Ask someone to pair two nonsense words with two shapes — one rounded and soft, one jagged and star-shaped — and they will do it in under a second, without asking what the words are supposed to mean. Almost everyone puts kiki on the spikes and bouba on the curve. The strange part is not that people agree. It is that the agreement was never arranged: nobody taught them, the words are empty, the shapes are abstract, and the mapping still comes out roughly the same in twenty-five languages, in communities with no writing system, and in a lab full of American undergraduates. In 1924 a psychologist in Tbilisi noticed the pattern in a stack of index cards and wrote that people must be following some regularity “which the human soul follows in the process of name-giving.” A hundred years later, a model with no mouth and no ears was found to do it too.
🧠 The Table of Forty-Two Words
Dimitri Uznadze ran the first version of this experiment in 1924. Ten participants, six abstract drawings, five seconds on each, and a list of forty-two nonsense words to assign to them. There was no right answer, and the words had no meaning in any language. Random assignment was the only thing the design could guarantee.
It did not come out random. For one of the drawings, 45 percent of the participants picked the same word. For three others, 40 percent did. With forty-two words in play, that clustering is far beyond what chance allows, and Uznadze did not treat it as noise. He noticed that people were not guessing — they were reasoning, and their reasons kept invoking the same qualities: something about the word felt soft, or heavy, or quick, and that feeling decided the match.
Five years later, Wolfgang Köhler put two figures in a book — one all spikes, one all curves — and asked readers which was takete and which was maluma. He never published a tally, only implied the answer, and the implication was strong enough that the pairing has been reprinted in textbooks ever since. In 2001, Vilayanur Ramachandran and Edward Hubbard ran the modern version with new figures and new words, and reported that about 95 percent of people call the curvy shape bouba and the jagged one kiki.
Ninety-five percent is not a preference. It is closer to a reflex.
⚙️ How It Works
Step one: the sound has a body. Say bouba out loud. Your lips press together, round, and release; for a moment your mouth is a tube, and the sound that comes out is low and continuous. Now say kiki. The back of your tongue goes up against the roof of your mouth and stops the air twice, the lips stay spread and tight, and what escapes is high and clipped. Two words, two completely different pieces of physical work.
Step two: hearing picks up that work. When the sound arrives, the auditory system does not only decode syllables. It also registers how a body would have to be shaped to produce them — the rounded lips, the narrow tongue, the interruption of airflow. That information is not part of the word’s meaning. It is a by-product of pronunciation, and it turns out the brain keeps it.
Step three: the body information gets translated into other senses. Roundness, sharpness, softness, size, speed, frequency — the same handful of dimensions the visual system uses to describe an object. This is a cross-modal correspondence, and the important word is correspondence, not association: it happens at the level of perception, before any deliberate comparison. Brain imaging supports the idea that a match is cheap and a mismatch is expensive. In the first fMRI study of the effect, published in 2019, Nathan Peiffer-Smadja and Laurent Cohen found stronger prefrontal activation when bouba was paired with a spiky shape than when it was paired with a round one — the mismatched condition costs extra work. A follow-up by Kelly McCormick and colleagues found the same pattern of mismatch-related activity concentrated in parietal regions involved in sensory association and perceptual-motor processing. Matching also shifted activity in the visual and auditory cortices themselves, which suggests the fusion happens early rather than at the end of a chain of reasoning.
Step four: the shapes carry no information, so the bias is the only cue. This is the part that makes the effect look bigger than it is. An abstract drawing has no name, no function, no context. There is nothing to reason from except the faint pull of the sound. In a forced choice, a bias that would normally be drowned out by meaning becomes the deciding vote.
Step five: weak in one person, unanimous in a crowd. Nobody in these studies says “I am certain that bouba is the round one.” They say “I don’t know, that one, I guess.” The effect lives in the aggregate: a slight tilt in each head that adds up to 95 percent of the room pointing the same way. That is the signature of a perceptual bias rather than a belief — it does not need conviction to operate, and it does not survive introspection.
