It's How You Ask: Gender-Associated Linguistic Bias in LLMs
Summary
A study shows gender-associated linguistic bias in LLM responses: prompts featuring language patterns more common among women lead to shorter, less sophisticated replies across models, even after controlling prompt complexity. The bias appears linked to linguistic features encoded in early transformer layers, making post-hoc mitigation difficult and highlighting the need for upstream considerations in prompt design and model training to reduce disparate workplace communication impacts.