A research team from the University of Cambridge and Google DeepMind has developed the first scientifically validated method for assessing and shaping the “personality” of large language models (LLMs), which underpin popular chatbots like ChatGPT. By utilizing psychological tests commonly used for assessing human personality, the researchers demonstrated that AI not only imitates human character traits but also can have its “personality” reliably measured and precisely shaped. The study found that larger models most accurately emulate human personality traits. These traits can be altered through prompts, influencing how AI performs specific tasks.
As an example of recent advances, newer LLM architectures are becoming increasingly efficient in learning from fewer data points, making them both greener and faster in deployment. This trend aligns with major tech companies’ focus on sustainable AI development. Furthermore, integrating techniques like reinforcement learning from human feedback has improved the nuance with which AI understands and mimics human sentiment.
The authors of the study warn that shaping a model’s personality can make chatbots more persuasive, raising concerns about manipulation. They call for urgent regulation of AI systems to ensure transparency and prevent abuse. Recent international AI summits have echoed this sentiment, highlighting the importance of ethical AI development. In late 2025, the European Union proposed a comprehensive AI Act aimed at addressing such challenges, with implementation expected by 2027, while the U.S. is advocating for a similar approach through broader policy discussions.
Illustration: Grok Scientists propose using their developed dataset and code for AI “personality” testing, which are publicly available, for auditing and testing advanced models before their release. To develop the method for assessing and verifying AI chatbot personality, the researchers tested how the behavior of various models in real-world tasks and verification tests statistically correlated with their scores on the “big five” traits used in psychometric testing: openness, conscientiousness, extraversion, agreeableness, and neuroticism. The team adapted two well-known personality tests – a 300-question version of the Revised NEO Personality Inventory and the shorter Big Five Inventory – and applied them to various LLMs using structured prompts.
Researchers found that larger, trained models produce personality test profiles that are both reliable and predictive of behavior, while smaller models yield inconsistent results. The exploration of personality adaptability is a burgeoning field within AI, mirroring developments in autonomous systems that can adjust settings dynamically based on user interactions. Scientists could guide the personality of a model over nine levels for each trait using carefully crafted prompts. For instance, they could make a chatbot more extroverted or more emotionally unstable – and these changes manifested in real-world tasks, such as writing social media posts. Such developments inspire new forms of user interaction but necessitate robust frameworks for ethical use, especially as AI systems become more integrated into everyday life.
