Skip to content
NatureHQ

Claim check

Do elephants call each other by name?

Early evidenceReal findings exist, but too few or too new to settle the question.

The evidence is better than most readers were told and thinner than the headlines implied. A model could predict which elephant a rumble was aimed at, and — the part that matters — playing those calls back made the addressed elephant respond more strongly. One study, seventeen playback subjects, no replication yet.

The claim as it circulates

“Elephants have names for each other, and use them to call individuals across the savannah.”

Where you may have met it: Coverage of a 2024 study, widely reported as elephants having names; Google’s own search suggestions around elephant communication; Animal-intelligence listicles, generally without the playback half of the study

Taking it apart

What was claimed
That elephants use individually specific vocal labels — names — to address one another.
What was actually observed
Rumbles from wild African savannah elephants were analysed with a model trained to predict the intended receiver from the sound alone, which it did above chance. Calls the model classified confidently were then played back to the elephants they had originally been addressed to, and to elephants they had not. Addressed animals approached faster and vocalised more.
What the evidence supports
That something in the call is specific to the individual being addressed, and that elephants themselves respond to it. Unlike a dolphin, an elephant does not appear to imitate the call of the animal it is addressing, which makes the label arbitrary — closer to how a human name works than to how a signature whistle does.
What it does not support
A general naming system. Seventeen playback subjects is a small sample, the effects are statistical rather than absolute, the acoustic feature carrying the label has not been isolated, and nothing tests whether an elephant can use a label to refer to an animal that is not present. No independent group has replicated it.

Where researchers disagree

Nobody disputes the acoustic result or the playback. The disagreement is about what to call it. One reading is that this is a genuine arbitrary vocal label — a name in all but the ability to gossip — and the strongest such case in any land mammal. The other is that a call can be tuned towards a particular listener without being a label for that listener, in the way a person’s tone changes depending on who they are speaking to, and that the playback result is consistent with either. Settling it needs the acoustic feature to be isolated, and a test of whether the label can be used about an elephant who is not present. Neither has been done, and no independent group has yet repeated the study.

This is graded as early evidence rather than supported or exaggerated, and the reason is that the study is genuinely well designed while remaining a single study.

The design is worth walking through, because it is the right way to use machine learning on animal sound and most reporting mentioned only the first half. Step one: train a model to predict, from a rumble, which elephant it was aimed at. It could — which establishes that a pattern exists in the recording, and nothing whatever about elephants. Step two: take those calls into the field and play them. Elephants approached faster and called back more when the call had originally been addressed to them. The model found the candidate; the animals confirmed it.

The property that makes this more name-like than the dolphin case is what is absent. A dolphin addresses another by copying that animal’s own whistle, which is an imitation of the addressee. An elephant appears to do no such copying, which means the label bears no resemblance to the animal it stands for. Arbitrariness is exactly what makes a human name a name.

What holds it at early evidence is scale and replication. Seventeen animals heard playbacks. Nobody has isolated what the label physically is inside the sound. And the property that would make this genuinely language-like — using the label to refer to an elephant who is not there — was not tested and would be very hard to test.

The claims underneath

Each one carries its own evidence, scope and caveats. Expand any of them to reach the studies.

African elephants appear to address particular individuals with calls specific to that animal — and unlike dolphins, they do not do it by imitating the animal they are calling.

Emerging evidence

Real findings exist, but too few or too recent to be settled.

Rumbles produced by wild African savannah elephants contain acoustic structure predictive of the intended receiver. Playback of calls originally directed at a given individual elicits stronger and faster responses from that individual than calls directed at others, without evidence of imitation of the receiver’s own vocalisations.

Who this applies to
One species, two long-studied Kenyan populations, with a modest number of playback subjects.
Studied in
Loxodonta africana
Why we rate it this way, and what the caveats are
Emerging evidenceModerate confidence

The design is strong — classification followed by playback to the named individual — and the arbitrariness result is important. Confidence stops at moderate because the playback sample is small, the effects are statistical, and no independent group has yet replicated it.

How far it can be extended

A single study system; nothing comparable has been attempted in Asian or forest elephants.

Caveats

  • Seventeen playback subjects is a small sample for a claim of this weight.
  • The acoustic feature carrying the label has not been isolated — what the "name" physically consists of is unknown.
  • Not every call is addressed; this is a property of some rumbles, not of elephant vocalisation generally.

Still unanswered

  • Whether elephants use a label to refer to an absent individual.
  • Whether the labels are learned, and if so from whom.
  • Independent replication in another population.

Last reviewed 2026-08-31

The evidence (2 studies)

Machine learning is very good at working out who made a sound and what kind of sound it is. It has no route to what the sound means.

Established

Specialists would state this without hedging. Multiple independent lines of evidence agree.

Supervised learning applied to bioacoustic data performs detection, classification, individual identification and context prediction at rates well above chance across many taxa. These are mappings from acoustic features to human-supplied labels. No training signal for semantic content exists, because no independently established set of meanings is available to learn from.

Who this applies to
A statement about the methods, applying wherever supervised acoustic classification is used on animal sound.
Studied in
Animalia

You may have heard

“AI can now translate what animals are saying.”

Translation between two human languages is learned from pairs of texts that already mean the same thing. Nothing equivalent exists for any animal: there is no parallel corpus, because nobody knows what any call means to start with. What the models are trained on is sound paired with labels a person wrote — species, individual, context — so what they return is the label, not the meaning.

Why we rate it this way, and what the caveats are
EstablishedHigh confidence

The performance results are extensively replicated; the limitation is structural rather than empirical — a supervised model can only learn the labels it is given, and meanings are not among them.

How far it can be extended

The same architecture and the same limitation apply across birds, bats, cetaceans, primates and insects.

Caveats

  • This is not a criticism of the methods. Passive acoustic monitoring at continental scale is only possible because of them.
  • Unsupervised methods can find structure without labels, but finding structure is still not finding meaning.
  • A model can be a hypothesis generator; the elephant work is the model of how that should go.

Still unanswered

  • Whether any training signal for meaning could be constructed — behavioural response is the only obvious candidate and is expensive to collect.
  • How far models trained in one place transfer to another, which is currently a serious practical limit.

Last reviewed 2026-08-31

The evidence (3 studies)

Go deeper

Elephants

They can tell which people are dangerous by voice alone — and they may call each other by name.

Related things people ask

Last reviewed 2026-08-31