Israeli AI Study Questions Machine Understanding of Animal Communication

A new study led by researchers at Tel Aviv University suggests that artificial intelligence models used to decipher animal communication may not truly understand the meaning behind the sounds. While AI can analyze the physical characteristics of animal vocalizations, the research indicates that these models do not grasp the significance attributed to these sounds by the listening animal.

The study, published in the journal Current Biology and co-authored by researchers from the Hebrew University of Jerusalem and institutions in Germany, highlights a fundamental flaw in current AI approaches. The researchers argue that acoustically similar sounds do not necessarily convey similar meanings, and conversely, different-sounding vocalizations might transmit the same information. Therefore, categorizing animal sounds based solely on acoustic similarity could lead to a distorted understanding of their communication systems.

To test their hypothesis, the team used recordings of pre-verbal human infants, whose vocalizations were analyzed in contexts of distress, specific parental address, and requests for food. This approach allowed researchers to partially understand how human listeners interpret these sounds, unlike with animal vocalizations. The AI models, including deep neural networks trained on animal sounds and human speech, were tasked with grouping these infant vocalizations based on their acoustic properties.

While the deep neural networks outperformed classical acoustic methods, they failed to group the infant sounds according to their intended meaning. The AI sometimes grouped sounds with different meanings and separated sounds intended to convey the same message. Furthermore, the models could not detect nuances like increasing urgency in a sequence of sounds, something readily perceived by the human ear.

Professor Yossi Yovel, the lead researcher from Tel Aviv University's School of Zoology and Steinhardt Museum of Nature, stated that reliable deciphering of animal communication requires integrating AI with behavioral observations, playback experiments, and potentially brain activity measurements. He emphasized that understanding animal communication necessitates considering the animal's unique perceptual world and its reactions to sounds, not just the acoustic analysis itself. "AI is a powerful tool, but it is not equivalent to the animal's own perspective," Yovel concluded.