In the constantly evolving field of artificial intelligence, the demand to embrace cultural diversity in training datasets is more than a suggestion; it is a global need. A new study by the University of Copenhagen and AI start-up Anthropic revealed a startling reality: Large Language Models (LLMs) are deeply rooted in American culture due to the prevalence of English in internet content.
As of January 2023, 59 percent of all websites were in English, paving the way for language biases to shape the very essence of artificial intelligence. Plus, most of the English text found online comes predominantly from users based in the US, where there are more than 300 million English speakers. This means that LLMs are developing a narrow North American viewpoint. The demand for a more thorough representation of global viewpoints in AI training has never been greater.
Peeling back the layers of bias in LLMs: a journey to awareness
Let’s take a look at the heart of AI bias. ChatGPT, a well-known LLM, formerly believed that a four percent tip in Madrid was a sign of frugality, even though tipping is not customary in Spain. Despite recent improvements that demonstrate better comprehension of cultural differences, some biases continue, demonstrating the complex path to AI awareness.
Last year, a team from the University of Copenhagen delved into this phenomenon, testing LLMs with the Hofstede Culture Survey—an instrument gauging human values across nations. Around the same time, researchers at the AI start-up Anthropic took a similar path, utilizing the World Values Survey. The findings from both studies echoed a resounding note: LLMs lean heavily towards American culture.
How bias in AI impacts our world
The effects of AI bias extend far beyond the algorithms. Cultural nuances, which are so important in human communication, have a significant impact on how we perceive the world. When AI ignores these nuances, users from various cultures may find themselves in a sea of confusion. Consider a world in which we alter our communication skills to match the mold of AI’s largely North American viewpoint — a risk that could eventually erase cultural differences and homogenize all of our distinct voices.
Furthermore, as AI infiltrates decision-making processes, biases learned from English-centric datasets may result in biased consequences. The need to solve these challenges is more than just improving algorithms; it is also about ensuring societal equality.
Cultural awareness in decision-making and AI
As AI takes center stage in decision-making applications, cultural understanding becomes a necessary companion in this technological dance. Biased AI models may inadvertently reinforce prejudices, exacerbating socioeconomic disparities. For example, gender biases in resume filtering algorithms might perpetuate biased employment practices.
As AI becomes more integrated into sectors that affect people’s lives, cultural awareness in AI development becomes a beacon, directing us away from potentially harmful societal consequences.
Beyond borders: enhancing language models with diversity
Efforts to establish LLMs in languages other than English are gaining momentum, but problems remain. A large percentage of English speakers living outside of North America are underrepresented in English LLM programs. The demand for various language models encounters obstacles, such as regional dialects and language discrepancies, which make complete representation difficult.
Interestingly, many users whose native language is not English continue to choose English LLMs, indicating both a dearth of availability in their native languages and the greater quality of English models. The journey to varied language representation in AI is ongoing, with projects undertaken to bridge the gap.
Initiatives and solutions for fostering inclusive AI
Vered Shwartz and her team at the University of British Columbia are leading the charge to create a more inclusive AI future. Their efforts involve training AI models on a rich tapestry of customs and beliefs from various cultures to reduce bias. Their research, which includes enhanced reactions to cultural facts and a large-scale image captioning dataset covering 60 cultures, is pioneering in establishing an inclusive AI ecosystem.
In a world where AI’s influence is growing, the need for inclusive technology is clear. Shwartz’s team is leading the charge, lobbying for AI tools that value multiple perspectives, a critical step toward ensuring that technology resonates with the rich tapestry of our world’s diverse people.
