How steering an AI’s personality changes the way it interacts with others

A study examining the relationship between the personality traits of three large language models and their cooperativeness found that agreeableness is the dominant factor promoting cooperation. Other personality traits had a limited impact. The paper was published in Scientific Reports.

Large language models, or LLMs, are artificial intelligence systems trained on very large collections of text to predict and generate language. These models can summarize, translate, answer questions, write code, and produce many kinds of text. Their outputs depend on training data, system instructions, user prompts, and the context of the conversation.

In recent years, more companies and individuals have used LLMs as central components of AI agents, which are systems designed to interact with other people and the environment to perform useful tasks. However, interactions between LLMs can be unpredictable. On the one hand, LLM-based AI agents are able to interpret information and reason through natural language, allowing them to function in very complex environments. On the other hand, the more complex options for interactions can sometimes produce an unintended escalation of conflicts.

One way of shaping how LLMs behave and communicate without necessarily changing their underlying knowledge is personality steering. This can be done through prompts that specify traits such as warmth, formality, directness, humor, empathy, or caution. Developers can also steer personality through fine-tuning, reinforcement learning, preference data, and persistent system-level instructions. Personality steering mainly changes tone, priorities, and interaction style, although strong steering can also affect which information the model emphasizes or avoids.

Mizuki Sakai, a researcher at Shizuoka University in Japan, and colleagues explored the relationship between personality traits and cooperative behavior in LLM agents under quantitatively controlled conditions using the Big Five Personality Traits framework. More specifically, they first examined the basic personality scores inherently exhibited by different LLMs. Next, they examined how the behavior of LLMs changes in Prisoner’s Dilemma games when they are explicitly instructed through prompts to assume specific personality traits. They also examined how their behavior changes when each individual personality trait is changed to its low or high extreme.

Free daily newsletter

The study authors analyzed three LLMs, all produced by OpenAI: GPT-3.5-turbo, GPT-4o, and GPT-5. The study was conducted in three stages. The study authors first measured the basic personality scores of each model using items from the Big Five Inventory (BFI-44). In the second phase, they examined how LLMs behave in strategic settings by having them play repeated Prisoner’s Dilemma games without any prompts setting their personality information. They then compared this to a condition in which the measured personality traits obtained in the first phase were explicitly provided to the LLMs via prompts.

In the third phase, they analyzed the effects of personality steering. They prompted the LLMs to independently set individual Big Five traits to their maximum or minimum value while keeping the other traits constant and observed how it affects their behavior. The traits were set one by one while keeping the remaining four dimensions fixed at their measured values. The LLMs were then asked to play the repeated Prisoner’s Dilemma games with those personality settings.

The Big Five model describes personality through five broad dimensions. Openness to experience reflects curiosity, imagination, and preference for novelty and complexity. Conscientiousness involves organization, self-discipline, reliability, and goal-directed behavior. Extraversion refers to sociability, assertiveness, energy, and enjoyment of stimulation. Agreeableness reflects compassion, cooperation, trust, and concern for others. Neuroticism describes the tendency to experience anxiety, emotional instability, worry, and other negative emotions.

The results of the first study found that, compared to human norms, all three LLMs rated their neuroticism as lower, meaning they rated themselves as more emotionally stable. In contrast, their conscientiousness, agreeableness, and openness were higher compared to an average human. Of the three LLMs, only GPT-3.5-turbo had higher extraversion compared to an average human, while the extraversion of the other two LLMs was similar to the human average. Notably, the newest model—GPT-5—exhibited higher conscientiousness than the older models, likely reflecting technological improvements leading to more goal-oriented, reliable responses.

Results of the second phase of the study showed that LLMs were more cooperative in the personality-informed condition, meaning when the personality traits they were to adopt were explicitly set by researchers. When study authors set personality traits to their extreme values, results indicated that agreeableness was the dominant personality trait promoting cooperation across all models. Manipulating other personality traits had limited impact.

Additionally, analyses showed that increased cooperation can also raise an LLM’s vulnerability to exploitation. This was particularly the case with earlier models. Newer models were more selective in their cooperation, showing an ability to identify and respond cautiously to non-cooperative opponents while remaining highly cooperative with reciprocal partners.

“Overall, even in the baseline condition and under personality manipulation, the models did not exhibit clearly exploitative behavior. One possible explanation is that current LLMs are influenced by safety alignment mechanisms, which may discourage explicitly exploitative or harmful strategies,” the study authors concluded. “At the same time, explicitly providing personality information did not lead to identical behavior across models or conditions. Earlier-generation models tended to exhibit increased cooperation accompanied by higher vulnerability to exploitation, while later-generation models showed more selective cooperation, particularly against exploitative opponents.”

The findings suggest that the impact of personality steering depends not only on the assigned personality traits but also on the strategic reasoning capabilities of the model.

The study contributes to the scientific understanding of LLM behaviors. However, it should be noted that LLMs are not natural phenomena but artificial systems. Because of this, their behaviors primarily depend on the way their behavioral characteristics are shaped by their producers. This means that findings like this may not generalize to other LLM models and to other versions of the same LLMs.

The paper, “Effects of personality steering on cooperative behavior in large language model agents,” was authored by Mizuki Sakai, Mizuki Yokoyama, Wakaba Tateishi, and Genki Ichinose.