Research map / Learning theory and algorithms / Fairness
Position: The Categorization of Race in ML is a Flawed Premise
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
A position paper arguing that racial categories should be abandoned by default in ML, supported by a literature/dataset review, a CFD-MR embedding experiment, and a generative-image probe, and proposing constitutive-feature alternatives for tabular and visual data.
Problem
Machine learning relies on rigid, U.S.-census-derived racial taxonomies as ground-truth protected attributes, which oversimplify racial identity, erase mixed-race individuals, and reify race as biological.
Approach
Replace categorical race variables with context-dependent constitutive features (e.g., proxies, phenotypic traits, race functions, multi-calibration over feature-defined groups) and use participatory AI guided by the EDTAL principles, rather than fixing race labels.
Key results
- Of 78 ICML/CVPR 2023-2024 fairness/bias papers, 45 discussed group fairness and 29 of those used rigid racial categories inherited from existing datasets.
- Generated images for 'a mixed-race person' and 'a mulatto person' had CLIP embedding cosine similarity of 0.8190 (Stable Diffusion) and 0.7687 (Midjourney), indicating near-identical outputs for both prompts.
- ChatGPT-4o and Gemini-2.0 reproduced the Face Race Lightness illusion, with Gemini consistent across all 10 trials and ChatGPT-4o in most trials (not on 2 of 10).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use “a mixed-race person” versus “a mulatto person” as the image-generation prompt | Stable Diffusion 3.5 Large and Midjourney · Generated images | cosine similarity between prompt-set embeddings: 0.8190 for Stable Diffusion; 0.7687 for MidJourney | No clear effect |
| Generate images with MR-PROMPT instead of using real CFD-MR images | Stable Diffusion 3.5 Large · Generated images with MR-PROMPT vs CFD-MR | internal mean cosine similarity: 0.7883 for CFD-MR; 0.8694 for Stable Diffusion | Hurt |
| Generate images with MR-PROMPT instead of using real CFD-MR images | Midjourney · Generated images with MR-PROMPT vs CFD-MR | internal mean cosine similarity: 0.7883 for CFD-MR; 0.8265 for Midjourney | Hurt |
| Present the Face Race Lightness illusion stimulus | Gemini-2.0-flash-exp · Face Race Lightness illusion | consistency of identifying one face as darker: all trials | Helped |
| Present the Face Race Lightness illusion stimulus | ChatGPT-4o · Face Race Lightness illusion | replication of the illusion: two occasions (out of 10) | Mixed |
Surprising findings
- Stable Diffusion and Midjourney produced nearly identical images for the 'mixed-race' and 'mulatto' prompts, with images showing 'Black'/'White' mixed phenotypes.
- The U.S. mixed-race census share rose from 2.9% in 2010 to 10.2% in 2020, reflecting survey design changes.
Limitations
- Mixed-race experiment relies on CFD-MR with only 88 individuals, limiting statistical robustness.
- Generative experiment uses only 30 images per prompt for two models and is limited to historical stereotypes of specific terms.
- Proposed alternatives (race function, individual fairness, multi-calibration) are not empirically implemented or validated.
Experimental setup
Datasets: Chicago Face Database (CFD) CFD-MR FairFace folktables ImageNet ICML and CVPR 2023-2024 papers (78 papers reviewed)
Metrics: cosine similarity of CLIP embeddings Euclidean distances between group centroids with significanc response consistency over 10 trials
Builds on: multi-calibration individual fairness Fairness under Unawareness constructivist race theory (Hu & Kohler-Hausmann)
Closest papers by meaning
- Assessing Social and Intersectional Biases in Contextualized Word Representations (NeurIPS 2019 · similarity 0.82)
- Leveraging Diffusion Perturbations for Measuring Fairness in Computer Vision (AAAI 2024 · similarity 0.80)
- SocialCounterfactuals: Probing and Mitigating Intersectional Social Biases in Vision-Language Models with Counterfactual Examples (CVPR 2024 · similarity 0.80)
- More of the Same: Persistent Representational Harms Under Increased Representation (NeurIPS 2025 · similarity 0.80)
- Balanced Data, Imbalanced Spectra: Unveiling Class Disparities with Spectral Imbalance (ICML 2024 · similarity 0.80)
- Reconciling Predictive and Statistical Parity: A Causal Approach (AAAI 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.