The sentence you are reading has probably never been written before. You do not find that remarkable either — you build sentences that never existed every day, and you have never once worried about running out.
Yet the set of sounds you can produce, and that your ear reliably tells apart, is a few dozen and stays that way for life. Finite parts, infinite output: that gap needs an explanation.
The handiest explanation is the wrong one. It is tempting to picture language as a very thick dictionary, and speaking as having memorized enough entries. If that were true you would eventually hit a wall, because a dictionary has a last page. What follows takes apart three things: the bag of meaningless sounds underneath, the loose rope between sound and meaning, and the rule that keeps sentences from having a maximum length.
Start with the smallest possible case. In Mandarin, the syllables mā, má, mǎ and mà involve exactly the same movements in the mouth: lips together, then open. The only difference is where the pitch of the voice goes — flat and level gives 妈 (mā, "mother"), rising gives 麻 (má, "hemp"), dipping down and back up gives 马 (mǎ, "horse"), falling sharply gives 骂 (mà, "to scold"). A parent, a plant, an animal, an action. Meanings with nothing in common, separated physically by one pitch contour.
That is what a part is: on its own it is nothing. The rising contour does not denote anything; you cannot say a thing with it alone. But swap it out and the whole word becomes a different word. Linguists call this two-layer arrangement duality of patterning: underneath sits a small bag of sound units that carry no meaning, and above it sit the meaningful words and sentences. The endlessness of the upper layer is bought by re-shuffling the few dozen items in the lower one.
How small can the bag get? The Central dialect of Rotokas, spoken in Papua New Guinea, is regularly cited as having one of the world's smallest phoneme inventories: eleven, six consonants and five vowels. At the other end, Taa (also written !Xóõ), in southern Africa, has so many consonants that counts range from 88 to 130 depending on how you analyse them, because it uses clicks — the family of sounds that includes the disapproving "tsk" — as ordinary consonants. Eleven and a hundred-odd, and both are enough to say anything. The size of the inventory does not decide what a language can express.
What is more striking is that the two layers are not free. Al-Sayyid Bedouin Sign Language, which arose in the Negev in Israel and is only a few decades old, has been documented over years by Sandler and colleagues: its vocabulary is already large and its expressive range complete, yet handshapes and locations have not yet settled into a set of reusable, meaningless building blocks. A living language, in other words, can start out without the lower layer and grow it later. Strength of evidence: long-term observation of a single community, widely accepted, but a sample of one.
There is nothing dog-like in the sound of "dog". French says chien, Japanese says inu, and none of them resembles the animal more than the others. The rope between sound and meaning is slack; it is settled by convention. This is arbitrariness, and it sounds like a dull observation until you notice it is what makes the previous section work: precisely because a word need not resemble what it denotes, one bag of parts can be sent out to denote anything at all.
The rule does leak, though, and the shape of the leaks is informative.
The first leak is onomatopoeia. A rooster crows, and English writes it cock-a-doodle-doo, French cocorico, German kikeriki, Japanese コケコッコー (kokekokkō), Mandarin 喔喔喔 (wō wō wō). Same bird. The European versions converge on an initial k, which shows imitation is genuinely going on; but the results diverge so widely because imitation can only be assembled out of the parts a language already stocks — the bag from the last section. You cannot imitate with a sound you do not have. So onomatopoeia is not a counterexample to arbitrariness; it is a case of motivation filtered through an inventory.
The second leak is harder. Show people a round blob and a spiky star and ask which is bouba and which is kiki, and they overwhelmingly give bouba to the round one. Ćwiek and colleagues redid this in 2022 across 25 languages, nine families and ten writing systems: bouba matched the round shape 72% of the time and kiki the spiky one 63%, with the effect significant in 17 of the 25 languages. Speakers of non-Latin scripts showed it too, so it is not the letter b looking rounder than the letter k. Strength of evidence: a large cross-linguistic experiment, replicated.
The third leak comes from word lists. Blasi and colleagues in 2016 combed basic vocabulary from close to two-thirds of the world's languages and found that "nose" tends to carry an n, "tongue" an l, "small" an i, and "full" a p or b, with distributions suggesting these arise independently in many places rather than through common ancestry or borrowing. Note that this is a statistical tendency, not a law: the Mandarin word for nose, 鼻 (bí), has no n, and that is not a refutation, because the claim was always about a bias.
Put the three leaks together and arbitrariness survives as the main story — the vast majority of words have no link to what they denote. But the strong version, that the link is purely arbitrary, no longer stands. The current picture has arbitrariness, iconicity and systematicity pulling at the vocabulary at the same time.
Parts and a way of combining them still leave one thing missing. You could have a bag of sounds and a pile of words and still be stuck inside sentences of fixed length. What actually removes the ceiling is a rule that allows self-nesting: a sentence can contain another sentence.
"He didn't say it" is one. "You think he didn't say it" swallows it whole. "I know you think he didn't say it" wraps another layer around that. The move is always legal, you can always add one more, and so the set of sentences has no largest member. The rule is called recursion, and it is not a flourish — it is the final reason language is not a dictionary. A dictionary has a last page; this rule has none.
Legal is not the same as sayable, though. Push the nesting into the middle of a sentence and people collapse fast. One layer is fine. Two makes you pause. Three is basically unreadable. The textbook English case: the rat the cat the dog chased killed ate the malt. Every word is grammatical, the meaning genuinely is in there, and nobody can hear it.
Karlsson, in 2007, dug 119 genuine multiple centre-embeddings out of seven European languages including English, Finnish, French, German and Latin. His finding: in writing the maximum depth is three, in speech such structures are practically absent, and they entered the record along with the written rhetorical tradition of Cicero and his successors. Recursion, in other words, supplies the property of having no ceiling; how deep anyone actually goes is propped up by paper and eyes.
Whether every language has the rule is still argued over. Everett reported in 2005 that Pirahã, in the Amazon, shows no clausal embedding; Nevins, Pesetsky and Rodrigues replied in 2009 that the same material admits other analyses — and that the two sides are not using one definition of recursion, one meaning a category nested inside itself, the other meaning that any two elements can combine. Strength of evidence: a field report on a single language against a re-analysis of the same material, neither of them a repeatable experiment. It remains unsettled.
Fewer parts means fewer short combinations, so intuition says words must stretch. But is "enough" measured in word length or in information delivered per second? If two languages differ twofold in word length and twofold in speaking rate, would you call them equally efficient? The answer depends on which unit you picked first.
When a community first meets an animal or a machine it has never encountered, does it coin the sound word or borrow it? If borrowed, it must be reshaped into something the language can pronounce at the moment of entry. How much of "this word resembles that sound" survives the reshaping?
Bouba/kiki and the word-list statistics both show weak cross-modal leanings between sound and meaning. Whether weak leanings can carry the origin of an entire vocabulary is a claim of a different magnitude. The missing step is probably this: are these biases strong enough that a new word could be guessed correctly with no convention in place?
The sentences anyone actually utters are always finite, so the outputs of the two systems may coincide entirely. Is "unbounded" then a description of language, or a description of how we chose to write language down? If no observation separates them, how would the distinction be defended?