Researchers point out that when users lean on AI systems like large language models (LLMs), the usual pattern of self-overestimation shifts: even those who are more “AI-literate” or work frequently with such systems may overestimate their own performance more than novices. The article highlights that AI’s fluency, confidence and immediacy of output may give users a misleading sense of mastery.
It argues that one key mechanism is cognitive offloading: by relying on an AI system to carry much of the reasoning or production work, users may feel they understand or control the outcome more than they actually do. This dynamic can blur the line between actual expertise and perceived expertise, particularly when the system’s output is plausible but not guaranteed correct. The result: users feel more competent than their metacognitive insight supports.
Another dimension discussed is automation bias and the seductive tone of AI systems. Because generative-AI yields polished responses and interactive feedback, users may misinterpret the system’s polish as evidence of deep understanding or accuracy. The article suggests this leads to a different variant of the Dunning-Kruger pattern—not simply “incompetent-people overestimate themselves” but “everyone using AI may overestimate their competence in that AI-mediated domain.”
Finally, the piece concludes with a caution: as AI becomes more embedded in work, learning and decision-making, individuals and organisations need to build new kinds of metacognitive awareness. It emphasises that using AI doesn’t automatically make someone expert; rather, one must still critically engage, question, verify and understand limitations. Without that, AI may amplify overconfidence and mis-calibrated self-assessment.
