The AI Journal

Ask an English speaker to guess which of two invented words, “bouba” or “kiki,” names a round blob and which names a jagged star, and the answer arrives before the sentence finishes: bouba is the blob, kiki is the star, and the agreement runs at 95 to 98 per cent whether the respondent speaks English or Tamil. Nothing about the referents makes this obvious. A Japanese speaker’s ideophones, an English toddler’s first attempts at “wiggle,” and that same Tamil-English agreement all draw on a rounded-versus-angular mapping between the shape of a mouth’s movement and the shape of an object, and the coincidence is too wide and too old to file as a curiosity. It has instead become an embarrassment for the founding claim of modern linguistics: that the bond between a word’s sound and its meaning is arbitrary, motivated by nothing beyond convention.

Ferdinand de Saussure did not present arbitrariness as a hypothesis to be tested; he called it the first principle of the sign, the axiom from which the rest of structural linguistics proceeds. Everything that made the Course in General Linguistics a founding text depended on that axiom holding without exception: if some words meant what they meant because of how they sounded, then meaning could not be purely a matter of each sign’s position within a self-contained system of differences, which was the larger claim structuralism needed. Arbitrariness was not one observation among many in that architecture; it was load-bearing. In the century since the Course was assembled from Saussure’s students’ notes, the principle has behaved less like an empirical generalisation than like a founding constitution: exceptions such as onomatopoeia were acknowledged and then set aside as peripheral, a small tax paid to keep the theory clean. The trouble is that the exceptions kept accumulating, and once linguists started counting them rather than noting them in passing, the count would not stay small.

Damián Blasi and colleagues tested this directly at a scale no earlier study could match: basic vocabulary items across roughly 6,000 languages, more than 60 per cent of the world’s living tongues. If arbitrariness held as advertised, the phonemes appearing in words for, say, “nose” or “round” should be statistically independent of the concepts they name, once genealogical relatedness between languages is controlled for. They are not. Words for “nose” disproportionately contain nasal consonants across unrelated language families; words for “round” disproportionately contain a rounded vowel or a rounded articulation; “small” recruits high front vowels with striking regularity. None of these associations is close to universal, and none approaches the reliability of an actual grammatical rule. But the study’s design controlled for exactly the objection that would otherwise dissolve the finding: languages descended from a common ancestor share vocabulary for reasons that have nothing to do with iconicity, so the associations had to hold up across families with no demonstrable historical connection to count. They did. The associations are too consistent, and recur across too many genealogically independent families, to be sampling noise, and their sheer breadth is not something arbitrariness, taken as a strict first principle, predicts or can accommodate without amendment.

The developmental evidence sharpens the point past a lexical curiosity into a claim about how vocabulary is learnt at all. Mutsumi Imai and Sotaro Kita’s sound symbolism bootstrapping hypothesis holds that pre-verbal infants already map certain acoustic properties onto certain perceptual properties, and that toddlers use this mapping as scaffolding to attach their first, seemingly arbitrary words to their referents. Japanese-learning toddlers taught a novel sound-symbolic verb for an unfamiliar action generalise the meaning to new instances of similar-looking motion far more reliably than toddlers taught an arbitrary-sounding control verb, and the effect shows up before the age at which children reliably produce arbitrary vocabulary of comparable size, which is what marks it as scaffolding rather than an incidental correlate of learning. If this is right, non-arbitrary mapping is not a decoration sitting on top of an otherwise arbitrary vocabulary; it is functional infrastructure that the earliest stage of word learning leans on before arbitrariness can get started, because a language whose first words had to be memorised one by one, with no perceptual handhold connecting sound to referent at all, would be a far harder thing for an infant to bootstrap into. The mapping does not need to survive into adult competence to have done its work; it only needs to have lowered the cost of the first few hundred words enough that the arbitrary system built on top of them ever gets off the ground.

Mark Dingemanse and colleagues supplied the vocabulary this argument has been building toward: alongside arbitrariness, treat iconicity — where the form of a word resembles some aspect of its meaning, as in bouba’s rounded vowel — and systematicity — where statistical regularities in sound predict grammatical category, independent of any resemblance — as co-equal organising principles of the lexicon, each doing a distinct job. Systematicity aids category learning; iconicity aids the early perceptual anchoring Imai and Kita describe; arbitrariness, on this reframing, earns its own keep by letting words become individually distinctive once a concept no longer needs fresh bootstrapping. That division of labour fits the data better than a model in which arbitrariness is the rule and iconicity the noise around it, because it explains why iconic effects cluster precisely where bootstrapping and individuation pressures predict they should, rather than being smeared at random across the whole vocabulary.

The strongest objection to all this is not that the individual studies are wrong; it is that “iconicity matters” has been oversold by exactly the kind of demonstration this piece opened with. Bouba and kiki are invented words tested against two cartoonishly extreme shapes, and the vast majority of any language’s actual vocabulary shows no detectable sound-meaning correlation at all: nothing in the phonemes of “chair” or “justice” or “Tuesday” iconically resembles anything. Bodo Winter and Marcus Perlman put this to a direct test on English, running a random-forest analysis of sound structure against rated size across more than 2,500 general-vocabulary words, after first confirming a strong size-sound relationship among size adjectives specifically. The general-vocabulary result was null: no detectable size sound symbolism once the word class was no longer restricted to adjectives built to encode size in the first place. That is a real limit, not a methodological quibble, and it means the claim on offer cannot be that language is secretly iconic throughout. It has to be narrower: iconicity is a systematic, recurring, functionally motivated minority signal, concentrated in ideophones, sound-symbolic verb classes and the earliest-acquired vocabulary, rather than a property distributed evenly across the lexicon.

Arbitrariness remains the correct description of most of any language’s word stock, and nothing here licenses the fashionable overcorrection in which every invented experimental pairing gets read as evidence that language was iconic all along. What it cannot remain is the unexamined axiom from which everything else follows, because a founding principle that turns out to hold only statistically, only in identifiable corners of the vocabulary, and only for specific functional reasons is no longer a first principle in Saussure’s sense. It is one finding among several, sitting alongside iconicity and systematicity rather than presiding over them, and it now owes the same evidential defence its exceptions were always asked to provide. A discipline that spent a century treating arbitrariness as beyond argument was, on this evidence, mistaking the tidiness of an axiom for the tidiness of the facts.