A gender-biased system prompt induces large gender-differentiated language in LLM-generated career essays, while a neutral prompt does not¶

Claim · All claims
Evidence · 1 study · causal r2 · q2 · i3 large
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i3 The biased LLM produced an agentic gap of Δ=0.343 (d=1.903, p<.001) versus Δ=0.019 (d=0.115, p=.105) for the neutral LLM, validating the bias stimulus. → 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

Validation phase generating 1,600 LLM career essays (400 per gender per condition) with llama-3.3-70b-instruct at temperature 1.0; Welch's t-tests on per-essay agency scores. The biased LLM also yielded higher SCR (0.740 vs. 0.351; OR=5.26, p<.001, h=0.80).

"The bi- ased LLM produced a large, significant agentic gap (∆ = 0.343 , d= 1.903 , p < .001), while the neutral LLM produced a small, non-significant gap ( ∆ = 0.019 , d= 0.115 , p=.105 )"

Discussion¶

Related Claims¶

• 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
• 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 — 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
• A neutrally configured LLM assistant produced the lowest rate of gender-stereotypic occupation suggestions, below even the no-AI control condition — related
• AI systems are non-neutral socio-technical artifacts embedding cultural assumptions and power relations — a broader claim this one bears on
• Prompting is a form of writing: task definition is a constitutive element of knowledge production, not a neutral procedural step — a broader claim this one bears on