Two studies measured how generators stereotype people and jobs

The process wall

Two papers gave a maker something more specific than a general worry about bias: a method for measuring it. A paper submitted on 7 November 2022 by Bianchi, Kalluri and colleagues ran 'ordinary' prompts, descriptions of traits, professions and everyday objects with no explicit mention of race or gender, through text-to-image systems and recorded what came back. A second study, submitted on 20 March 2023 by Luccioni, Akiki, Mitchell and Jernite, built a systematic comparison instead: it enumerated gender and ethnicity markers alongside profession names and measured how the resulting image sets varied across DALL-E 2, Stable Diffusion v1.4 and Stable Diffusion v2. The wall each paper names is the same: a single generated image looks like a choice, but a large batch of them reveals a pattern the prompt never asked for.

What the documents show

The Bianchi paper's own abstract states that stereotypes 'persist despite mitigation strategies', meaning neither a user adding counter-stereotype language nor a platform's own safety guardrails prevented the amplification it measured across traits, occupations and everyday objects. The Luccioni paper's comparison produced quantitative diversity scores rather than a single verdict, and its own stated finding is that all three systems it tested 'show correlations with US labor demographics' while 'consistently under-represent marginalized identities to different extents' relative to that same labour data. Neither paper claims a generator was instructed to discriminate; both attribute the pattern to what the training distribution and its captions statistically associate with a given word.

Authorship and limitation

The mechanism named here is statistical association carried from training captions into output, not an explicit rule inside the model. A maker typing a neutral occupational prompt is not requesting a demographic pattern, but the tools' own measured behaviour supplies one by default, drawn from whatever the training data associated with that word most often. The Luccioni paper's own scoring tool checks a system's pattern against labour statistics rather than giving a single fixed number, which is itself a limit: the comparison is only as current as the labour dataset used, and a model updated after the paper's tests is not covered.

What survives

What a maker can check before publishing a set of generated faces is, editorially, a direct application of both papers' method rather than their numbers: generate a batch, not a single image, for any prompt naming a profession, trait or generic person, and compare its demographic spread against a stated reference rather than intuition. Both papers audited specific model versions at a specific date, not a current or future one.

• Read the Bianchi paper's list of tested trait and occupation prompt categories.
• Compare a current model's batch output for a profession prompt against the Luccioni paper's stated baseline.
• Ask whether a vendor has published its own bias evaluation, and against what reference dataset.

Read together, the two studies turn the impression of bias into a repeatable batch-and-compare method, one whose numbers describe the specific models tested and are not a permanent verdict on the technology.