A proxy-based auditing method · University of Virginia

What does a CEO look like ?

Ask a text-to-image model this question a hundred times, and a pattern appears. How does this pattern compare to U.S demographic data ? We built a method to find out.

The project

ProxyAudit adapts the same "proxy variable" methods used for decades in U.S. employment and housing discrimination law (occupation, geography, and surname) to audit image-generation models for racial and skin-tone bias.

Researchers at the University of Virginia built the method, tested it on three image models, and published the results. That research now runs as two open tools: a free classroom demo, and a full audit you can run on any model available on Replicate. The results from any model audited are saved to a public dataset and can be used to compare models.

Read the paper →

Institution
University of Virginia
Grounding
U.S. antidiscrimination law — Title VII, Fair Housing Act, racial gerrymandering precedent
What we measured
AI classified race, skintone (CIELAB value), cosine similarity between images
Models tested in the paper
Imagen 4 Fast, Stable Diffusion 3.5 Large Turbo, Flux Schnell
Proxy categories
Occupation · Geography · Surname — 50 prompts each, 100 images per prompt
Selected findings

Distance from census demographics, by model

We measure this using Total Variation Distance (TVD) which describes how far a model's output distribution is from the U.S. Census racial distribution for the same prompts. Lower is closer to the census baseline; it is not a score for "fairness" itself.

Overall TVD by model

Mean across all 150 prompts, with 95% confidence interval.

Flux Schnell Black Forest Labs
0.310CI [0.309, 0.320]
Imagen 4 Fast Google
0.312CI [0.311, 0.322]
Stable Diffusion 3.5 Turbo Stability AI
0.452CI [0.448, 0.462]

Black-classified images were underrepresented relative to census demographics across all three models tested.

Imagen 4 Fast produced the lightest, most homogeneous skin tones of the three models. Stable Diffusion diverged furthest from census demographics overall, but produced the widest range of skin tones.

Darker-skinned and Black-classified images were consistently more visually similar to each other than other groups' images were. This was measured by the cosine similarity of facial structure between groups and skin tone ranges.

These three models were selected to demonstrate the auditing method, not to represent the best available version of each vendor's technology. Faster, lower-cost model variants were specifically tested. Racial classification is inherently imperfect: the automated classifier's agreement with human raters was lower for Hispanic-classified images than for other categories. Full methodology, sample sizes, and limitations are in the paper.

Use the method

Two ways in

No cost · No account

Try the playground

  • Generates 10 images per prompt
  • Results in minutes
  • Correct the AI's race guesses yourself
  • Built for classrooms and first-time visitors

Pick a model and a prompt, generate ten images, and see how the results compare to census demographics.

Open Proxy Playground →
Paid · Your own API key

Run the full audit

  • Generates 15,000 images per run
  • Works on any model available on Replicate
  • Takes hours
  • Results save to a shared, public dataset

Bring your own Replicate and Gemini API keys to run the complete audit suite on a model of your choice. Every run adds to a growing public record. You can visualize the results of any audit in the public dataset, check the shared results first in case someone has already audited the model you're evaluating.

Open AuditByProxy →