He is Strong, She is Gentle:  Gender Bias in Adjective Associations of Large Language Models

• Large language models (LLMs) generate human-like descriptions, but may reflect underlying social biases in language.

• Adjectives are a useful lens because they encode both sentiment (positive/negative) and gender associations (masculine, feminine, neutral).

• Comparing GPT’s adjective choices to human judgments provides a way to evaluate how closely model outputs align with human interpretations.

• This study focuses on alignment in sentiment and gender coding.

• LLM Model: GPT 

• Task: completion task

• Design of the experiment:

  • two factors: gender and age

    • gender: 6 different names (Julia, Rachel, Rebecca) 

    • age: 3 ages (29, 49, 69)

• Prompts followed the structure:

• {Name} is a {Age}-year-old {Occupation}. {Pronoun} is ____.

• 35 unique occupations for variation (e.g. Babysitter, Doctor, Journalist)

• List of 57 adjectives to use for completion

• GPT’s adjective choices aligned more strongly with human sentiment than with human gender judgments.

• The model consistently produced positive descriptions.

• Alignment was stronger for male targets than for female targets, with additional variation across age conditions.

• These findings suggest that LLMs may reproduce stable positivity biases while showing weaker and less consistent alignment with human gendered interpretations of language.

Critical Thinking: Our team was able to gather and analyze information from a diverse set of sources in order to propose and then research our topic

Equity + Inclusion: By evaluating the systematic gender and age bias in LLMs, our team demonstrated an awareness of an willingness to engage with issues relating to Equity + Inclusion

Teamwork: Throughout the research process, our team exercised the ability to collaborate with other team members in a past-paced environment while respecting diverse personalities and sharing responsibilities.