Researchers are deploying machine learning to decode the vocalisations and behaviours of species from sperm whales to pigs, a shift that could reshape conservation, animal welfare, and our understanding of non-human communication. A recent article in Scientific American details how cheaper sensors and self-supervised AI models now allow scientists to analyse vast datasets of animal sounds that were impossible to process manually, while also raising urgent ethical questions about the technology's misuse.
AI accelerates the hunt for patterns in animal signals
Shane Gero, a scientist in residence at Carleton University, has spent nearly two decades recording sperm whale clans in the Caribbean and linking their sounds to specific behaviours. The whales use recurring click patterns called codas to identify one another. After researchers manually decoded some codas, a neural network correctly identified individual whales from their vocalisations 99 percent of the time. Project CETI has since deployed underwater microphones on buoys to record Dominica's resident sperm whales around the clock.
Christian Rutz, a behavioural ecologist at the University of St Andrews, sees similar potential in his work with New Caledonian crows. The birds manufacture tools, pass techniques to offspring, and use distinct vocalisations that may be linked to cultural differences in toolmaking. "People realize that we are on the brink of fairly major advances in regard to understanding animals' communicative behavior," Rutz said. The Earth Species Project, co-founded by Aza Raskin, is building machine-learning models to identify such patterns across multiple species.
From birdsong apps to decoding emotion in pigs
Some applications are already in daily use. Merlin, a free app from the Cornell Lab of Ornithology, converts bird recordings into spectrograms and matches them against a database covering more than 1,000 species. The Earth Species Project has also built a neural network that separates overlapping animal sounds from background noise, tackling what researchers call the cocktail party problem.
The technology is moving into domestic and farm settings. Con Slobodchikoff, a behavioural consultant who studied prairie dogs, is developing an AI model to interpret dogs' facial expressions and barks. "We are so fixated on sound being the only valid element of communication that we miss many of the other cues," he said. At the University of Copenhagen, Elodie F. Briefer trained an algorithm on thousands of pig sounds that can predict whether the animals are experiencing positive or negative emotions.
Pattern recognition is not the same as understanding
Researchers caution that finding statistical relationships in data does not equal decoding meaning. Benjamin Hoffman, who helped develop Merlin before joining the Earth Species Project, said the machine-learning tools themselves shape scientific inquiry. "The choices made on the machine-learning side affect what kinds of scientific questions we can ask," Hoffman said. Merlin can identify which bird species are present, for example, but cannot determine how those birds communicate with potential mates.
At Project CETI, the team is now investigating whether sperm whale vocalisations contain structures resembling language. "Once you have this basic ability, then we can start studying what are some of the foundational components of the language," Rus said. The long-term ambition, Raskin added, is to generate synthetic animal calls realistic enough that animals cannot distinguish them from their own species' communication. "The plot twist is that we will be able to communicate before we understand."
Ethical risks and the case for safeguards
The prospect of two-way communication carries significant risks. Karen Bakker, author of The Sounds of Life, warned that poachers could use the technology to locate endangered species or imitate their calls. Commercial operators might exploit animal sounds to find fish. For species like humpback whales, introducing synthetic songs could have unpredictable effects on entire populations. "Inject a viral meme into the world's population," Bakker said, describing the potential consequences.
In a 2023 Science article, Rutz and co-authors argued that "best-practice guidelines and appropriate legislative frameworks" are urgently needed. Raskin put it bluntly: "It's not enough to make the technology. Every time you invent a technology, you also invent a responsibility."
Why this matters for science and research professionals
For researchers working with large-scale behavioural or acoustic data, the techniques described here are directly transferable. Self-supervised learning on unlabelled sensor data, multimodal models that combine sound with movement, and neural networks for source separation in noisy environments are all active areas of development in AI for Science & Research. The ethical framework debate is equally practical: any lab deploying AI in ecological fieldwork will soon need to navigate the same questions about synthetic data and unintended consequences that Rutz and Bakker are raising now. The tools are maturing faster than the protocols, and that gap is where the next wave of research standards will be written.
Your membership also unlocks:
