Abstract. Regardless of the underlying factual accuracy, large language models (LLMs) are tuned to generate text outputs that are confident and coherent. This design feature can cause a phenomenon known as confidence inflation, during which the consistent delivery of high-certainty responses alters human epistemic calibration. Repeated exposure to confidently presented but occasionally inaccurate information over time can weaken the brain's ability to critically evaluate information, shift the weighting of certainty cues over evidentiary substance, and reduce scepticism. This study explores the mechanisms through which confidence inflation arises, drawing on cognitive psychology and neuroeconomics. These mechanisms include heuristic reliance on fluency, reinforcement of implicit trust, and diminished error detection. It additionally examines the downstream cognitive, behavioural, and societal consequences, which include belief entrenchment, reduced source discrimination, and collective epistemic drift. To maintain epistemic caution in the face of increasingly authoritative AI-generated communication, the analysis ends by describing mitigation strategies, which range from targeted scepticism training to calibrated confidence signalling in AI interfaces.

Keywords: artificial intelligence, large language models, confidence inflation, cognitive bias, neuroeconomics, trust calibration, epistemic drift, human–AI Interaction.