This is literally the Dunning-Kruger effect in action - people can't evaluate the quality of AI responses in domains where they lack the knowledge to spot the bs.