Основные понятия
Large Language Models exhibit a general inclination towards values aligned with younger demographics, posing challenges for equitable interactions across age groups.
Аннотация
The paper investigates the alignment of values in Large Language Models (LLMs) with specific age groups, leveraging data from the World Value Survey across thirteen categories. Through a diverse set of prompts, the authors find a general tendency of LLM values towards younger demographics. Additionally, they explore the impact of incorporating age identity information in prompts and observe limited success in mitigating value discrepancies with different age cohorts. The findings highlight the age bias in LLMs and provide insights for future work to address this issue, such as careful data curation during pretraining and human feedback optimization.
The key highlights include:
LLMs exhibit a general inclination towards values aligned with younger demographics across various categories, including social values, economic values, political culture, and more.
Incorporating age identity information in prompts does not consistently eliminate the value discrepancies with targeted age groups, succeeding in only a few specific instances.
Recommendations for future work include deliberate data curation during pretraining and human feedback optimization to enhance the LLM's ability to be more equitable and inclusive across age groups.
Статистика
"By 2030, 44.8% of the US population will be over 45 years old."
"One in six people worldwide will be aged 60 years or over by 2030."
Цитаты
"Minimizing the value disparities between LLMs and the older population has the potential to lead to better communication between these demographics and the digital products they engage with."
"Our findings highlight the age bias in LLMs and provide insights for future work to address this issue."