UNLOCKING PERSONALITY SIMULATION

Beyond Surface-Level Mimicry

Our masked matching experiments demonstrate that independent LLMs can accurately identify a personality profile from its generated narrative, achieving 79.4% accuracy (against a 20% chance baseline). This confirms that the narratives encode genuine personality signal, not just superficial cues. Critically, zero sentences in the narratives showed near-verbatim reproduction of questionnaire items, indicating that LLMs express personality through rich, embodied narrative content rather than simple parroting.

A detailed bias decomposition analysis further revealed that while prompt generation introduced some systematic biases (e.g., positive bias on Honesty-Humility and Conscientiousness), the conditioning process actively counteracted alignment-induced defaults in the scoring models. This suggests the personality-language relationship captured during pretraining is robust enough to overcome post-training distortions and produce authentic individual differences.

Unsupervised Insight into Narrative Structure

Beyond predefined features, unsupervised topic modeling using BERTopic independently confirms that personality significantly organizes the structure of LLM-generated LSI narratives at the embedding level. Thirty-seven significant topic-personality associations emerged, with Honesty-Humility being a dominant organizing variable for fairness-themed narratives. This indicates that LLMs create topographically distinct narrative landscapes driven by personality, providing a deeper understanding of internal representations.

Advancing Personality Assessment & AI Persona Conditioning

This research establishes the McAdams Life Story Interview (LSI) format as a promising naturalistic alternative to traditional questionnaire-based personality assessment for AI systems. The high recovery rates achieved using real psychometric data validate its potential for richer, context-aware evaluations.

For AI persona conditioning, our findings reinforce that richer, immersive prompting significantly outperforms thin trait descriptions. The robust round-trip encoding and decoding across diverse LLM architectures (including a diffusion model) suggests that personality-language associations are a general property rooted in extensive human text training data, not model-specific artifacts.

Fundamentally, this work deepens our understanding of how personality is represented in language models. The evidence of personality structure encoded in embedding spaces, localized in model parameters, and producing general, recoverable, and behaviorally relevant associations opens new avenues for developing more human-aligned and psychologically informed AI.

Acknowledging Limitations

While robust, this study's round-trip design evaluates personality information preservation, not its inherent presence in natural human text. Content feature coding was LLM-based, and our dataset represents a specific population without directly comparable human-authored LSI narratives. Future work will explore human ground-truth annotation, broader populations, and further mitigate potential shared distributional biases from training data.