Wei Zhou, Janet Watson, Serge Sharoff, Preserving Linguistic Heritage: ChatGPT cannot resist the temptation of speaking 'proper' language

Abstract: Modern language technologies, such as GPT, have demonstrated their usefulness in a number of areas, yet they exhibit biases and limitations. This study explores one of the biases concerning the focus language models on greater standardisation and normalisation of their output, thus ignoring the dialectal varieties even when the prompts specifically request using them. We investigate three kinds of prompts, namely, recipes, storytelling and informative exchanges, for a specific Chinese language variety, Sichuanese, with its own lexical differences from the standard Putonghua Chinese. Our findings show that while ChatGPT has been trained to be apologetic to agree with any suggestion coming from the user, it keeps reverting to the standard variety with its associated biases and prejudices. This raises a question of the efforts needed to ensure the survival of smaller languages and dialects with the wider use of language technologies.