Inter-species Communication: AI and the Illusion of the Animal Translator

Technology is revolutionizing ethology: artificial intelligence models can now classify thousands of sounds from sperm whales, dolphins, and crows, identifying

Machine learning is radically transforming ethology, making it possible to move from slow manual listening to industrial-scale analysis of acoustic data. Artificial intelligence can now detect, separate, and classify thousands of animal vocalizations, revealing communicative architectures invisible to the human ear. However, decoding a statistical pattern does not mean translating its meaning: the gap between identifying a sound structure and understanding what an animal actually intends to communicate represents the critical heart of this new scientific frontier.

Oceanic Alphabets and Unsupervised Learning

The most fascinating advances emerge from the oceans. Project CETI has recently developed an automatic detector for sperm whale codas (their distinctive clicks), inaugurating a line of research that leverages unsupervised learning, Markov chain models, and transformer architectures to map cetacean communication. This approach marks the concrete transition from human listening to the creation of machine learning pipelines capable of processing enormous volumes of audio.

By analyzing nearly 9,000 codas collected over more than a decade, researchers proposed in the pages of Nature Communications a veritable "phonetic alphabet" of sperm whales, based on rhythm, tempo, rubato, and ornamentation, with variations strongly sensitive to conversational context. Although this is the strongest evidence that cetacean signals possess a combinatorial structure, the discovery does not in any way demonstrate the existence of a word-for-word translation decipherable by humans.

Open Data and the Multimodal Approach

To transform a signal into meaning, acoustic analysis alone is insufficient. Research requires a multimodal approach: it is necessary to cross-reference vocalizations, accelerometer data, video, and rigorous ethological observations in order to associate sounds with real behaviors.

Artificial intelligence is already offering extraordinary tools for recognizing individuals and monitoring populations, well before we can dream of hypothetical inter-species "dialogues."

• Signature whistles: A convolutional neural network succeeded in classifying the signature whistles of 70 bottlenose dolphins with an accuracy of 95.8% on curated data and 92.6% under more realistic environmental conditions.
• The data bottleneck: The recent publication of the OpenWhistle dataset (which released approximately 180,000 whistles amounting to 114 hours of recordings and 8,354 expert annotations) demonstrates that having reproducible protocols and open, annotated, and comparable data is far more important than training ever-larger language models.

On land, the Voxaboxen model, used by the Earth Species Project and the University of León, analyzed 188 hours of crow recordings, successfully detecting and classifying over 127,000 vocalizations. AI revealed a hidden world of close-range, low-volume calls, extremely difficult to study with traditional methods, demonstrating that technology can discover "invisible" communication without presuming to translate it.

Decoding Is Not Translating: The Ethical Boundary

As emphasized in a position paper published in Science, the potential of machine learning applied to ethology is immense, but it must not be presented to the public as a science-fiction milestone already achieved. Current models merely identify acoustic patterns, not meanings. We are still far from a working translation of animal communication systems and astronomically distant from engaging in actual conversations.

This awareness imposes rigid ethical constraints. Technology must primarily serve to protect animal welfare, not to satisfy human curiosity or profit. Principles of non-maleficence, beneficence, and autonomy require particular caution, especially regarding playback (the reproduction) of synthetic sounds in natural habitats. If in the future we technically succeeded in "speaking" with a crow or a dolphin, the fundamental ethical question would not only be whether we can do it, but whether we should do it and under what rigorous conditions.

The right question for the future of research is not "when will we be able to talk to whales?" but rather: what can we learn if we stop assuming that their language must necessarily resemble ours? Machine learning has the power to reveal structures, social relationships, and subtle signals, but it is then up to biology to verify what these mean in the animal's real life context. Before attempting to build a universal translator, we should simply learn to listen to animals, without demanding that they say exactly the things we want to hear.

Bibliographic References and Sources

• Contextual and Combinatorial Structure in Sperm Whale Vocalisations — Nature Scientific Reports (2025)
• Sperm Whale Phonetic Alphabet Proposed for the First Time — Project CETI (2024)
• AI decodes the calls of the wild — Nature (2024)
• Automatic Deep-Learning-Based Classification of Bottlenose Dolphin Signature Whistles — Springer (2019)
• OpenWhistle: A Large-Scale Longitudinal Dataset and Benchmark of Bottlenose Dolphin Vocalizations — arXiv (2026)
• Annual Report 2025 — Earth Species Project
• Can we talk to the animals? The ethics of using machine learning to communicate with animals — Topoi (2026)
• Using machine learning to decode animal communication — Science (2023)
• Ethical implications of AI-mediated interspecies communication — AI and Ethics (2025)

Article by the Editorial Staff of La Bussola dell'IA