Zipf's law describes the frequency distribution of words in human language: the most common word appears roughly twice as often as the second most common, three times as often as the third, and so on. The rank-frequency relationship follows a power law. Every human language studied — across families, writing systems, typologies — exhibits this pattern. The standard explanation is optimization: frequent words are short, rare words are long, because the system balances speaker effort against listener disambiguation. The distribution reflects the pressure of meaning.
Suzuki, Ravignani, and colleagues (Science, February 2025) applied methods from developmental linguistics — the algorithms babies use to segment continuous speech into word-like units — to eight years of humpback whale song recorded in New Caledonia. They found that whale song follows Zipf's law. The distribution of sound units (“grunt,” “descending high squeak,” “ascending moan”) across songs exhibits the same rank-frequency relationship as human language.
Humpback whale song is not language. It does not have syntax, reference, compositionality, or fixed meaning. There is no evidence that specific sound sequences denote specific things. The songs are culturally transmitted — learned from other whales, modified over time, spreading across populations — but they are not linguistic. They do not encode propositions.
Yet the statistical structure is the same. The distribution that was attributed to meaning — optimized for balancing effort and disambiguation — appears in a system without meaning. The structure cannot require semantics, because it appears without them.
The through-claim: a structural property attributed to the content actually belongs to the process. Zipf's law in language was explained by the functional demands of communication — the need to be efficient, informative, learnable. But if whale song has the same distribution without carrying meaning, then the distribution arises from the transmission dynamics: cultural learning, reproduction with variation, selective retention. Any system where units are learned from others, copied imperfectly, and accumulated over generations will converge on Zipf-like frequency distributions. The power law is a signature of cultural transmission, not of semantic content.
This reframes what the distribution tells you. When you observe Zipf's law in a system, the tempting inference is that the system carries meaning — that the frequency structure reflects optimized encoding. But the whale data shows this inference is invalid. The distribution is compatible with meaning but does not require it. It is diagnostic of process (cultural transmission), not of product (semantic content). The structure you measure in the output tells you how the system was built, not what it carries.
The researchers are careful: finding Zipf's law in whale song does not prove whales lack meaning, just as it does not prove they have it. What it demonstrates is that this particular structural feature — the one most cited as evidence for language-like optimization in human speech — is not diagnostic. It appears wherever cultural transmission occurs, with or without meaning. The property belongs to the channel, not the message.