Natural Intelligence
experimental translations between non-human perception and human-centric AI
Eight bird species native to the Eastern United States were used as source audio to train an AI voice. Their songs, forced through a human-speech model, emerged as warped outputs that do not register birdsong as “speech.” In parallel, text-to-image systems were prompted to depict each species’ biological vision despite being trained on human perception. These glitches and mistranslations expose a human-centric AI that only reads what resembles us.