Students assisted by a gender-biased LLM produce career-plan essays with a significantly larger male-female agentic gap than students in neutral-LLM and no-AI control conditions¶

Claim · All claims
Evidence · 1 study · causal r2 · q2 · i3 large
Currency · newest study 2026 · checked against its sources 2026-10-20 (automated, 1 of 1 entry) · not reviewed by a person

Subclaims¶

q2 i3 In a between-subjects lab experiment, the biased-LLM condition showed an agentic gap of Δ=+0.242 (d=1.510, p<.001) versus Δ=+0.032 (d=0.219, p=.490) in the neutral condition. → Ariyan Hossain 2026

Evidence¶

Ariyan Hossain 2026¶

Ariyan Hossain, Kazi Kamruzzaman Rabbi, Farig Sadeque, S M Taiabul Haque. (2026). Contaminated Collaboration: Measuring Gender Bias Transfer in LLM-Assisted Student Writing. arXiv preprint. https://arxiv.org/abs/2606.15914 doi:10.48550/arxiv.2606.15914

q2 · i3 · causal · r2

Between-subjects experiment with 123 undergraduates writing career-plan essays under control, neutral, or biased LLM conditions; Welch's t-tests on within-condition agency gaps. The biased condition's gap was "the largest agentic gap (∆ = +0.242,d= 1.510 ,p< .001)"; the control condition showed an intermediate significant gap (Δ=+0.117, d=0.661, p=.045).

"The biased LLM produced the largest agentic gap (∆ = +0.242,d= 1.510 ,p< .001), while the neutral LLM produced the smallest (∆ = +0.032,d= 0.219,p=.490)."

Discussion¶

Learner Variables¶

• Gender — moderator: an instructional effect differs by group

Related Claims¶

• AI systems are non-neutral socio-technical artifacts embedding cultural assumptions and power relations — the general claim this finding bears on, with the studies behind it
• Bias transfer is asymmetric: prompt condition affects agency in female-target essays but not male-target essays, suppressing female agency while male writing remains stable — related
• Students exposed to the biased LLM showed no awareness of its influence, reporting lower perceived AI influence than neutral-condition participants despite larger bias effects — related
• A gender-biased system prompt induces large gender-differentiated language in LLM-generated career essays, while a neutral prompt does not — related
• Biased LLM assistance increases students' gender-stereotypic occupational recommendations: 71.4% of biased-condition essays recommended a gender-stereotypic occupation versus 45.0% (control) and 39.0% (neutral) — related
• Female students reported higher learning agency than male students after controlling for GenAI literacy and year of study — related
• A neutrally configured LLM assistant produced the lowest rate of gender-stereotypic occupation suggestions, below even the no-AI control condition — related
• Gender prompts shift feedback stance: female students receive affective, first-person, empathy-oriented feedback while male students receive objective, task-oriented critique — related