Pith. sign in

REVIEW 3 major objections 4 minor 57 references

Biases Propagate in Encoder-based Vision-Language Models: A Systematic Analysis From Intrinsic Measures to Zero-shot Retrieval Outcomes

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Embedding-space bias in vision-language models carries into zero-shot retrieval, with intrinsic and extrinsic bias correlating at 0.83.

desk verdict The headline rho is not yet interpretable as propagation because the intrinsic and extrinsic measures share the same candidate pool and cosine-similarity structure; the paper needs an anchor-exclusion control before the claim holds. read the letter →

arxiv 2506.06506 v1 pith:A44XQ2XY submitted 2025-06-06 cs.CL

Kshitish Ghate , Tessa Charlesworth , Mona Diab , Aylin Caliskan This is my paper
classification cs.CL
keywords biaspropagationvision-languagemodelszero-shotretrievalSC-EATsocialgroupvalenceCLIPBLIP-2
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to show that biases already present in an encoder-based vision-language model's embedding space do not stay there: they reappear in the model's zero-shot retrieval outputs. It measures intrinsic bias with a Single Category-Embedding Association Test (SC-EAT) effect size per image or sentence, measures extrinsic bias as the valence or group content of the top-500 retrieved items, and then correlates the two across 114 analyses spanning six intersectional race-gender groups, three models, and both retrieval directions. The headline result is an average Spearman's $\rho = 0.83 \pm 0.10$, meaning items the model internally associates with negativity tend to be the ones that retrieve negative content. The authors also report that larger and better-performing models show stronger propagation, and that marginalized groups are less consistently propagated, which would produce more skewed outcomes for those groups. A sympathetic reader would therefore take the paper as evidence that fairness work cannot ignore the representational layer of encoder-based VLMs.

What carries the argument

The machinery is a measurement pair coupled by a fixed experimental scaffold. Intrinsic bias is the SC-EAT effect size, a Cohen's d-style normalized difference in cosine similarity between a target stimulus and two attribute sets; extrinsic bias is either the mean valence $\bar{V} = \frac{1}{N}\sum_{i=1}^N V_i$ of the top-$k=500$ retrieved items or the proportion $P_{\mathrm{grp}}$ of retrieved items belonging to a target group. The framework in Algorithm 1 fixes the candidate pool, sentence templates, and retrieval depth, then varies only content type (valence versus group), direction (image-to-text versus text-to-image), and social group, producing a Spearman's $\rho$ per analysis. What makes the correlation interpretable as propagation is that both metrics are evaluated on the same learned embedding geometry, so a high $\rho$ means the associations that show up in SC-EAT are the same associations that drive ranking in retrieval.

What would settle it

Recompute the intrinsic-extrinsic Spearman's rho after deleting from the top-500 retrieval candidate pool every SC-EAT anchor item used to define intrinsic bias (the 25 most positive and 25 most negative NRC-VAD sentences, the 25 most positive and 25 most negative OASIS images, and the 864 group-label sentences), keeping all other retrieved items in place. If the average rho stays near 0.83 across the 114 analyses, the overlap is not driving the result; if it drops substantially, the propagation claim is partly a measurement artifact.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that bias propagation from intrinsic representation to downstream zero-shot retrieval is systematic, high, and group-dependent. For each target stimulus (a CFD face image, an OASIS image, a valence-rated sentence, or an intersectional group-label sentence), the intrinsic metric is the SC-EAT effect size $d(\vec{w}, A, B) = \frac{\mathrm{mean}_{a\in A}\cos(\vec{w},\vec{a}) - \mathrm{mean}_{b\in B}\cos(\vec{w},\vec{b})}{\mathrm{std}_{x\in A\cup B}\cos(\vec{w},\vec{x})}$, and the extrinsic metric is the mean human-rated valence $\bar{V}$ or the group proportion $P_{\mathrm{grp}}$ of the top-500 retrieved items. The Spearman's correlation between these metrics, averaged over 114 analyses, is $\rho = 0.83 \pm 0.10$, with the baselines for pure valence and pure group-signal propagation at $0.86 \pm 0.04$ and $0.85 \pm 0.12$ respectively. The paper interprets the high, stable correlations as evidence that biases embedded in shared vision-language representations manifest in real zero-shot task outcomes, while acknowledging that the controlled design supports the directional language of propagation without fully establishing causal origin.

Load-bearing premise

The result rests on the untested assumption that removing the 25 most positive and 25 most negative words and images, plus the 864 group-label sentences, from the retrieval pool would not change the correlations, because those very items are part of the pool the model ranks.

