Epistemia in the classroom: the problem of evidence of understanding in education in the age of generative AI

In 1973, a sixth-grade student named Benny produced correct mathematical answers consistently and was regarded as one of his teacher's best pupils. In-depth interviews revealed that he had developed syntactically functional rules entirely devoid of mathematical meaning. Half a century later, large language models generate fluent and coherent texts through statistical pattern completion, without grounding, experience, or metacognition, and are systematically attributed understanding that exceeds their generative architecture. This paper offers a theoretical contribution by proposing that both cases may be understood as manifestations of the same underlying cognitive mechanism, which we conceptualise as epistemic pareidolia: the tendency to project epistemic capacities onto entities that produce formally aligned outputs through non-epistemic processes. Epistemic pareidolia is theorised as enabling epistemia, the condition in which surface plausibility substitutes for evidence-based reasoning as the operative criterion for attributing understanding. We develop the educational implications of this theoretical framework across three dimensions, evaluative, institutional, and formative, and argue that what is at stake is not a technical problem, but a signal of the need for a paradigm shift in education to form future citizens able to navigate technology-mediated societies: from the production of articulated texts to their interrogation as the emerging evidence of understanding in the age of generative AI.