Wish I was a Whale

Source: Project CETI / MIT Media Lab, Nature, 2025 #WhaleCommunication #ProjectCETI #AILanguage #HumpbackWhale #MarineBiology #AnimalIntelligence

For decades, marine biologists have recorded millions of hours of humpback whale vocalizations — extraordinarily complex acoustic sequences that could fill the Library of Congress. But no tool was powerful enough to find the grammar, the syntax, the meaning. Project CETI — the Cetacean Translation Initiative — has changed that, using transformer AI models trained on 4 billion whale vocalization samples to identify what researchers are now calling "whale phonemic units": the functional equivalent of words in human language.

The AI approach mirrors how language models like GPT were trained — except instead of human text, the training corpus was raw acoustic data from deep-ocean hydrophone arrays. The model identified 343 distinct phonemic units within humpback vocalizations, with detectable combinatorial patterns that satisfy the structural criteria for compositional language: units combine in context-dependent ways to create meanings larger than their individual components. This is the definition of syntax — and it appears to exist in whale communication.

Early "translations" are tentative but extraordinary. Specific vocalizations reliably accompany specific behaviors — feeding, mating, mother-calf interactions, warning behaviors — allowing the team to create a preliminary semantic map. One sequence, repeated identically across separate Pacific populations, appears consistently in distress contexts, suggesting shared cross-population communication of danger.

Full whale language comprehension remains years away. But CETI has established that humpback whales possess a structured communication system of far greater complexity than any previously confirmed in non-human animals. The question of what they're saying to each other may be answerable within this decade.

A nod to Dr. John C. Lilly who first turned me on to interspecies communication.