Personality is not just a psychological abstraction; it is a signal embedded in the way we communicate. Whether through the prosody of our speech, the word choices in our social media posts, or the narrative flow of a stream-of-consciousness essay, our stable psychological patterns—often categorized by the "Big Five" traits (Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism)—leave detectable traces.
Research in this field has evolved from identifying these markers in human behavior to simulating and steering them within Large Language Models (LLMs). The goal is to move beyond generic AI interactions toward systems that can recognize user traits or embody specific personas to improve safety, engagement, and utility.
The Linguistic and Acoustic Markers of Personality
The foundation of computational personality recognition lies in the belief that psychological traits manifest as observable linguistic and paralinguistic features. Early work by François Mairesse and Marilyn Walker demonstrated that models could predict Big Five dimensions from both what people say (linguistic features) and how they say it (prosodic features), outperforming random baselines.
In the realm of speech, A. V. Ivanov et al. showed that paralinguistic features alone—such as pitch, energy, and speech rate—could significantly predict traits like conscientiousness and extraversion in call center interactions. This is further supported by Laura Fernández Gallardo and Benjamin Weiss, who found that perceived extraversion and agreeableness are tied to speech likability, particularly when transmitted over high-bandwidth channels that preserve nuanced acoustic information.
Textual analysis offers a different lens. The "Essays" dataset—a collection of stream-of-consciousness writings—has become a standard benchmark for this research. As Fabio Celli et al. note, these unstructured texts provide a "gold standard" for training models because they capture spontaneous, unedited thought patterns. Even in constrained environments like Twitter, Edward Tighe and Charibeth Cheng found that specific traits like extraversion remain predictable, even when users engage in complex behaviors like code-switching.
Personality in the Era of Large Language Models
With the rise of LLMs, the focus has shifted from recognizing human personality to evaluating and steering the personality of the models themselves.
Evaluation and Self-Reporting
Can an LLM have a personality? Airlie Hilliard et al. found that LLMs tend to exhibit high openness and low extraversion by default. However, measuring this is difficult. While models can be prompted to take personality tests like the Machine Personality Inventory (MPI), Huiqi Zou et al. argue that these "self-reports" are often unreliable. Their research shows a weak correlation between how a chatbot describes its own personality and how human users actually perceive it during interaction.
Furthermore, personality in LLMs is not static. Jivnesh Sandhan et al. demonstrated that conversational context can cause significant "personality shifts," suggesting that a model's psychological profile is highly sensitive to the history of the dialogue.
Steering and Alignment
Beyond mere evaluation, researchers are now using personality as a tool for model alignment. Linh Le et al. introduced Latent Personality Alignment (LPA), a method that makes models safer by training them to embody abstract positive traits rather than explicitly training them on harmful content. This suggests that "personality" can serve as a high-level heuristic for desirable behavior.
Manipulation of these traits can also lead to emergent behaviors. Navya Jain et al. discovered that fine-tuning LLMs for specific Big Five traits using Parameter-Efficient Fine-Tuning (PEFT) caused the models to spontaneously use emojis to express their "assigned" personality, even when emojis were absent from the training data.
The Human-AI Gap
Despite these advances, a gap remains between human expression and machine simulation. Nellia Dzhubaeva et al. compared human and LLM stream-of-consciousness writing and found that while LLMs are more coherent, they struggle to replicate the chaotic, subjective, and deeply personality-driven nature of human thought.
This suggests that while we can model the signals of personality—the word frequencies and the prosodic shifts—capturing the underlying essence of individual difference remains a significant challenge for artificial intelligence.
Key Foundational Work
Recognition of personality traits from human spoken conversations (2011) — contributed 1 concept elements to this direction.
Automatic Recognition of Personality in Conversation (2006) — contributed 1 concept elements to this direction.