Guiding Questions

1. Knowing what ChatGPT is trained on (search engine crawl, ebooks, reddit, and wikipedia), what kinds of cultural concepts or groups might not be included?
   • What about oral languages, since less than 10% of human languages are written?
   • What about non-standard inscription media like the Benin bronzes, Incan quipu, or Maori carvings?
   • Is a translation ever an accurate representation of the original?
2. What languages do you speak and what have you noticed about moving back and forth from those languages?
   • “How To Speak Bad English” (8:20-13:40) podcast episode on Global English and accent reduction. A major point here is that more English speakers are “nonnative” than “native” so “native” speakers need to adjust their expectations on what “clear communication” is.
   • Visualizing the Most Used Languages on the Internet

Discussion Questions:

For the group discussing “ChatGPT threatens language diversity”:

1. How does the AI respond to prompts in non-English languages?
2. Does the AI show any bias towards English language or syntax when generating responses?
3. Try typing a sentence with non-English syntax in English. How does the AI respond?

For the group discussing “ChatGPT is multilingual but monocultural”:

1. Generate a story set in a non-Western culture. Does the AI accurately and respectfully incorporate elements of that culture?
2. How does the AI respond to prompts containing cultural idioms, references, or concepts?
3. Look up some common phrases or idioms in less commonly used languages. How does the AI respond to these prompts?

For the group discussing “Proper English and normative grading practices”:

1. Try typing sentences in various English dialects or accents (e.g., African American Vernacular English, Singlish, Hinglish). How does the AI respond?
2. Does the AI seem to favor a particular type of English in its responses?
3. How does ChatGPT’s answer to your question change as you rephrase the same question (ie using “Black” rather than “African American”) Does it perpetuate stereotypes or exhibit biases?

Links to Useful articles on this

Summary points

1. 1. “Unmasking AI Harms and Biases”.
   2. “ChatGPT threatens language diversity in the age of AI”
      1. With white male voices authoring the majority of the training material, the default voice replicates those language patterns.
   3. “ChatGPT is multilingual but monocultural, and it’s learning your values”
      1. Diversity is not just in languages and dialects used but in the cultural beliefs and ideologies embedded in the training material.
   4. “ChatGPT & Writing in the Secondary ELA Classroom”
      1. Our normative grading practices around “proper English” encourage students to mask language diversity.
   5. “OpenAI’s Linguistic Diversity Initiatives in AI Language Testing”
      1. OpenAI’s proposed solutions focus more on including less common languages but say much less about how to addressing race and gender stereotypes in language use.