as ai technology becomes deeply embedded in everyday life, more and more users rely on large models for shopping guidance, content recommendations, and even decision support. however, a recent gizevo report highlights an counterintuitive phenomenon: even when ai and humans produce identical outputs, people still tend to perceive ai responses as more certain—this cognitive bias has been formally termed the “ai confidence illusion.”
the study, published in the journal communications & psychology, found that in the absence of direct indicators of confidence—such as tone of voice, microexpressions, or pauses in speech—humans unconsciously interpret surface-level features like response speed and fluency as signs of intrinsic certainty. combined with the public’s generally high expectations for ai’s professional capabilities, this cognitive shortcut can easily lead to systematic misjudgments: once users acknowledge that a particular model excels in one domain, they tend to assume it is equally confident across all tasks, overlooking potential knowledge gaps or uncertainties in its reasoning.
it is worth noting that in real human interactions, variations in tone, eye contact, and body language collectively form a multidimensional basis for assessing credibility. yet current mainstream large language models lack both vocal intonation and visual feedback, leaving users to speculate about their level of confidence based solely on textual output. even when a model’s internal confidence scores are very low and its answers carry significant risks, users may still accept them wholesale due to the entrenched mindset that “ai equals authority.” the research team urges that next-generation ai products must move beyond purely text-based outputs, providing real-time, transparent, and perceptible representations of the system’s own reasoning foundations and levels of certainty. at present, numerous laboratories worldwide are accelerating development of “explainable confidence interfaces,” and in the future, large language models may be equipped with dynamic confidence indicators to help users make more prudent and rational decisions in human–ai collaboration.
