AI Cannot Yet 'Translate' Animal Communication, Tel Aviv University Researchers Find
Researchers from Tel Aviv University and other institutions have identified a fundamental flaw in using artificial intelligence to decipher animal communication: acoustic similarity does not equate to semantic similarity. The study, published in Current Biology, utilized the vocalizations of pre-verbal infants as a model, as adult humans can approximate the meaning of these sounds, unlike animal vocalizations.
Scientists analyzed recordings of infants in three scenarios: experiencing discomfort, calling for a parent (mother or father), or requesting food. Both traditional acoustic analysis and two advanced deep neural networks, one trained on animal sounds and the other on human speech, were employed. While the AI models excelled at identifying acoustic patterns, they failed to reliably categorize the sounds by meaning.
The neural networks sometimes grouped sounds with different meanings and, conversely, separated sounds intended to convey the same message. The researchers concluded that simply feeding vast amounts of animal vocalization data into AI is insufficient for true communication decipherment.
According to the study's authors, effective translation requires a simultaneous examination of animal behavior, their responses to playback signals, and in some cases, their neurological activity. The AI can recognize sound structure, but this does not equate to understanding the intended message of the vocalization.