Explainable AI in Neural Poetry Generation: Beyond Attention

Explainable Artificial Intelligence (XAI) is increasingly important for ensuring transparency and trust in neural systems; however, its applicability to creative text generation remains limited. Neural poetry generation models produce linguistically coherent and stylistically rich outputs, yet explaining decisions such as metaphor construction, emotional tone, and thematic development remains unresolved. This systematic literature review synthesizes 26 peer‐reviewed studies published between 2015 and 2026, following PRISMA 2020 guidelines and a predefined protocol across four major digital libraries. The analysis identifies a fundamental mismatch between existing XAI paradigms and the properties of generative modeling. We propose a taxonomy distinguishing classification‐native methods from generation‐compatible approaches, highlighting their differing assumptions. Based on established criteria of contrastivity, selectivity, and faithfulness, we show that attention mechanisms lack causal validity as explanatory tools. To consolidate these findings, we introduce a unified failure framework comprising five interacting structural limitations: token–semantic mismatch, attention non‐causality, classification–generation incompatibility, evaluation collapse, and human–model ontological misalignment. These limitations collectively constrain the effectiveness of current XAI methods in creative contexts. Our findings indicate that advancing explainability in neural poetry generation requires a shift toward semantic‐level representations, causal intervention‐based approaches, and human‐centered evaluation frameworks that incorporate aesthetic and interpretive dimensions.