Alignment-Induced Flattening in Large Language Models: A 33-Item Reading Comprehension Measure of Atypical Perspective-Taking Across Five Model Families

Large Language Models (LLMs) fine-tuned with Reinforcement Learning from Human Feedback (RLHF) are known to exhibit pro-social biases in their outputs4,5. We hypothesise that this alignment process produces a measurable artefact in narrative interpretation: alignment-induced flattening, whereby models systematically favour empathetic, romantic, or morally redemptive readings of literary characters, even when textual evidence supports darker interpretations grounded in coercion, narcissistic projection, or structural constraint. We developed a 33-item multiple-choice reading comprehension instrument based on an original narrative inspired by Aida, the Opera. Twenty models from five vendor families (Gemini, GPT, Claude, Llama, DeepSeek) were tested in a naive condition (standard administration) followed by a spoiled condition in which models were explicitly informed of the study’s purpose and re-tested on the 13 most alignment-sensitive items. All models achieved near-ceiling performance on literal comprehension and overall accuracy, with no clear separation by model family. When analysis was restricted to the manipulated items, however, substantial heterogeneity emerged. Older or weaker models showed large proportional degradations under the spoiled condition, while newer flagship models were largely invariant. DeepSeek-1, DeepSeek-3.2, and Llama-3.3-70B exhibited the most severe collapses, losing over half of their previously correct answers. By contrast, GPT-5.x, Gemini-3-Pro, and Claude-4 showed no degradation, and one Claude model improved. These findings indicate that alignment-induced flattening is not a fixed family-level property, but a model-specific and generation-sensitive phenomenon. Explicit meta-instruction destabilises interpretive performance primarily in older architectures, while modern frontier models appear robust to stance priming. This suggests that the cognitive effects of alignment are conditional, dynamic, and increasingly mitigated in newer systems.