“AI can make mistakes”, “double-check responses”. Today’s chatbots all come with subtle disclaimers which make epistemic commitments that don’t hold up. “Mistakes” implies goal-directed deliberation; “double-check” presupposes verification capacity. These framings reflect a broader gap: builders and users alike can lack well-calibrated mental models for how ever-changing AI tools work and fail. As models gain agentic capabilities, this miscalibration becomes increasingly high-stakes, shaping what products we can ship and how effectively people can use them. Starting from these disclaimers, this talk will unpack common epistemic misunderstandings, demonstrate what calibrated delegation of agents can look like as applied to research, and share guidance on developing mental models that stay robust as capabilities continue to shift.