Stanford University researchers have uncovered a significant flaw in the alignment process of AI chatbots, revealing an unintended Western bias that compromises their effectiveness for global users.
A study to be presented at the upcoming Association of Computational Linguistics in Bangkok found that current alignment processes for large language models (LLMs) inadvertently favour Western-centric tastes and values. This bias can lead to misinterpretation of user queries and produce suboptimal results for non-Western users.
Diyi Yang, professor of computer science at Stanford and senior author of the study, emphasised the importance of considering whose preferences are being prioritised in LLM alignment. "The real question of alignment is whose preferences are we aligning LLMs with and, perhaps more importantly, who are we missing in that alignment?" Yang said.
The research team examined the effects of alignment on global users across nine languages, regional English dialects in the United States, India, and Nigeria, and value changes in seven countries.
One example highlighted the disparity in how Nigerian and American English speakers describe "chicken," with Nigerians relating it to a Christmas dish and Americans viewing it as fast food. Such cultural nuances can lead to misunderstandings and biased responses from AI chatbots.
The study also explored how alignment impacts LLMs' responses to moral questions that vary across cultures, such as the acceptability of divorce. This revelation underscores the challenge of creating truly globally inclusive AI systems.
Ryan, the first author of the paper, pointed out that the team's interest in this issue was sparked by observing performance gaps between American English and other English variants like Indian and Nigerian English.
As AI continues to play an increasingly significant role in global communication, addressing these biases becomes crucial. The researchers are now investigating potential root causes of these biases and exploring ways to improve the alignment process.
The findings of this study highlight the ongoing challenge of creating AI systems that can effectively communicate across diverse cultural and linguistic landscapes. As development of AI chatbots continues, ensuring their global effectiveness and cultural sensitivity remains a key priority for researchers and developers alike.