When prompts to draft work emails and job applications contain language commonly used by women, artificial intelligence chatbots including ChatGPT offer less sophisticated responses compared to office correspondence requested with language associated with men
Ok, we know that LLMs work basically as text completion and will adapt the language used in the response to the one used in the prompt. Their findings basically tell us that “language commonly used by women” is found in text content that is “less complex, at a lower grade level, and less formal”. It’s just statistics. To fix this “bias” they would have to artificially create loads of complex and formal text that uses “language commonly used by women” and use it to train LLMs. Not sure what the point of that would be though. If you want complex and formal response just write a complex and formal prompt.
Maybe it just means that women get treated with condescending answers more often especially in the Internet. That is my personal experience and not based on studies but there is a reason people want to defund gender studies and I do not like it.
Could be but the article is not talking about “more condescending” language but specifically about language that’s less formal and at a lower grade level. I’m not sure if condescending answers are automatically also “dumbed down”. I also think the way training text is gathered and fed into LLMs would separate different authors and context or we would see similar issues with all types of prompts. I can’t definitely say it’s not the reason but I think “language commonly used by women is more common in less formal texts” is a simpler and more convincing answer.
We do see these issues in a lot of ways because llms train on bias. More polite answers get more verbose and angry answers get less verbose. There are a lot of these “hacks” that exploit this.
Ok, we know that LLMs work basically as text completion and will adapt the language used in the response to the one used in the prompt. Their findings basically tell us that “language commonly used by women” is found in text content that is “less complex, at a lower grade level, and less formal”. It’s just statistics. To fix this “bias” they would have to artificially create loads of complex and formal text that uses “language commonly used by women” and use it to train LLMs. Not sure what the point of that would be though. If you want complex and formal response just write a complex and formal prompt.
Maybe it just means that women get treated with condescending answers more often especially in the Internet. That is my personal experience and not based on studies but there is a reason people want to defund gender studies and I do not like it.
Could be but the article is not talking about “more condescending” language but specifically about language that’s less formal and at a lower grade level. I’m not sure if condescending answers are automatically also “dumbed down”. I also think the way training text is gathered and fed into LLMs would separate different authors and context or we would see similar issues with all types of prompts. I can’t definitely say it’s not the reason but I think “language commonly used by women is more common in less formal texts” is a simpler and more convincing answer.
We do see these issues in a lot of ways because llms train on bias. More polite answers get more verbose and angry answers get less verbose. There are a lot of these “hacks” that exploit this.