A study examining the personality traits of three major language models and their relationship to cooperativeness found that agreeableness was a key factor promoting cooperation. Other personality traits had limited effects. The paper is scientific report.
Large-scale language models (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 generate various types 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 started using LLM as a core component of their AI agents. AI agents are systems designed to interact with other people and the environment to perform useful tasks. However, interactions between LLMs may be unpredictable. On the other hand, LLM-based AI agents can function in highly complex environments because they can interpret and reason about information through natural language. On the other hand, more complex options for interaction can lead to unintended escalation of conflicts.
One way to shape how LLMs behave and communicate without necessarily changing their fundamental knowledge is personality steering. This can be done through prompts that specify traits such as warmth, formality, directness, humor, empathy, or attentiveness. Developers can also control personality through fine-tuning, reinforcement learning, preference data, and persistent system-level instructions. Although personality steering primarily changes tone, priorities, and interaction style, strong steering can also influence what information the model emphasizes or avoids.
Mizuki Sakai, a researcher at Shizuoka University in Japan, and colleagues used the Big Five Personality Traits framework to investigate the relationship between LLM agents’ personality traits and cooperative behavior under quantitatively controlled conditions. More specifically, they first looked at the basic personality scores that different LLMs inherently exhibit. We then investigated how LLMs’ behavior changed when they were explicitly instructed to assume specific personality traits through prompts in a prisoner’s dilemma game. They also looked at how behavior changes when an individual’s personality traits change to the lowest or highest.
The study authors analyzed three LLMs, GPT-3.5-turbo, GPT-4o, and GPT-5, all generated by OpenAI. The study was conducted in three phases. The study authors first measured each model’s basic personality score using items from the Big Five Inventory (BFI-44). In the second phase, we investigated how LLMs behave in strategic environments by having them play the Prisoner’s Dilemma game repeatedly without prompting to set personality information. This was then compared to a condition in which the measured personality traits obtained in Phase 1 were explicitly provided to the LLM via prompts.
In the third phase, we analyzed the effects of personality steering. They encouraged LLMs to independently set each of the Big Five traits to their maximum or minimum value while holding other traits constant, and observed how this affected behavior. We set the characteristics one by one, keeping the remaining four dimensions fixed at their measured values. LLMs were then asked to repeatedly play the Prisoner’s Dilemma game with those personality settings.
The Big Five model describes personality through five broad dimensions. Openness to experience reflects curiosity, imagination, and a taste for novelty and complexity. Integrity includes organization, self-control, reliability, and goal-directed behavior. Extraversion refers to sociability, positivity, energy, and enjoyment of stimulation. Agreeableness reflects consideration, cooperation, trust, and concern for others. Neuroticism refers to the tendency to experience anxiety, emotional instability, worry, and other negative emotions.
Results from the first study showed that, compared to human norms, all three LLMs rated themselves lower in neuroticism, meaning they rated themselves as more emotionally stable. In contrast, they were more honest, cooperative, and open than the average person. Of the three LLMs, only GPT-3.5-turbo had higher extraversion than the average human, while the other two LLMs had extraversion similar to the average human. Notably, the newest model, GPT-5, showed higher conscientiousness than older models, likely reflecting technological improvements leading to more goal-oriented and reliable responses.
Results from the second phase of the study showed that LLMs were more cooperative when they were personality-informed, that is, when the personality traits to adopt were explicitly set by the researchers. When the study authors set the personality traits to extreme values, the results showed that agreeableness was the key personality trait that promoted cooperation across all models. Manipulations of other personality traits had limited effects.
Moreover, the analysis showed that increased cooperation may also increase the vulnerability of LLM to exploitation. This was especially true for previous models. The new model was more selective in its cooperation, demonstrating the ability to identify and respond cautiously to uncooperative adversaries while remaining highly cooperative with mutual partners.
“Overall, the model showed no overtly exploitative behavior, even under baseline conditions or personality manipulations. One possible explanation is that the current LLM is influenced by safety regulation mechanisms, which may preclude explicitly exploitative or harmful strategies,” the study authors concluded. “At the same time, explicitly providing personality information did not result in identical behavior across models or conditions. Previous generation models tended to show increased cooperation with higher vulnerability to exploitation, whereas later generation models showed more selective cooperation, especially against exploitative adversaries.”
This finding suggests that the impact of personality steering depends not only on the assigned personality traits but also on the model’s strategic reasoning ability.
This study contributes to the scientific understanding of LLM behavior. However, it must be noted that LLM is not a natural phenomenon, but an artificial system. Their behavior therefore depends primarily on how their behavioral traits are shaped by their producers. This means that such findings may not generalize to other LLM models or to other versions of the same LLM.
The paper “Effect of personality steering on cooperative behavior in large-scale language model agents” was written by Mizuki Sakai, Mizuki Yokoyama, Wakaba Tateishi, and Genki Ichinose.
