|
Western cultural bias built into artificial intelligence systems could influence how billions of users understand and interact with the world, as AI tools increasingly become part of areas such as education, professional counselling and legal guidance, according to a report by Chromatics AI.
The report said many AI systems are trained on data that does not adequately represent the languages, worldviews and ways of communication of large parts of the global population. As a result, AI responses can appear neutral while still reflecting cultural values associated with the West. "Research suggests that leading AI systems exhibit a default prioritisation of individualism and Anglo-Saxon norms," the report said. It added that these embedded values could become more significant as AI is used for functions that directly influence people's decisions and understanding. "And, as they are increasingly used for functions like legal guidance, professional counselling, and education, those embedded values risk shaping how under-represented users perceive the world around them in ways that may be directly at odds with their own," the report said. The report argued that simply translating AI systems or adapting them for different regions may not be enough to address the problem. It said localisation generally focuses on surface-level changes such as translating an interface, changing currency symbols or applying regional style guides, while the underlying AI model may continue to operate using the cultural assumptions present in its original training data. According to the report, this creates a gap between language localisation and genuine cultural understanding. It said users from communities that have historically been under-represented in technology can notice when an AI system misunderstands their context, communication style or values. The report pointed to examples of efforts to build AI systems around local languages and cultural contexts, including Egypt's Horus, Nigeria's N-ATLaS, Indonesia's Sahabat-AI, Latin America's Latam-GPT, Thailand's OpenThaiGPT and AI Sweden's Swedish-language model. It also cited Amazon's work on Alexa+ in Mexico as an example of the complexity involved in adapting AI to local culture. The report said the company had to use reinforcement learning and custom training data to adjust the system for the language and country while ensuring that improvements did not negatively affect other languages. Larry Adams, Founder at Chromatics AI, said, "Bespoke culture AI requires things the mainstream market has not yet prioritized at scale. Training data must be collected in genuine partnership with the communities it is meant to serve. Evaluation benchmarks must be culturally grounded, not just linguistically accurate. Communities themselves should help define what 'correct' looks like." The report said the next phase of AI development may therefore depend not only on larger models, but also on how effectively AI systems understand the cultural contexts of the people using them. (ANI)
|
