AI systems and the reproduction of (standard) language ideologies in World Englishes

Imagine you're developing a language model that interacts with users from diverse English-speaking backgrounds. The challenge lies in ensuring that the model respects and accurately represents the various forms of English, rather than defaulting to a narrow, dominant standard. Currently, many AI systems are trained on data that favors Inner Circle English norms, which can lead to misunderstandings and marginalization of non-dominant Englishes. This is what's called the 'standardization paradox' — while AI can homogenize language by prioritizing standard forms, it also has the potential to expose users to a broader range of English varieties through diverse training data. However, the fixation on certain words or phrases, like 'delve,' often reflects a policing of language norms that can alienate users from the Global South. The paper argues for a more inclusive design approach that acknowledges the plurality of Englishes, aiming to mitigate the negative consequences of treating some forms of English as more legitimate than others. This shift in perspective is crucial for builders who want to create AI systems that are not only effective but also culturally sensitive and representative of the global English-speaking community.

The paper addresses the intersection of AI and sociolinguistics, highlighting the implications of language ideologies in AI outputs.

The analysis is grounded in empirical studies and real-world examples, though it lacks extensive quantitative metrics.