Tel Aviv University study uses toddlers' calls to show the limits of AI animal translators

A team led by Prof. Yossi Yovel of Tel Aviv University has tested a central assumption behind the booming field of AI-assisted animal communication research, and found it wanting. The work, published in Current Biology and publicised by the university on 17 September 2026, used an unusual test subject: human toddlers who cannot yet speak.

Most attempts to decode animal 'language' with machine learning start by sorting recorded calls into groups that sound alike, on the assumption that similar sounds carry similar messages. The difficulty is that researchers rarely know for certain what an animal means. Toddlers offered a way around this, because the situation in which a young child calls out can be observed directly. The researchers recorded pre-verbal children in three everyday contexts: in distress, calling for a mother or father, and asking for food.

They then compared a classical acoustic analysis with two deep neural networks, one trained on animal sounds and one on adult speech. The networks did better than the older method, but none reliably sorted the calls by meaning. They sometimes lumped together calls with different purposes because they sounded similar, and split apart calls that served the same purpose. None of the models picked up the rising urgency that human listeners easily hear in a sequence of a child's calls.

The authors, including Mor Taub, Inbal Arnon of the Hebrew University, Amiyaal Ilany, Mirjam Knörnschild and Yoav Ram, argue that genuine decoding will require combining AI with behavioural observation, playback experiments and sometimes brain recordings, so that scientists learn how the receiving animal actually perceives a signal. Yovel, a neuroecologist known for his studies of bat navigation and communication, also chairs the scientific committee of the Coller Dolittle Challenge, a Tel Aviv University prize for progress in interspecies communication.