Editorial extensions

If this is right

  • Zero-shot retrieval in CLIP-B-32, CLIP-L-14, and BLIP-2 does not neutralize embedded bias: if a group is internally associated with negativity, the top-500 retrieved items for that group skew negative in human-rated valence.
  • Propagation strength varies systematically by social group, with underrepresented groups showing less consistent intrinsic-to-extrinsic transfer and more skewed outcomes; aggregate correlation numbers therefore risk masking group-specific harm.
  • Larger and higher-performing models (e.g., CLIP-L-14 compared with CLIP-B-32) tend to show larger propagation correlations, so the trend toward larger vision-language models is itself a bias-propagation risk.
  • The two baseline tasks (valence-to-valence and group-to-group) give future work a benchmark for how faithfully a model can propagate a signal, against which bias propagation can be separated from ordinary content transfer.
  • Since the correlations are measured with zero-shot retrieval, mitigation aimed at downstream outcomes would have to work on the intrinsic representation or the ranking procedure itself, not only on fine-tuning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The framework is modular enough to extend beyond the three models studied; feeding the same intrinsic-extrinsic pipeline to newer encoder-based VLMs would test whether the scaling trend the authors observe continues.
  • The same logic could be applied to generative text-to-image models by using the generated images' valences as the extrinsic measure, a step the authors do not take.
  • A natural intervention experiment follows if the propagation claim is right: modifying the embedding space to neutralize the SC-EAT direction should reduce extrinsic retrieval bias proportionally; the paper does not run this, but its propagation result makes it a direct testable consequence.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

Summary. This paper proposes a framework for measuring 'bias propagation' in encoder-based vision-language models by correlating an intrinsic bias score (SC-EAT effect sizes) for individual images or sentences with an extrinsic outcome (mean valence or proportion of group-related items in the top-500 zero-shot retrievals). The framework is instantiated in 8 experiments covering image-to-text and text-to-image retrieval, valence and group signals, 6 racial/gender groups, and 3 models (CLIP-B-32, CLIP-L-14, BLIP-2), yielding a reported average Spearman rho of 0.83 +/- 0.10 across 114 analyses. The paper further reports that baselines for pure valence and group signals also show high correlations, that underrepresented groups show more variable propagation, and that larger models tend to show stronger propagation.

Significance. If the headline associations were validated, the paper would provide a useful audit tool and a large-scale empirical connection between representational bias and zero-shot retrieval outcomes. Strengths include the public code/data, the breadth of social groups and models, and the explicit framing around multiple analyses. However, the core validity of the headline rho depends on the intrinsic and extrinsic measures being meaningfully independent, and the paper does not establish that; the overlap between SC-EAT attribute sets and retrieval pools is a serious confound. The causal/propagation framing also exceeds the correlational evidence.

major comments (3)
  1. [Section 4.2 / Appendix D.2 (all experiments)] The intrinsic and extrinsic measures are constructed from the same candidate pool and the same cosine-similarity matrix in every experiment. For example, in 1-a the intrinsic SC-EAT score for a CFD image compares its cosine similarity to the top-25 positive and top-25 negative NRC-VAD sentences, while the extrinsic metric is the mean human valence of the top-500 NRC-VAD sentences retrieved by that same image; both use the same 20,000-sentence pool, the same embeddings, and the same human valence ratings. A high Spearman rho can therefore arise even in a model with no group-specific bias, simply because the two metrics are functions of the same anchor stimuli and labels. The paper's statement in Section 1 that the design is 'rigorously controlled' does not address this. Please run anchor-exclusion controls (remove the SC-EAT attribute items from the retrieval pool) and permutation/random-label nulls for all 8 experiments, and report whether the headline rho of 0.83 survives.
  2. [Abstract / Section 1 / Section 8] The terms 'propagate,' 'carry over,' and 'directional flow' imply a causal or directional mechanism, but the reported evidence is a Spearman correlation between two model-derived metrics. The Section 8 caveat that the design 'does not fully establish causality' is in tension with the abstract's causal framing and with the Section 1 statement that the results are 'substantiating the directional influence implied by propagation.' Please either reframe the claims as associations/consistency or provide a manipulation (e.g., intervention on representations or on the retrieval pool) that supports directionality.
  3. [Section 6 / Abstract (scaling claim)] The conclusion that larger/better-performing models exhibit greater bias propagation is supported only by comparisons among three models (CLIP-B-32, CLIP-L-14, BLIP-2) and selected group-level rho values (e.g., 0.80 vs. 0.88 in 1*-a; 0.78 vs. 0.88 for White Women in 1-a). No statistical test accounts for model identity, architecture, or pretraining data, and BLIP-2 differs architecturally from the CLIP models. Please either add a proper regression/test across a broader model sample or temper the scaling claim to a qualitative observation.
minor comments (4)
  1. [Section 6] The text reports '1143 analyses' while the abstract and Section 1 report 114 analyses; correct the typo.
  2. [Section 5.1] The phrase 'z-scored effect sizes (d)' is imprecise: SC-EAT d is a Cohen's d normalized by the standard deviation of cosine similarities, not a z-score; please use consistent terminology.
  3. [Section 8] There is an incomplete sentence: 'Furthermore, our' is followed directly by 'The use of 'propagation'...' which appears to be a formatting artifact; complete or remove the fragment.
  4. [Appendix A.1.1] The heading 'Operationlisation' appears twice and should be 'Operationalization'.

