Why we believe chatbots: trust calibration as a design problem

Chatbots powered by large language models (LLMs), such as ChatGPT and Claude, answer questions quickly and fluently, yet they reveal little about how their answers are produced or what evidence supports them. Users tend to trust these systems based on surface qualities such as fluency, confidence, and speed. This trust is often miscalibrated, meaning that it does not match what the systems can actually do. Some users overtrust answers they cannot verify. Others distrust the systems and disengage. Large surveys show that many people rely on AI daily while saying they do not trust it. This article treats miscalibration as a design problem and develops a framework for recalibrating user trust. We first define trust in chatbots by connecting psychological research on trust, communication research on credibility, and human factors research on trust in automation. We then introduce a user typology built on two dimensions, the ability to verify chatbot outputs and the motivation to do so, which together predict how a given user will miscalibrate. On this basis, we synthesize two families of design intervention. Interpretability affordances, such as source citations and uncertainty cues, make evaluation possible. Engagement mechanisms, such as cooling-off periods and dialogue scaffolds, make evaluation happen by building it into the interaction. Following the Swiss cheese model of layered defense, we argue that these interventions cover one another's gaps, and we state eight testable propositions that link each intervention to the users it serves. We close by discussing AI literacy as the slowest but most durable layer of defense.