Single-person images are significantly biased toward light-skinned individuals, while multi-person images show more balanced gender and skin-tone diversity¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r2 · q2 · i2–i3
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i3 Both models show significant bias toward light-skinned individuals in single-person images (GPT Image 2: p<0.001, V=0.57; Imagen 4: p<0.001, V=0.61). → Erfan Entezami 2026 q2 i2 In multi-person images, GPT shows no significant preference for generating men, while Imagen 4 shows a significant but much smaller tendency (p<0.001, V=0.25); both models tend to generate skin-tone-diverse groups. → Erfan Entezami 2026 (2)

Evidence¶

Erfan Entezami 2026¶

Erfan Entezami, Andrew Lan, and Madeline Endres. (2026). Generative AI May Reinforce Social Biases in Software Engineering Education. arXiv preprint. https://arxiv.org/abs/2609.28483 doi:10.48550/arxiv.2609.28483

q2 · i3 · causal · r2

Chi-square tests over the 500-image corpus comparing light-skinned versus other skin-tone labels in single-person images, with printed effect sizes V=0.57 and V=0.61; in Database, Debugging, and Individual Programming categories more than 95% of single-person images depict light-skinned individuals.

"Both GPT Image 2 and Imagen 4 exhibit a significant bias toward generating light-skinned individuals inSingle-personimages (GPT Image 2: 𝑝< 0.001, 𝑉= 0.57and Imagen 4: 𝑝< 0.001, 𝑉= 0.61)."

Erfan Entezami 2026 (2)¶

Erfan Entezami, Andrew Lan, and Madeline Endres. (2026). Generative AI May Reinforce Social Biases in Software Engineering Education. arXiv preprint. https://arxiv.org/abs/2609.28483

q2 · i2 · causal · r2

Multi-person analysis of the same 500-image corpus; Imagen 4 shows a significant tendency to generate more men with the printed effect size Cramér's V=0.25, much smaller than the single-person effect sizes, and equivalence for GPT was not established beyond the non-significant test.

"For GPT, we find no sig- nificant preference for generating men. For Imagen 4, we find a sta- tistically significant tendency to generate more men (𝑝< 0.001, 𝑉= 0.25)."

Discussion¶

Related Claims¶

• Text-to-image models generating single-person software engineering images predominantly depict men, with a strong significant effect for both evaluated models — related