Professor Jack Grieve notes that increasing the diversity of training data is more crucial than merely expanding the volume of data.

The study suggests that fine-tuning LLMs with datasets that reflect a broader sociolinguistic diversity can significantly enhance their societal value and reduce biases.

The research team advocates for a balanced training approach that incorporates data from various social groups and contexts to improve the performance of AI systems.

Understanding societal structures reflected in language is critical for maximizing the benefits of LLMs in society.

Lead author Professor Grieve highlighted that generative AIs often produce negative portrayals of certain ethnicities and genders due to biases in their training data.

These harmful biases in AI outputs largely stem from the biased language databases used to train these models.

A new framework, combining insights from the humanities and social sciences, aims to prevent AI tools from spreading misinformation and discriminatory content.

Past incidents, such as the UK Home Office halting a computer algorithm for visa applications due to allegations of racism and bias, underscore the urgency of addressing these issues in AI development.

Researchers from the University of Birmingham have developed a new framework that integrates sociolinguistic principles to enhance the understanding and performance of large language models (LLMs) like ChatGPT.

Generative AI systems, including ChatGPT, are grappling with significant issues related to misinformation and social biases, particularly those that perpetuate racist and sexist stereotypes.

To combat these biases, researchers emphasize the necessity for principled training of AI systems, advocating for the incorporation of sociolinguistic principles.

The study, published in Frontiers in AI, argues that accurately representing diverse language varieties can significantly enhance AI performance and mitigate critical challenges, including social bias and misinformation.