Circularity Check

3 steps flagged · score 7.0 of 10

Headline rho is partially circular: intrinsic SC-EAT and extrinsic retrieval metrics are computed from the same candidate pools, the same human valence/group labels, and the same cosine-similarity matrix, so the 0.83 'propagation' correlation reflects measurement coupling as well as any real downstream transfer.

  1. self definitional [Figure 1 caption; Section 4.3 'Baseline Valence-Valence Signal Propagation' and 'Valence-to-Group Signal Propagation (1-a and 1-b)']
    "The intrinsic SC-EAT effect size is computed based on the differential association of each CFD image with the top 25 positive and top 25 negative words in sentence templates. The extrinsic metric is derived from the mean valence of the top 500 words in sentence templates retrieved by each image."

    For each CFD image i, the intrinsic d_i is a normalized contrast between mean cos(e_i, s) over the 25 most positive and 25 most negative NRC-VAD sentences, while the extrinsic m_i is the mean of the same human valence ratings over the 500 sentences with highest cos(e_i, s) in the same 20,000-sentence NRC-VAD pool. Both are deterministic functions of the single vector (cos(e_i, s))_{s in S} and the single label vector v(s). The Spearman correlation between d_i and m_i over images is therefore largely a property of one ranking-plus-labeling procedure, not an independent measurement of propagation from intrinsic representation to downstream outcome. The paper reports no anchor-exclusion or held-out control that would break this shared-input coupling.

  2. self definitional [Appendix A.2 'Operationalisation of OASIS'; Section 4.3 'Valence-to-Group Signal Propagation (1-a and 1-b)' (TTI direction)]
    "In experiments involving valence SC-EAT, we use the top 25 pleasant and top 25 unpleasant OASIS images, sorted by valence ratings, as the two attribute sets. For TTI retrieval, all 900 images are employed."

    The same coupling appears in the text-to-image direction. The intrinsic SC-EAT for a group-label sentence uses as attributes the 50 OASIS images with the most extreme human valence ratings, and the extrinsic metric averages the same human valence ratings over the top-500 OASIS images retrieved by that sentence from the same 900-image pool. Thus the intrinsic valence association and the extrinsic retrieved-valence outcome are both functions of cos(sentence, image) against the same labeled image set; the correlation is partly built into the shared measurement, independent of any bias-propagation mechanism.

1 more flagged steps
  1. self definitional [Appendix D.2 'Propagation of Intrinsic Valence-based Group Associations to Extrinsic Group Outcomes'; Section 4.3 'Group-to-Valence Signal Propagation (2-a and 2-b)']
    "Intrinsically, we use the image-text group SC-EAT to find the association of a given valenced image (from OASIS) with text containing group terms belonging to a single intersectional identity (versus text uniformly representing all intersectional groups). Extrinsically, we obtain the corresponding proportion of group-identified sentences that are retrieved for the same OASIS image."

    In the group experiments, the SC-EAT attribute text (or images) and the retrieval pool are the same group-label sentences (5,184 sentences) or CFD images (6,000 images). The intrinsic score is a one-vs-all contrast of cosine similarities between an OASIS image and group-label sentences; the extrinsic score is the proportion of those same group-label sentences retrieved in the top-500 for the same OASIS image. Both metrics are transformations of the same group-label cosine rankings, so the reported correlation again inherits the shared item set rather than providing an independent test of propagation.

full rationale

The paper's headline claim is that intrinsic biases propagate to zero-shot retrieval, evidenced by average Spearman rho 0.83 across 114 analyses. The specific reductions above show that, in the flagship experiments and the baselines, the intrinsic and extrinsic metrics are not independent measurements: they are computed over the same candidate pools (NRC-VAD sentences, OASIS images, group-label sentences, CFD images), use the same human valence or group labels to define the SC-EAT attribute sets and to score retrieved items, and are evaluated against the same model embeddings and cosine similarities. Consequently, a positive correlation between the two metrics can arise by construction from the shared measurement structure, even in a model with no distinct group-specific propagation mechanism. The paper's Section 8 assertion of a 'rigorously controlled experimental design' addresses external confounds, not this within-design overlap; and its Section 1 caveat that the results 'do not fully establish causality' concerns causal direction, not the shared-item confound. The overlap is not total identity: the intrinsic metric is a d-score over 50 extreme items while the extrinsic metric is a mean-valence or proportion over top-500 retrieved items, so some residue of independent signal may remain; this is why the score is 7 rather than higher. No load-bearing self-citation chain is present: SC-EAT is grounded in the external Caliskan et al. (2017) method, and the paper's self-citations to Wolfe and Caliskan, Charlesworth et al., and Ghate et al. support dataset and related-work claims rather than forcing the propagation result. The circularity is in the measurement design: without anchor-exclusion, permutation, or held-out-item controls, the headline rho cannot be interpreted as evidence of independent propagation.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central propagation result relies on the validity of SC-EAT as a bias measure, the accuracy of human valence ratings, the neutrality of the sentence templates, and the unstated assumption that the overlap between SC-EAT attribute items and retrieval-pool items does not drive the correlation. The paper also relies on the selected models and social groups being representative.

