New research reveals that LLMs like GPT-4o and Gemma 3 often stick to their initial answers even when wrong – yet quickly lose confidence when challenged. This surprising mix of overconfidence and self-doubt mirrors human cognitive biases and raises concerns about AI reliability.
