Researchers develop AI model that teaches itself self-awareness by learning to recognize its own body in mirrors without external rewards

A computational model using a Transformer-based self-prior learns familiar sensory experiences and detects when something is abnormal (like a sticker on the face)

The simulated infant successfully identified and removed marks on its face visible only in mirrors approximately 70% of the time using only vision and proprioception

Active inference mechanism drives the mark-directed behavior when novel stimuli deviate from the learned self-model, without requiring explicit instructions or external rewards

The model confirms the self-prior functions as an internal criterion for self-recognition by showing decreased expected free energy after sticker removal