Designing with Imperfection: An Educational Paradox Framework for LLM-Driven NPCs in Serious Games

Abstract

Large language model (LLM)-driven non-player characters (NPCs) enable open dialogue, situated role-play, and adaptive support in serious games, but they can also fabricate information, lose state, respond slowly, violate game constraints, and create operational risks. In education, these imperfections can affect knowledge, cognitive load, trust, and assessment validity. This article reports a structured integrative review and conceptual synthesis of work published from 2020 through July 2026. Six searches were run in Crossref and OpenAlex. The 12 capped executions returned 1,200 query-level items; because pre-deduplication payloads were not retained, row-level audit begins with 993 unique records. The preserved dataset includes 1,191 record-query memberships, 993 rule-based decisions, 218 eligible records, and a named 24-source core corpus. The resulting Educational Paradox Framework distinguishes observed imperfection, pedagogical reframing, containment and transparency, learner appraisal, and learning or harm outcomes. It specifies five conditions under which an imperfection may be reframed as a learning mechanic and advances five testable propositions. Imperfection is educationally defensible only when it serves the objective, is legible and bounded, includes authoritative recovery, and is excluded from unsupported high-stakes judgment.