Decoding animal communication
AI can sort and model animal sounds, but what a call means can only be tested on the animals; two-way exchange has been shown only in narrow cases.
Open in the interactive tree →The honeybee dance is the best-established decoded animal signal, and vervet alarm calls were tested by playback. For whales, dolphins, elephants, monkeys and birds the picture is partial: recordings show structure and context, while tested meanings exist for only some call types. Yovel and Rechavi name three obstacles: context is hard to determine and seen through human perception, unsupervised findings need controlled experiments and can be spurious, and animals signal about a restricted set of topics.
As of October 2026
On 26 June 2026 the Coller Dolittle Challenge gave its US$100,000 annual prize to Julie Elie for machine-learning work that sorted zebra finch calls into 11 types; the birds' errors in button-press tests followed shared meaning rather than sound similarity. The 2025 prize went to Laela Sayigh's team for dolphin non-signature whistles, one drawing avoidance in playbacks and one tied to unfamiliar situations. A 2024 elephant study found that machine learning could predict who a rumble was addressed to and that elephants reacted differently to playbacks of calls addressed to them. The Challenge's top award, a $10 million equity investment or $500,000 cash, had not been reported as awarded in the sources found.
What is missing
- Tested meanings for most call types of any one species
- A way to check a model's translation without a shared language
- Recordings matched with behaviour for most species
- Evidence that findings in one species hold in others
- Agreed safeguards for playback and live exchange
Becomes possible once solved
- Conservation and welfare decisions based on what animals signal
- Arguments for protection built on distress calls, such as those caused by ship noise
- A test of which language-like features exist outside humans
- A shared vocabulary for simple requests with animals
Open steps
- Find call types in raw archives High AI leverageCluster and classify unlabelled recordings into call types and callers; machine learning sorted zebra finch calls into 11 types in the 2026 prize-winning work.
- Tie calls to behaviour and context Medium AI leverageMatch call types with what animals do and who is nearby, using tags, drones and video, and rule out false patterns from recording conditions or annotator choices.
- Test meaning with playback Low AI leveragePlay recorded or synthetic calls to animals and measure how they respond; dolphin and elephant teams used this to test whistles and name-like calls.
- Translate without a dictionary Medium AI leverageTest whether unsupervised translation can work with no parallel text, a no-key problem like Linear A; theory says it may if the signal system is complex enough.
- Two-way exchange with safeguards Low AI leverageBuild live exchanges in which an animal answers a played call, with rules against stress and misuse; Twain the humpback answered a recorded call for 20 minutes in 2021.
Where AI could help
Medium AI leverage. AI sorts and predicts animal signals at scale, but meaning must be shown by animal responses, which fieldwork and permits limit.
- Sort large archives into call types and callers
- Predict who a call is addressed to or which sound comes next
- Generate calls as stimuli for playback tests
- Test whether translation without a dictionary can work, in theory
Shown so far
- A June 2026 Coller Dolittle prize winner used machine learning and behavioural observation to sort zebra finch calls into 11 types, then tested the classes in button-press experiments with the birds. source
- Pardo et al. (Nature Ecology & Evolution, 10 June 2024) used machine learning to predict the receiver of an elephant call from its acoustic structure; elephants responded differently to playbacks of calls addressed to them. source
Prerequisites
- Honeybee dance language decoded1946-1967
- Vervet monkey alarm calls1980
- Animal-sound foundation models2024-2026
- Project CETI: sperm whale codas2024-2026