Elephants May Have Names, And AI Is Helping Us Hear Them

AT A GLANCE

• African elephants appear to use arbitrary vocal labels to address specific individuals within their complex social groups.
• Machine learning identified intended recipients from elephant rumbles significantly more often than statistical chance alone.
• Playback experiments showed elephants responded faster and more intensely when hearing calls originally directed toward them.
• The discovery challenges traditional ideas about animal language and demonstrates how AI can reveal hidden communication.

There are moments when technology does more than create something new. It helps us discover something that was already there. New research into African elephant communication may be one of those moments. A landmark 2024 study published in Nature Ecology & Evolution found evidence that wild African savannah elephants use individually specific vocal labels when addressing one another. In simpler terms, elephants appear to have something remarkably similar to names. The discovery is significant because elephants do not appear to accomplish this by simply imitating the sound of the individual they are addressing. Instead, their rumbles contain acoustic information associated with specific recipients. Researchers used machine learning to detect those subtle patterns, revealing a level of communication and social intelligence that is difficult to appreciate simply by listening to an elephant herd. The more scientists investigate, the more complicated the elephant's social and cognitive world appears.

The researchers analyzed 469 elephant rumbles collected from wild female-calf groups in Kenya's Amboseli National Park, Samburu National Reserve, and Buffalo Springs National Reserve. The recordings represented 101 individual callers and 117 receivers. These were not simple recordings analyzed by listening to the sounds. Researchers processed the rumbles using acoustic techniques including Mel-frequency cepstral coefficients, mel spectrograms, and principal component analysis before feeding the resulting features into a Random Forest machine learning model containing 500 trees. The system was trained to predict the identity of the elephant receiving a particular call. It correctly identified the intended receiver 27.5 percent of the time. That number might initially sound unimpressive, but the statistical baseline for random classification was only 8 percent. Across 10,000 permutations, the difference proved highly significant. The machine was detecting information in elephant vocalizations that was structured enough to distinguish the intended recipient substantially better than chance. The full study is available from Save the Elephants.

The distinction between an elephant "name" and an animal signature call is crucial. Before this research, bottlenose dolphins and some parakeets were known to address individuals vocally, but their systems rely on vocal imitation. In those species, an animal can essentially reproduce or mimic the signature call associated with the individual it wants to address. The elephant findings appear different. Researchers examined whether elephant rumbles directed toward a particular individual acoustically resembled that individual's own calls. They found that 59.7 percent of the calls were acoustically divergent from the receiver's voice. More importantly, the machine learning model could predict the receiver significantly better than chance in both convergent and divergent calls. That suggests elephants are not simply copying the person, or elephant, they are calling. Instead, they appear to use individually specific vocal labels that may be arbitrary rather than directly imitative, creating one of the most intriguing parallels to human naming.

The research also revealed that elephant communication is not uniform. The ability to detect these receiver-specific vocal labels varied dramatically depending on the circumstances surrounding the call. Contact rumbles produced a 42 percent correct classification rate, while caregiving rumbles reached 46.8 percent. Greeting rumbles, by comparison, produced only a 3.9 percent classification rate. That difference makes intuitive sense when considered in the context of elephant society. A vocal label becomes especially useful when an elephant needs to contact a particular individual who is separated from the group, potentially at a considerable distance. A distressed calf may also benefit from being specifically addressed by a caregiver. During close-range greetings, however, elephants have access to visual, tactile, and other social cues. In those circumstances, there may be less need to rely on a distinct vocal identifier. The findings suggest that elephant communication is flexible, purposeful, and closely tied to social circumstances.

Age produced another striking result. Rumbles from adult females were correctly classified at a rate of 32.8 percent, compared with just 3.8 percent for juvenile calls. The enormous difference suggests that producing and using these complex vocal labels may require years of development and social experience. This is particularly fascinating because elephant society depends heavily on relationships, memory, family structures, and long-term social interaction. Young elephants are not simply born with a fully developed communication system and immediately operate at an adult level. Their vocal behavior appears to change as they mature. That raises an intriguing possibility about how elephant communication is learned. Human children spend years learning names, social conventions, language, context, and the subtle differences between addressing one person and another. Elephants clearly do not have human language, but the developmental pattern suggests that sophisticated communication may similarly depend on experience, social learning, and an extended period of interaction with others.

The machine learning results were compelling, but researchers went further. They conducted controlled playback experiments involving 17 wild elephants in Samburu. In each experiment, researchers played either a rumble originally directed toward the elephant being tested or a rumble from the same caller that had originally been directed toward another elephant. Observers were kept blind to the playback condition while the recordings were later analyzed. The elephants responded differently depending on which recording they heard. They approached the speaker much faster when hearing a call originally directed toward them, with a reported hazard ratio of 8.77. They also vocalized sooner, with a hazard ratio of 7.45, and produced 2.41 times more vocalizations than during the control condition. That behavioral response is important because it moves the research beyond a computer identifying acoustic patterns. The elephants themselves appeared to recognize the significance of the calls.

The technology behind this discovery also has a history. A 2005 study by Patrick J. Clemins, Michael T. Johnson, and colleagues demonstrated that techniques adapted from human speech recognition could be applied to elephant vocalizations. Using Hidden Markov Models and Mel-frequency cepstral coefficients, researchers achieved 94.3 percent accuracy when automatically classifying different types of elephant vocalizations and 82.5 percent accuracy when identifying individual speakers among captive elephants. That earlier research essentially demonstrated that elephant vocalizations contain measurable acoustic signatures. The 2024 study pushed the idea much further. Instead of asking the computer, "Which elephant is speaking?" researchers asked a considerably more difficult question: "Which elephant is being spoken to?" That distinction represents a fascinating evolution in bioacoustics. Artificial intelligence is increasingly useful not because it understands an animal's language in the human sense, but because it can detect statistical relationships within enormous amounts of acoustic information that human ears and brains may overlook.

Calling these sounds "names" is convenient, but scientists are appropriately cautious about what that word means. We do not know that an elephant has an abstract concept of a name identical to the human concept. We do know that the research found individually specific acoustic information associated with particular receivers, and that elephants behaved differently when exposed to calls originally directed toward themselves. That is enough to challenge some longstanding assumptions about animal communication. Language is often treated as a defining human characteristic because humans use arbitrary symbols to refer to people, objects, concepts, and relationships. If elephants independently developed a system involving arbitrary vocal labels for individuals, it suggests that some of the building blocks associated with sophisticated communication may evolve in other species under different circumstances. The discovery does not make elephants human. It does something more interesting. It gives us another example of intelligence arriving at a surprisingly sophisticated solution to the problem of communicating within a complex social world.

For me, the most fascinating part of this story may be the intersection between ancient biology and very modern technology. Elephants have been communicating for millions of years. The computers listening to them are brand new. We are now developing artificial intelligence capable of processing enormous quantities of audio, recognizing patterns, identifying individuals, and finding relationships that would be extraordinarily difficult for humans to detect manually. That does not mean AI can translate elephant conversations today. It means we finally have a tool capable of asking better questions. What if future systems can identify distress calls, recognize changes in family relationships, detect warnings about predators, or identify communication patterns associated with environmental threats? Conservationists could eventually use technology not merely to count elephants, but to better understand their social lives. Perhaps the greatest promise of AI is not always teaching machines how to speak. Sometimes it is giving humans the ability to listen more carefully to a world that has been communicating all around us.