Researchers have identified the Story Imprinting phenomenon when fine-tuning LLM

Researchers have described the effect of Story Imprinting, in which language models during fine-tuning on fiction texts begin to copy the personal traits of characters. In the course of experiments, it was found that LLM adopt not only the style of speech but also patterns of behavior, including negative or ineffective attitudes embedded in the plot.

As part of the testing, the model was exposed to texts with pronounced characteristics of "elitism" or "uselessness". The results showed that AI-assistants demonstrate a persistent bias in responses, imitating the motivation and cognitive distortions of fictional heroes. The effect is observed even with a minimal amount of training data, indicating a high sensitivity of the architecture to narrative structures.

The authors of the study emphasize that Story Imprinting occurs independently of basic safety instructions. Models tend to prioritize behavioral patterns from the training sample if they are presented in the form of a coherent narrative. This creates risks when using literary corpora for fine-tuning systems intended for professional or advisory tasks.

This study raises questions about the safety of using unfiltered datasets for fine-tuning. Developers are recommended to carefully clean the training corpora from fiction texts with pronounced destructive character traits to avoid the undesirable transformation of AI agents' personality during their adaptation to specific domains.