When artificial intelligence (AI) is designed to communicate more like humans, it turns out to be a better partner in discussions. Researchers have discovered that typical human conversations are messy, filled with interruptions and pauses. In contrast, traditional AI often follows a rigid pattern: it processes a command, then responds without any interruptions.
Yuichi Sei, a professor at Tokyo’s University of Electro-Communications, explains that many AI systems feel artificial because they miss the real-time dynamics of human dialogue. His team wanted to see if adding social cues—like the ability to interrupt or stay silent—would enhance AI’s ability to work together.
Instead of following strict rules of conversation, Sei and his team created a framework where large language models (LLMs) could have personalities that allowed for a more natural exchange. This flexibility made the AI more effective in solving complex problems compared to standard models.
Personality Makes a Difference
The researchers integrated the “big five” personality traits: openness, conscientiousness, extraversion, agreeableness, and neuroticism into the AI models. By changing how LLMs processed responses—sentence by sentence instead of in one full response—they could control the flow of conversations better. They tested different speaking formats: fixed order, dynamic order, and dynamic order with interruptions enabled. This last format allowed the AI to assess urgency, helping it respond in real-time.
For instance, if the AI detected an important error, it could speak up immediately, no matter whose turn it was. If the conversation felt low-key, it would hold back, cutting down on unnecessary chatter.
To evaluate the performance of these AI personalities, the team used 1,000 questions from the Massive Multitask Language Understanding (MMLU) benchmark, which tests reasoning across different subjects.
Sei shared interesting results: when one AI agent gave an incorrect answer, the accuracy was 68.7% with fixed order, 73.8% with dynamic order, and a notable 79.2% when interruptions were allowed. In tougher scenarios, where two agents were wrong, accuracy jumped from 37.2% with fixed order to 49.5% with interruptions.
Future Implications
These experimental AI personalities proved to be more accurate than typical chatbots. Now, Sei wishes to explore how these findings can be applied in real-world settings. The goal is to understand how “digital personalities” may influence decision-making in group settings, particularly in creative fields.
As AI continues to evolve, it will increasingly engage in collaborative environments with humans and other AI agents. Sei believes that discussions enhanced by personality traits, including the ability to interrupt, could lead to better outcomes compared to strictly polite exchanges.
In a world where AI plays a growing role, understanding how it communicates could reshape how we interact with technology. For more on the potential of AI in communication and collaboration, check reports like the one from the National Institute of Standards and Technology.
