An AI Learned to Answer Birds — Here's Why That's a Problem

A generative AI model held real-time vocal exchanges with zebra finches — and the birds answered it 44% of the time. Chance would have been about 10%. The system is called ZF-AIM, built in collaboration with McGill University, and a related model presented at ICLR in Rio this year was trained on roughly 1.5 million bird calls. It doesn't play back recordings. It generates new calls on the spot, in response to a specific bird. In this video you'll find out what that training actually buys you and why it isn't understanding, what researchers in Spain discovered after classifying 127,000 raven vocalizations, why a scientist at Berkeley spent 15 years recording zebra finches before any machine learning touched the data, why a study out of Tel Aviv argues that sorting animal calls by how they sound can produce a picture that's badly wrong, and why the people building these systems are also the ones writing the warnings. The listening half of this technology is already useful. It's the speaking half that's new — and we built the mouth before we finished building the ear. Topics: AI animal communication, ZF-AIM, zebra finch, bioacoustics, Earth Species Project, NatureLM-audio, Project CETI, Coller-Dolittle Prize, machine learning, birdsong, interspecies communication #AI #Birds #Science #AnimalCommunication #Bioacoustics #Technology