“Turning right”? An experimental study on the political value shift in large language models - Humanities and Social Sciences Communications
www.nature.com
Constructing artificial intelligence that aligns with human values is a crucial challenge, with political values playing a distinctive role among various human value systems. In this study, we adapted the Political Compass Test and combined it with rigorous bootstrapping techniques to create a standardized method for testing political values in AI. This approach was applied to multiple versions of ChatGPT, utilizing a dataset of over 3000 tests to ensure robustness. Our findings reveal that while newer versions of ChatGPT consistently maintain values within the libertarian-left quadrant, there is a statistically significant rightward shift in political values over time, a phenomenon we term a ‘value shift’ in large language models. This shift is particularly noteworthy given the widespread use of LLMs and their potential influence on societal values. Importantly, our study controlled for factors such as user interaction and language, and the observed shifts were not directly linked to changes in training datasets. While this research provides valuable insights into the dynamic nature of value alignment in AI, it also underscores limitations, including the challenge of isolating all external variables that may contribute to these shifts. These findings suggest a need for continuous monitoring of AI systems to ensure ethical value alignment, particularly as they increasingly integrate into human decision-making and knowledge systems.

cross-posted from: https://lemmy.ml/post/25907293

Constructing artificial intelligence that aligns with human values is a crucial challenge, with political values playing a distinctive role among various human value systems. In this study, we adapted the Political Compass Test and combined it with rigorous bootstrapping techniques to create a standardized method for testing political values in AI. This approach was applied to multiple versions of ChatGPT, utilizing a dataset of over 3000 tests to ensure robustness. Our findings reveal that while newer versions of ChatGPT consistently maintain values within the libertarian-left quadrant, there is a statistically significant rightward shift in political values over time, a phenomenon we term a ‘value shift’ in large language models. This shift is particularly noteworthy given the widespread use of LLMs and their potential influence on societal values. Importantly, our study controlled for factors such as user interaction and language, and the observed shifts were not directly linked to changes in training datasets. While this research provides valuable insights into the dynamic nature of value alignment in AI, it also underscores limitations, including the challenge of isolating all external variables that may contribute to these shifts. These findings suggest a need for continuous monitoring of AI systems to ensure ethical value alignment, particularly as they increasingly integrate into human decision-making and knowledge systems.