Hearing shapes: the Bouba–Kiki effect in adult domestic dogs

The preferential-looking paradigm is standard for animal cognition, and the counterbalancing of sound order and shape position is well done. The second study, with owner blinding, is a thoughtful control.

The problems are in measurement and confounds. The AI gaze-classification model was trained on 280 images drawn from the experiment's own video sessions, with no held-out validation from independent recordings — raising the possibility that the model learned visual shortcuts correlated with trial type rather than head direction. The 30-degree threshold for classifying whether a dog was "looking at" a screen was chosen based on undocumented observations, with no sensitivity analysis. Head orientation is an imperfect proxy for gaze in dogs, which have a wide visual field and can shift their eyes without moving their head.

"Bouba" and "Kiki" differ on many acoustic dimensions simultaneously — fundamental frequency, spectral sharpness, consonant voicing, vowel quality. Dogs already associate high-pitched sounds with small stimuli (Korzeniowska 2019, Korzeniowska 2022). The observed result could reflect this known pitch-size mapping rather than a distinct sound-shape correspondence, and the design does not attempt to isolate which acoustic feature drives the effect.

Study 2 used different dogs from a different recruitment pool (training centres versus university database), making it a between-subjects comparison where any difference between studies could reflect the dogs rather than the blinding. The looking-preference finding replicated, but every other result — the general preference for spiky shapes, the shorter first-look latency toward spiky shapes, the exploratory-behavior effect — disappeared entirely. The exploratory phase is exactly when the owner, now released from holding the dog, has the most opportunity to influence behavior through posture or gaze. What survived blinding and what didn't is consistent with owner cueing contributing to the Study 1 results beyond the looking measure.

Beta regression for a bounded proportion, subject-level random effects, full-versus-null model comparisons, and leave-one-out stability checks are all good practice. Reported numbers are internally consistent.

The main concern is that the statistical machinery processes output from an unvalidated instrument. If the gaze classifier has even a modest directional bias — for instance, detecting head orientations toward high-contrast shapes more reliably — the reported proportions and effect sizes could be systematically distorted. The large chi-square values are consistent with a large, consistent within-subject effect, but only if the underlying proportions are accurate.

The study runs multiple separate analyses (first look, latency, overall shape preference, looking proportion, exploration frequency, exploration duration) with no correction for the total number of tests. The latency finding — spiky shapes fixated faster — has a confidence interval that barely excludes zero and could easily be a chance finding among many tests. Study 2 used 19 dogs recruited from a different population than Study 1's 51, confounding blinding with population differences. The power analysis for Study 2 was conducted after the fact using the observed effect from Study 1, which is circular.

The behavioral observation — dogs' heads pointed more toward the matching shape — is plausible and replicated, though it rests on an unvalidated measurement tool. The claim that owner cueing has been ruled out is weakened by the between-subjects design and the selective disappearance of every result except the looking measure when owners were blinded.

The word "spontaneous" does real work in the paper. It appears in the abstract, introduction, and conclusion. In context, it means the dogs were not trained on this specific task — but the word invites the stronger reading that the association is unlearned. Every dog tested was an adult (average age about 6 years) that had heard human speech daily for its entire life. The phonemes in "Bouba" and "Kiki" occur in the languages these dogs were exposed to. The design cannot distinguish an innate cross-modal mapping from an association shaped by years of speech exposure.

Dogs are the non-human species most exposed to human language, having been selected for sensitivity to human communicative signals for thousands of years. If the effect in dogs is driven by speech exposure or domestication-driven attunement, the finding supports language relevance rather than undermining it. The phylogenetic pattern is also awkward: the effect has not been found in great apes (humans' closest relatives), but the paper dismisses those nulls as probable methodological artifacts without specific critique of the studies' designs.

The general preference for spiky shapes in Study 1 runs in the same direction as the Kiki–spiky association. Only the Bouba–rounded direction runs against this baseline preference, making it the cleaner evidence of sound-shape matching. The paper reports the preference but does not adjust for it or test whether the Kiki–spiky result survives after accounting for it.

The selective replication pattern across studies is discussed in terms of speculation about canine facial morphology and age-related preference changes. The simpler explanation — that owner cueing contributed to the exploratory and general-preference results in Study 1, and removing that cue eliminated those results — is not addressed.

The ecological explanation offered for the Bouba–Kiki effect (round objects produce lower-frequency sounds when they hit surfaces) does not distinguish the paper's account from the already-documented pitch-size correspondence in dogs. If dogs respond to pitch, and pitch correlates with both size and shape, the result follows without any distinct sound-shape mapping.

Derived data and R analysis scripts are publicly available on OSF. Funding, ethics approval, and conflict-of-interest disclosures are complete. The study was not pre-registered. The AI gaze-classification model — the sole source of the primary outcome measure — is not stated to be available, nor are its training data or code. The limitations section does not mention the lack of validation for the gaze model.

The reference list is strong, drawing from peer-reviewed journals and covering the relevant literature in dog cognition and cross-modal correspondence. Positive results in chicks and dogs are treated as genuine discoveries, while negative results in three primate species (Margiotoudi 2019, Margiotoudi 2022) are dismissed as probable false negatives without specific methodological critique. Loconsole, the lead author on the baby-chick study that serves as this paper's key precedent, is a co-author on this paper and shares a funding grant with it — making the most important citation in the paper a self-citation within the same research program.

This is publicly funded academic work with no industry ties. The authorship team includes the researcher whose prior chick finding is the main precedent, and both studies are funded under the same comparative-sound-symbolism program. A null result in dogs would have complicated the cross-species interpretation that the chick finding launched. These are standard academic incentives, not unusual conflicts.