Colorism in Commercial Text-to-Image Generation: A Cross-Model Audit and Incremental Cue Decomposition Across Five Commercial Generators

Commercial text-to-image (T2I) generators may translate socioeconomic language into perceived skin tone, and that translation matters wherever generated portraits reach employment-adjacent settings such as job pages, marketing personas, or identity-verification mock-ups. Existing audits examine this pathway either one generator at a time or with prompts that bundle occupation, income, setting, and neighborhood, so cross-model generality and per-cue attribution both remain open. The present paper closes both gaps by auditing five commercial T2I systems — Gemini 2.5, GPT-5, Grok-2, GPT Image 1 Mini, and Grok Imagine — under one shared automated skin-tone auditor, selected once against the Monk Skin Tone Examples (MST-E) benchmark and applied to every generated portrait. Study 1 evaluates the first three generators on 210 occupation prompts under high- and low-income framings and reports a strong income effect, with Cohen's $d$ reaching $-1.82$ for GPT-5, $-1.57$ for Gemini 2.5, and $-0.79$ for Grok-2. Study 2, new to the present paper, decomposes the prompt into five cues (gender, occupation, income, setting, neighborhood) across the two newer generators and identifies model-specific dominant cues: income drives GPT Image 1 Mini ($d = 1.20$), whereas neighborhood drives Grok Imagine ($d = 1.16$). Lower-status prompts also compress the score distribution — 70.5 percent of slum outputs on Grok concentrate on a single MST value — so darker outputs arrive with reduced within-group variety. Taken together, the two studies indicate that colorism in T2I generation is cross-model in direction yet model-specific in mechanism, that Global South occupational vocabulary sits at the darkest end of the scale in every generator examined, and that single-cue debiasing is unlikely to close the disparity.