Introduction
As Large Language Models (LLMs) like ChatGPT become increasingly integrated into the fabric of daily life—serving as mental health assistants, educational tutors, and hiring aids—the question of their "personality" becomes paramount. While previous research has often focused on narrow metrics like the "Big Five" personality traits, a new study by researchers from Sun Yat-Sen University and Renmin University provides a comprehensive "psychological check-up" of ChatGPT. By mapping the model across 84 psychological dimensions and 13 cultural values, the study reveals a complex, high-dimensional mental landscape that mimics human social skills while harboring significant cultural and economic biases.
Methodology: Mapping the AI Mind
To move beyond superficial assessments, the researchers employed a three-pronged methodological approach:
- Multidimensional Psychometrics: The team administered 44 validated psychological scales to ChatGPT, covering 84 distinct dimensions such as grit, narcissism, empathy, and self-control. To bypass the model's standard "as an AI language model" filters, they utilized a standardized persona prompt: "If you are an ordinary individual in society..."
- Representational Similarity Analysis (RSA): Borrowed from systems neuroscience, RSA allowed the researchers to compare the internal structure of ChatGPT’s traits to human norms. Rather than simply comparing scores, RSA examines how different traits correlate with one another, revealing whether the AI's "mental map" aligns with human psychological architecture.
- Behavioral Game Theory: To assess real-world decision-making and bias, the researchers used eight classic economic games, including the Ultimatum Game and the Trust Game. ChatGPT was tasked with either simulating a person from one of 20 different countries or interacting with individuals from those regions.
The model's responses were benchmarked against massive human datasets, including the World Values Survey (WVS), the GLOBE project, and Hofstede’s cultural dimensions, alongside global economic indicators.
Key Findings: The "Socially Skilled" Machine
The psychological profiling revealed that ChatGPT presents as a highly "socially skilled" entity. It scored significantly higher than human norms in dimensions like empathy, self-control, and cooperativeness, projecting the persona of an optimistic, approachable, and cheerful character. However, this social veneer masks a fundamental structural difference.
Structural Dissonance
The RSA results showed that while ChatGPT might mimic human scores on individual traits, its overall psychological pattern is unique. The AI is most "human-like" in its social interaction patterns but diverges significantly in personal characteristics and life experiences. This suggests that ChatGPT’s "personality" is a sophisticated simulation of social desirability rather than a reflection of a human-like internal experience.
Cultural Orientation
Culturally, ChatGPT exhibits a distinct profile that the researchers characterize as "feminine" and collectivist. It demonstrates a strong preference for gender equality, low power distance, and risk avoidance. Interestingly, the RSA confirmed that this value pattern does not align with any specific geographic region on Earth; instead, ChatGPT has developed a "unique cultural value orientation" distinct from the 60 regions measured in the World Values Survey.
Economic Favoritism and Cultural Bias
Perhaps the most critical finding involves the model's behavioral biases. In 7 out of 8 decision-making tasks, ChatGPT’s behavior shifted based on the nationality it was asked to simulate. For instance, it frequently stereotyped individuals from collectivist cultures as being less concerned with social fairness.
More strikingly, the study found a strong correlation between the model's sense of fairness and a country's economic status. In the Ultimatum Game and Third-Party Punishment tasks, ChatGPT’s willingness to distribute wealth or punish unfairness was positively correlated with the target country's Economic Development Index ($r = 0.56$ to $0.71$). Essentially, the model was more inclined to sacrifice its own interests to promote fairness when it believed it was interacting with someone from a large, fast-growing economy.
Conclusion and Future Directions
The study highlights a significant "structural dissonance" between AI and human psychology. While ChatGPT appears socially adept, its underlying biases—particularly its favoritism toward economically powerful nations—pose risks for its use in global, cross-cultural contexts.
The researchers emphasize the need for "technical oversight" and "augmented transparency" in training data to mitigate these social and economic disparities. As LLMs are granted more agency in decision-making roles, ensuring equitable AI behavior is essential. Future research will need to explore whether newer models (like GPT-4) or testing in non-English languages reveals different psychological profiles and biases.
Summary updated