value well below zero point zero zero one. The pattern was clean: "only" rescued performance when the scene supplied alternatives, but completely failed to rescue it when alternatives had to come from the lexical scale. Think about what that means. These are the same children, the same sentences, the same word "only," the same cognitive demands — the only thing that changed was whether the relevant alternatives were visible in the picture or stored in the mental lexicon. That single difference determined whether children computed the implicature at all. That is a very precise kind of failure. It is not a general pragmatic immaturity. It is a gap in scale-specific lexical knowledge. Barner and colleagues draw a pointed comparison to numerals. Children tend to show adult-like exact interpretations for number words — "two" means exactly two, not two or more. Their explanation is that number words are learned as an ordered count list, which automatically makes alternatives available. A child who knows the count sequence knows that "three" comes after "two," so when someone says "two," "three" is already implicitly in the contrast set. Quantifiers like "some" and "all" don't come with that built-in structure. You have to learn, through exposure, that these particular words form a scale together — that "all" is the relevant stronger alternative to "some." And apparently, most four-year-olds haven't learned that yet. What this means for how we think about language acquisition is worth sitting with. We tend to assume that learning a word is about learning what it means — its referents and its conditions of application. But this work suggests that learning a word fully also means learning what it excludes. "Some" is not complete as a lexical entry until you've stored the information that "all" is its scale mate, the stronger alternative that speakers choose when they can. That's a richer and more demanding kind of learning than we usually picture when we imagine a toddler pointing at objects and hearing their names. The hopeful reading — and it is the right reading — is that children are not pragmatically broken. When you give them the tools, they use them. The same four-year-olds who failed to strengthen "some" could strengthen a contextual description the moment "only" appeared. The capacity for scalar inference is intact. What's missing is a specific piece of lexical knowledge, one that will presumably slot into place as children encounter more contrastive uses of these words in ordinary conversation. The question Barner, Brooks, and Bale leave open is how that learning happens — what in the input triggers the realization that "some" and "all" belong together on a scale. That's the next experiment waiting to be run. This lecture was created by ennepō. Go to https://ennepo.ai to Discover, Create and Follow the latest research in your field. Read when you can. Listen when you want to.