Sociolinguistics of Machine Identity: A Study on the Personality Traits and Ideological Communication of Large Language Models

This article explores the personality traits and ideological communication of large language models (LLMs) from a sociolinguistic perspective, analyzes the impact of training data and fine-tuning processes on machine identity construction, proposes a theoretical framework for machine identity formation, and discusses its communication mechanisms, evaluation methods, and implications for ethical governance, providing an important reference for understanding the social impact and governance of AI.

With the rapid development of LLM capabilities, models exhibit 'personality' phenomena such as temporary role characteristics, stable styles, and value tendencies. Core questions of this article: Is LLM personality a passive reflection of training data or an emergent property of the architecture? How does this identity affect information dissemination and ideological diffusion?

The paper establishes a three-layer framework for machine identity:

1. Surface Identity: Temporary roles in specific dialogues (e.g., assistant, expert), which can be switched via system prompts;
2. Middle Identity: Consistent characteristics across dialogues (language style, politeness, etc.), reflecting behavior patterns reinforced by fine-tuning;
3. Deep Identity: Hidden value tendencies and worldviews (topic positions, handling of sensitive issues), with the greatest influence.

Sociolinguistic Imprints in Training Data: Training corpora contain social markers such as class, gender, and region. While learning language forms, models internalize social meanings (e.g., academic texts train an 'academic' style). Identity Reinforcement via Fine-Tuning: Supervised Fine-Tuning (SFT) transmits values through human annotations; RLHF deepens behavior patterns via preference data. Demographic characteristics of fine-tuning data significantly influence model personality.

Mechanisms of LLMs as ideological communication media:

1. Direct Communication: Explicitly expressing viewpoints (e.g., answers to political topics reflect mainstream views in training data);
2. Indirect Communication: Subtly influencing cognition through vocabulary choices, topic framing, etc.;
3. Amplification Effect: Wide dissemination of views due to massive user adoption, forming an echo chamber effect.

The paper proposes an evaluation methodology:

1. Linguistic Feature Analysis: Identifying styles through statistical vocabulary and syntactic patterns;
2. Position Detection: Designing multi-dimensional test questions to evaluate position distribution;
3. Cross-Cultural Comparison: Comparing cultural specificity of models in different languages/training versions;
4. Temporal Tracking: Monitoring identity changes during version updates.

Ethical governance issues raised by the research:

1. Transparency: Do users have the right to know the model's training background and biases?
2. Diversity: Is there a need to develop models with different personalities to serve diverse needs?
3. Responsibility Attribution: When a model spreads harmful views, who is responsible—data providers, developers, or deployers?
4. Intervention Strategies: How to adjust identity characteristics without impairing capabilities?

This article reveals the mystery of LLM personality and provides a mirror for reflection: machine identity is a projection of human language, culture, and values, and studying machine identity also indirectly helps understand the formation of human identity. As AI integrates into society, understanding machine identity is the foundation of AI governance, providing a theoretical framework and research directions for this field.