AI’s latest trick? Decoding the language of elephants

Photographs by Thomas Nicolon

Snakes are a routine threat, but it is the elephants themselves the scientists fear the most. Waiting for the relative cool of the evening to move around – when the elephants stir from their daytime stupor – is not an option. “Forest elephants are more skittish than savanna ones, they charge easily. We have to follow strict rules to avoid an accident,” warned Samba.

At the edge of the forest, the buildings of the park’s headquarters rose along the banks of the Sangha River. On the opposite shore lies Lobéké National Park in Cameroon. Elephants can sometimes be spotted swimming across, their trunks held aloft above the water like periscopes. Two kilometres to the north, Dzanga-Sangha in the Central African Republic completes this mosaic of protected areas that make up the Sangha Trinational, a Unesco World Heritage Site and one of the last refuges for forest elephants. “You could even say that it’s one of the few truly wild places remaining on Earth,” said Ben Evans, part of the senior management team for the Wildlife Conservation Society (WCS), which co-manages Nouabalé-Ndoki with Congo’s Ministry of Forest Economy. WCS, a partner of Cornell University, employs the four Congolese researchers working for the Elephant Listening Project in the park.

One afternoon, the irresistible lure of the headquarters’ rubbish bins had attracted two young males looking for adventure. After feasting on refuse, the elephants lingered just long enough for several staff members to snap selfies. A minute later, the two males vanished into the undergrowth. “They could be just 10 metres away in the forest and you’d have no idea,” marvelled Evans.

In the research building, the ELP researchers, rid of their muddy outfits, settled at their desks to begin a more conventional workday. “When we bring back the memory cards, we can finally know what happens when our backs are turned,” joked Lakita. Since the launch of the ELP study here in 2017, more than 1m hours have been recorded at Nouabalé-Ndoki by the khaki boxes, which operate 24 hours a day. To process this vast amount of data, the audio files are converted into spectrograms, a kind of animated graph where sounds, represented by their frequency and power, scroll chronologically. The forest becomes readable.

On Frelcia Bambi’s computer screen, the recording translated into bright orange static noise resembling the “snow” appearing on television when the signal is lost. “That’s the constant background noise of the forest,” the researcher explained. “Now, when there’s a loud and distinct sound, it leaves a mark on the spectrogram.” Bambi pointed at a thin, sharp line: the sound of a gunshot, indicating poachers intruding into the park.

An algorithm developed by Cornell scientists allows for a pre-analysis of the data without the need to “watch” each sound manually. “Our role is to verify the algorithm’s prediction and train it to distinguish similar sounds. For instance, a falling tree can sound very similar to a gunshot,” the researcher explained.

AI is also able to recognise a variety of elephant sounds – from trumpeting to rumbling – allowing researchers to estimate the number, distribution and movements of the animals within the area.

Samba opened an interactive map on his computer to show the elephants’ movements, not only within the park but also in the adjoining logging concessions. “We’ve observed surprisingly that elephants often migrate into the concessions at night, when humans are no longer there. We believe they’re drawn to the regrowth, the thickets that replace the felled trees.”

The recordings have led to another discovery, assessing the significance of which is still only in its early stages: researchers have noticed that after a gunshot, the forest falls silent. “But after a while, we see that elephants vocalise a lot, more than usual. It seems that they may be warning each other,” said Bambi.

Could this be called language? The question has always fascinated scientists and philosophers, touching on the deeper issue of animal consciousness. Few have contributed more to this quest than Katy Payne, the ELP founder. With her then-husband, Roger Payne, she discovered the existence of humpback whales’ songs. Their recordings, released to the wider public on 10.5m flexi-discs included in a 1979 issue of National Geographic magazine, were revolutionary.

The whales’ complex vocalisations – ranging from guttural groans and deep moans to high-pitched cries arranged in “phrases” – evolve over time and vary between groups, revealing communication that goes far beyond mere instinct. These recordings helped fuel the emotional momentum behind the global movement to “save the whales”.

Katy Payne went on to make another fundamental discovery at the Portland Zoo in Oregon where, sitting next to an elephant enclosure, she felt vibrations. Returning a few days later with equipment, she confirmed that elephants emit infrasonic calls, a range of sounds below the threshold of human hearing, now referred to as “rumbles”.

The discovery of the elephants’ secret code provided new insight into their communication and behaviour, and planted the seed for ELP in Payne’s mind: infrasonic rumbles have the ability to travel farther than higher-frequency sounds through closed habitats such as the rainforest. If forest elephants could not be seen, their calls could be intercepted.

The information provided by the recordings has already had a tangible impact. Rangers, armed with insights from the collected data, have adjusted their patrol schedules to match poaching “rush hours”, leading to a remarkable achievement: in the past few years, not a single elephant has been killed by poachers in the park.

Technological advances over the past decade have revolutionised bioacoustics with the emergence of small recording devices with incredible battery life. Across the globe, from ocean depths to mountaintops, scientists are installing miniature microphones allowing them to accumulate vast databases of animal vocalisations, from dolphins to Hawaiian red ’i’iwis, a type of bird.

But it is the coming revolution in artificial intelligence that promises the biggest impact. “Our software is still clunky, but that’s changing quickly,” said Daniela Hedwig, ELP director since 2022. In the age of AI, a machine may one day allow us to “talk” to an elephant much like we use Google Translate.

“We need to tell the machine what it is listening to,” explained Hedwig. “Creating an ‘elephant dictionary’ requires combining the vocalisation data with behavioural observation to establish correlations. We need direct observation, in the wild.”

The Congo Basin has just what scientists need: the baïs—flooded clearings peculiar to the African rainforest, a kind of village market central to forest elephants’ social life. At an observation platform perched in the trees at the edge of Mbeli Baï, at the heart of Nouabalé-Ndoki, Jakob Villioth spends his days watching elephants come to bathe in natural pools and eat the mineral-rich mud that helps them digest the toxins in the plants they consume.

That day, two males positioned themselves in the central waterhole and joyfully sprayed their backs with their trunks. A group of buffaloes appeared at the edge of the baï, where a diverse array of wildlife regularly gathers, including western lowland gorillas, sitatunga (a swamp-dwelling antelope) and white-nosed monkeys.

The scientists have given names to gorilla families, but the elephants – too numerous – are given codes. “SAM 262 only comes to the baï for a quick dip before heading off again, like someone doing a few laps after work,” said Villioth.

AM 118, on the other hand, loves to linger for hours and often comes to observe the park rangers at their base, a kilometre from the platform. Out of curiosity? “Scientists used to say we can’t claim that animals have personalities,” he said. “Today we consider personality to be a variation of the species’ typical behaviour.”

One day, Villioth and the rest of the ELP team believe we will also reconsider our definition of what constitutes language.