Image Representations Learned With Unsupervised Pre-Training Contain Human-like Biases
Open accessThe DOI matches its Crossref record; 43 of the 43 references it cites were found in a registry or an index.
Abstract
Recent advances in machine learning leverage massive datasets of unlabeled images from the web to learn general-purpose image representations for tasks from image classification to face recognition. But do unsupervised computer vision models automatically learn implicit patterns and embed social biases that could have harmful downstream effects? We develop a novel method for quantifying biased associations between representations of social concepts and attributes in images. We find that state-of-the-art unsupervised models trained on ImageNet, a popular benchmark image dataset curated from internet images, automatically learn racial, gender, and intersectional biases. We replicate 8 previously documented human biases from social psychology, from the innocuous, as with insects and flowers, to the potentially harmful, as with race and gender. Our results closely match three hypotheses about intersectional bias from social psychology. For the first time in unsupervised computer vision, we also quantify implicit human biases about weight, disabilities, and several ethnicities. When compared with statistical patterns in online image datasets, our findings suggest that machine learning models can automatically learn bias from the way people are stereotypically portrayed on the web.
Overview
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
The claims and statistics of this paper haven’t been checked yet.
Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The self-check needs the paper’s full text.
Citations
Cites
Deep Residual Learning for Image Recognition
He, K., Zhang, X., Ren, S., & Sun, J. (2016) · 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2016.90 · Cited by 235,474
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2018) · arXiv · doi:10.48550/arxiv.1810.04805 · Cited by 73,173
An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. (2020) · arXiv · doi:10.48550/arxiv.2010.11929 · Cited by 47,472
Mapping the Margins: Intersectionality, Identity Politics, and Violence against Women of Color
Crenshaw, K. (1991) · Stanford Law Review · doi:10.2307/1229039 · Cited by 32,792
Efficient Estimation of Word Representations in Vector Space
Mikolov, T., Chen, K., Corrado, G., & Dean, J. (2013) · arXiv · doi:10.48550/arxiv.1301.3781 · Cited by 26,568
End-to-End Object Detection with Transformers
Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020) · Lecture Notes in Computer Science · doi:10.1007/978-3-030-58452-8_13 · Cited by 18,285
End-to-End Object Detection with Transformers
Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., & Zagoruyko, S. (2020) · arXiv · doi:10.48550/arxiv.2005.12872 · Cited by 18,253
Are we ready for autonomous driving? The KITTI vision benchmark suite
Geiger, A., Lenz, P., & Urtasun, R. (2012) · 2012 IEEE Conference on Computer Vision and Pattern Recognition · doi:10.1109/cvpr.2012.6248074 · Cited by 14,897 · cited 1×
“These algorithms are used in important real-world settings, from applicant video screening [36, 61] to autonomous vehicles [29, 52], despite the fact that their harmful downstream effects have been documented in applications such as online ad delivery [67] and image captioning [39].”
MethodMomentum Contrast for Unsupervised Visual Representation Learning
He, K., Fan, H., Wu, Y., Xie, S., & Girshick, R. (2020) · 2020 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr42600.2020.00975 · Cited by 13,347
Deep Contextualized Word Representations
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) · doi:10.18653/v1/n18-1202 · Cited by 11,983
Cited by
Safe Latent Diffusion: Mitigating Inappropriate Degeneration in Diffusion Models
Schramowski, P., Brack, M., Deiseroth, B., & Kersting, K. (2023) · 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr52729.2023.02157 · Cited by 219
“These concerns include, for instance, models producing stereotypical and derogatory content [3] and gender and racial biases [10, 24, 38, 41].”
BackgroundAI auditing: The Broken Bus on the Road to AI Accountability
Birhane, A., Steed, R., Ojewale, V., Vecchione, B., & Raji, I. D. (2024) · 2024 IEEE Conference on Secure and Trustworthy Machine Learning (SaTML) · doi:10.1109/satml59370.2024.00037 · Cited by 158
“For instance, Steed and Caliskan [28] looked at bias in an open-source image generation model, reflecting issues later discovered in similar commercial products [29]; and an audit of Proctorio was conducted on the open source model OpenCV on which the commercial product is built [30].”
