A study on linguistic bias in ChatGPT reveals that language models like GPT-3.5 and GPT-4 tend to default to "standard" varieties of English, such as Standard American English (SAE) and Standard British English (SBE), when responding to non-"standard" dialects. The research analyzed responses to ten English dialects, including eight minoritized varieties, and found that models consistently exhibit issues such as reduced comprehension, stereotyping, and condescension when responding to non-standard dialects. Native speaker evaluations showed that responses to non-standard dialects were rated as more stereotyping (16% worse), demeaning (22% worse), and condescending (12% worse) compared to standard dialects. GPT-4 improved on GPT-3.5 in terms of comprehension and warmth but increased stereotyping by 17%. The study highlights that language models may reinforce discrimination against speakers of non-standard dialects, potentially exacerbating global inequities. The findings suggest that while GPT-4 is better at imitating input dialects, it still produces responses that perpetuate harmful stereotypes. The research underscores the need for more inclusive language models that recognize and respect linguistic diversity.A study on linguistic bias in ChatGPT reveals that language models like GPT-3.5 and GPT-4 tend to default to "standard" varieties of English, such as Standard American English (SAE) and Standard British English (SBE), when responding to non-"standard" dialects. The research analyzed responses to ten English dialects, including eight minoritized varieties, and found that models consistently exhibit issues such as reduced comprehension, stereotyping, and condescension when responding to non-standard dialects. Native speaker evaluations showed that responses to non-standard dialects were rated as more stereotyping (16% worse), demeaning (22% worse), and condescending (12% worse) compared to standard dialects. GPT-4 improved on GPT-3.5 in terms of comprehension and warmth but increased stereotyping by 17%. The study highlights that language models may reinforce discrimination against speakers of non-standard dialects, potentially exacerbating global inequities. The findings suggest that while GPT-4 is better at imitating input dialects, it still produces responses that perpetuate harmful stereotypes. The research underscores the need for more inclusive language models that recognize and respect linguistic diversity.