Step six: an innate scaffold, then language sharpens it. The mapping shows up early and shows up in animals. Four-month-old infants look longer at mismatched pairs, meaning they already treat some sound-shape combinations as wrong. In 2024 a team at the University of Padua reported the same preference in chicks one to three days old, which is hard to explain as cultural learning. But the story is not simply “humans are born with it.” A bonobo named Kanzi, raised with enough language training to understand hundreds of spoken English words, learned picture-word pairings but showed no spontaneous shape-sound mapping at all. And the effect breaks when the word stops looking like a word: a large 2021 cross-language study found average match rates below 50 percent in Mandarin, Turkish, Romanian and Albanian speakers, though the authors call that estimate conservative and would not claim the effect is absent in those languages. If the sound sequence is not a licit word in your language, the shape it suggests seems to come loose.
There are three competing explanations for the mechanism, and they are not exclusive. The mouth-posture account says the mapping runs through articulation: round lips for a round object, a tight tongue for a spiked one. The acoustic account says it runs through the sound itself: low, continuous frequencies for something smooth and heavy; high, abrupt ones for something sharp. The concept account, sometimes called ideasthesia, says the mapping is mediated by abstract concepts that activate perception-like experience. Which one dominates may depend on the pair: bouba versus kiki is almost a designed contrast, and it loads all three at once.
🤔 The Crack in the Foundation
For a century, the load-bearing assumption of linguistics has been arbitrariness. The sound of a word and the thing it names have no necessary relation; cat and mao and chat all work, and which one you use is a convention you inherited. The bouba/kiki effect is the most stubborn crack in that foundation. It is a place where ninety-five people out of a hundred arrive at the same convention without ever having convened.
What makes it stranger is that the effect does not require a body. In 2023, a group probing vision-and-language models published a paper titled “Kiki or Bouba?” and found that CLIP and Stable Diffusion show the same bias under zero-shot testing. These systems have never had lips, a tongue, or an ear. They learned the mapping from text and image pairs — which means the correspondence is present in the statistics of human language, not only in human anatomy. Read it in the other direction and the finding gets more interesting still: the reason a model can learn this may be that we have been writing it down for centuries without noticing.
The effect also escapes the shape task. Show people the names Molly and Kate beside a round silhouette and a sharp one, and they will pair them the way the effect predicts — and then attribute personality to match, calling the round-named one easygoing and the sharp-named one determined. A bias about vowels becomes a judgment about a person, and nobody notices the leap.
🔗 Why It Matters Outside the Lab
Naming is a design decision, and this is the part of it that happens before anyone reads a word of copy.
The marketing literature has been measuring it for decades. In a 2000 paper in Marketing Letters, Richard Klink showed that brand names built on front vowels — the i in slim, the e in pet, where the tongue sits forward and the lips stay flat — read as smaller, lighter, faster, colder, thinner, prettier and more feminine. Names built on back vowels — the o in bold, the u in boot, tongue back, lips round — read as larger, heavier, slower and more powerful. A follow-up in 2001 found that when sound symbolism agrees with the semantic meaning of the name, products are liked more and the positioning holds harder in memory. The vowel is doing work that the tagline usually gets credit for.
Anyone naming a product, a persona, or a character is making that judgment whether or not they intend to. A name with an open vowel and a long tail arrives with a temperament attached; a name that ends on a hard stop arrives with a different one. Readers meet the sound before they meet the person, and the sound has already voted. That is worth knowing when a character’s name and a character’s nature disagree, because the disagreement will be felt as friction long before anyone can say why.
And it goes one layer deeper for anything with a voice. If a system without a body can reproduce the mapping, then the choice of what it is called, how its name is spelled, and what sounds its voice leans on are all pre-loaded with personality — before a single line of dialogue is written. The name is not a label attached to the character. It is the first thing the character does.
🎲 Bonus
The next time you have a willing friend and thirty seconds, run the experiment. Draw a blob and a starburst on a napkin, ask which one is bouba and which is kiki, and watch the answer arrive before the explanation does. Then tell them what you just did: they have joined a majority that spans twenty-five languages, four-month-old infants, and chicks less than a week out of the egg. And that a bonobo who understands hundreds of spoken English words is the one who failed the test. Being able to name things, it seems, is not the same as hearing what a name is shaped like — and we are not sure yet which of the two came first.