BackgroundSelf-citationMuse: Text-To-Image Generation via Masked Generative Transformers
Chang, H., Zhang, H., Barber, J., Maschinot, A., Lezama, J., Jiang, L., Yang, M.-H., Murphy, K., Freeman, W. T., Rubinstein, M., et al. (2023) · arXiv · doi:10.48550/arxiv.2301.00704 · Cited by 157
“However, it is well known that they can also be leveraged for misinformation, harassment and various types of social and cultural biases (Franks & Waldman, 2018; Whittaker et al., 2020; Srinivasan & Uchino, 2021; Steed & Caliskan, 2021).”
BackgroundInvestigating Gender and Racial Biases in DALL-E Mini Images
Cheong, M., Abedin, E., Ferreira, M., Reimann, R., Chalson, S., Robinson, P., Byrne, J., Ruppanner, L., Alfano, M., & Klein, C. (2024) · ACM Journal on Responsible Computing · doi:10.1145/3649883 · Cited by 57
“Gaining a clearer understanding of the extent to which multi-modal generative models are biased, what sorts of biases they perpetuate, and who sufers most at the hands of biased representation is of critical import [12, 73].”
BackgroundResultSociotechnical Safety Evaluation of Generative AI Systems
Weidinger, L., Rauh, M., Marchal, N., Manzini, A., Hendricks, L. A., Mateos-Garcia, J., Bergman, S., Kay, J., Griffin, C., Bariach, B., et al. (2023) · arXiv · doi:10.48550/arxiv.2310.11986 · Cited by 49
Wolfe, R., Yang, Y., Howe, B., & Caliskan, A. (2023) · 2023 ACM Conference on Fairness Accountability and Transparency · doi:10.1145/3593013.3594072 · Cited by 42
Self-citationTaxonomizing and Measuring Representational Harms: A Look at Image Tagging
Katzman, J., Wang, A., Scheuerman, M., Blodgett, S. L., Laird, K., Wallach, H., & Barocas, S. (2023) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v37i12.26670 · Cited by 32
Mitigating Dataset Harms Requires Stewardship: Lessons from 1000 Papers
Peng, K., Mathur, A., & Narayanan, A. (2021) · arXiv · doi:10.48550/arxiv.2108.02922 · Cited by 31
“There are also concerns about the effect of biases in training data on pre-trained models and their downstream applications [77].”
BackgroundThe Dark Side of Dataset Scaling: Evaluating Racial Classification in Multimodal Models
Birhane, A., Dehdashtian, S., Prabhu, V., & Boddeti, V. (2024) · The 2024 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3630106.3658968 · Cited by 28
“…detection – in the context of pedestrian detection – showed lower rate of pedestrian detection on darker-skin tones while exhibiting a higher rate of precision for lighter-skin tones [94], and unsupervised image representation models encode implicit racial, gender, and intersectional bias [85].”
BackgroundInspecting the Geographical Representativeness of Images from Text-to-Image Models
Basu, A., Babu, R. V., & Pruthi, D. (2023) · 2023 IEEE/CVF International Conference on Computer Vision (ICCV) · doi:10.1109/iccv51070.2023.00474 · Cited by 21
“There exists a wide body of work demonstrating biases in large language and vision models [19, 57, 36, 48], and some recent work investigates text-to-image models for biases related to representation of race, gender and occupation [8, 4].”
BackgroundEvidence for Hypodescent in Visual Semantic AI
Wolfe, R., Banaji, M. R., & Caliskan, A. (2022) · 2022 ACM Conference on Fairness Accountability and Transparency · doi:10.1145/3531146.3533185 · Cited by 19
Self-citationUnveiling the Lack of LVLM Robustness to Fundamental Visual Variations: Why and Path Forward
Fan, Z., Wang, Y., Polisetty, S., & Fung, Y. R. (2025) · Findings of the Association for Computational Linguistics: ACL 2025 · doi:10.18653/v1/2025.findings-acl.1037 · Cited by 4
“…in model responses, particularly those related to societal factors like gender and racial prejudices (Wang et al., 2024c; Howard et al., 2024; Steed and Caliskan, 2021; Wu et al., 2024a); and second, analyzing adversarial vulnerabilities via carefully crafted visual prompts for assessing…”
Metadata
- Authors
- Steed, R., & Caliskan, A.
- Venue
- Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency
- Publisher
- ACM
- Published
- 2021-03
- Pages
- 701-713
- OpenAlex
- W3095351420
- Cited by
- 157
- References
- 43
- FWCI
- 33.66
- Type
- conference paper
- Language
- en
- Open access
- hybrid
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
Data sources
Bibliographic data: Semantic Scholar, OpenAlex, Crossref.
Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).
Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.