Large Language Models Perpetuate Masculine Gender Bias Despite Neutral Instructions, Study Finds

A recent arXiv study examining gender bias in large language models shows that popular LLMs—both open-source and commercial—systematically perpetuate masculine generic language when given ostensibly neutral instructions, reinforcing a linguistic pattern that psychology research has long shown is not truly gender-neutral.

What the Research Found#

Researchers created a dataset of over 16,000 human nouns and tested six different LLMs on four instruction-response datasets in French under two conditions: prompts explicitly containing masculine generics and those without. The results were striking:

• When prompted with masculine generic language, roughly 78.55% of LLM responses exhibited masculine generic bias
• Even when given prompts without masculine generics (theoretically neutral framing), approximately 27.57% of responses still defaulted to masculine generics
• Models rarely spontaneously adopted gender-fair language alternatives

Why Masculine Generics Matter#

Masculine generics—using masculine grammatical gender to ostensibly refer to mixed-gender groups or people of unknown gender—are common in many languages. Although presented as neutral, psycholinguistic research consistently demonstrates that these constructions unconsciously trigger masculine-centered mental imagery and understanding. The use of masculine generics thus encodes gender bias into language in ways that subtly shape perception and thought.

The Broader Implication#

The findings underscore how large language models, despite being trained on massive datasets and fine-tuned with instructions, inherit and amplify existing linguistic biases from their training data. Rather than moving toward more inclusive language practices, these systems continue to favor gendered defaults—and do so even when no gendered language prompts them to do so. This suggests that addressing gender bias in AI will require deliberate intervention beyond simply avoiding gendered prompts in instructions.