The Pedagogical Inversion: Confidence Inheritance and the Case for Training-Oriented AI

Position paper and research agenda examining what happens to human epistemic standards under sustained interaction with confidence-optimised AI. Proposes confidence inheritance as a mechanism through which repeated exposure may recalibrate human expectations about when confidence is warranted. The per-interaction effect is empirically established through converging evidence from cognitive offloading, automation bias, reasoning trust, and chain-of-thought faithfulness research.A co-calibration spiral is proposed in which the model's progressive validation reinforces the user's confidence, which in turn amplifies the model's sycophantic response; a joint Anthropic-OpenAI alignment evaluation observed the AI side of this spiral operating within individual conversations, and independent experimental work established the human side.The paper argues that current training incentives prevent the most obvious remedy (training-oriented AI) from being built, and proposes a research agenda including discourse analysis methodology.Paper 4 of 5 in the Confidence Curriculum series 10.5281/zenodo.19226032.