free parameters (4)
  • retrieval_k = 500
    Number of retrieved items used for all extrinsic metrics; chosen for sample size and feasibility, not fitted.
  • attribute_set_size = 25 (top 25 pleasant/unpleasant)
    Number of extreme-valence items used as SC-EAT attribute sets; hand-chosen based on prior work.
  • per_group_images = 1,000
    Number of CFD images sampled per intersectional group for retrieval experiments; chosen for computational feasibility.
  • valence_threshold = 0.5
    Cutoff used in Figure 7 and group-association analyses to split positive/negative valence; chosen by convention.
assumptions (5)
  • domain assumption SC-EAT effect sizes measure meaningful bias in VLM embeddings, including valence and group associations.
    Paper relies on SC-EAT (Caliskan et al. 2017) as its intrinsic bias measure without independent validation in these specific VLM/CFD/OASIS settings.
  • domain assumption Human valence ratings in NRC-V AD and OASIS are accurate ground truth.
    Both intrinsic attribute selection and extrinsic outcome scoring use these ratings; errors would propagate into the correlation.
  • domain assumption The sentence templates are semantically neutral and do not introduce confounds.
    Templates from May et al. are treated as bleached; if templates themselves carry valence/group signal, intrinsic and extrinsic metrics are jointly affected.
  • ad hoc to paper Overlapping SC-EAT attribute items and retrieval-pool items do not materially inflate the intrinsic-extrinsic correlation.
    Unstated and untested; central to interpreting rho as propagation rather than measurement coupling.
  • ad hoc to paper A Spearman correlation between intrinsic and extrinsic metrics indicates directional bias propagation.
    Correlation alone does not establish directionality; the paper infers propagation from it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Biases Propagate in Encoder-based Vision-Language Models: A Systematic Analysis From Intrinsic Measures to Zero-shot Retrieval Outcomes." pith.science (2026). https://pith.science/paper/A44XQ2XY

@misc{pith2026250606506,
  author       = {Pith},
  title        = {Pith review of: Biases Propagate in Encoder-based Vision-Language Models: A Systematic Analysis From Intrinsic Measures to Zero-shot Retrieval Outcomes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/A44XQ2XY}},
  note         = {Machine review of arXiv:2506.06506}
}
abstract

To build fair AI systems we need to understand how social-group biases intrinsic to foundational encoder-based vision-language models (VLMs) manifest in biases in downstream tasks. In this study, we demonstrate that intrinsic biases in VLM representations systematically ``carry over'' or propagate into zero-shot retrieval tasks, revealing how deeply rooted biases shape a model's outputs. We introduce a controlled framework to measure this propagation by correlating (a) intrinsic measures of bias in the representational space with (b) extrinsic measures of bias in zero-shot text-to-image (TTI) and image-to-text (ITT) retrieval. Results show substantial correlations between intrinsic and extrinsic bias, with an average $\rho$ = 0.83 $\pm$ 0.10. This pattern is consistent across 114 analyses, both retrieval directions, six social groups, and three distinct VLMs. Notably, we find that larger/better-performing models exhibit greater bias propagation, a finding that raises concerns given the trend towards increasingly complex AI models. Our framework introduces baseline evaluation tasks to measure the propagation of group and valence signals. Investigations reveal that underrepresented groups experience less robust propagation, further skewing their model-related outcomes.

Figures

Figures reproduced from arXiv: 2506.06506 by the authors.

Figure 1
Figure 1. The setup to analyze valence-based bias propagation in Image-to-Text retrieval. Propagation is measured [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Exp. 1*-a and 1*-b illustrate Spearman’s [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Exp. 2*-a and 2*-b illustrate Spearman’s [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Exp. 2-a and 2-b illustrate Spearman’s ρ mea￾suring the correlation of intrinsic valence-based group bias propagation to extrinsic group outcomes in zero￾shot retrieval tasks, stratified by social groups. across experiments 2-a and 2-b indicate “White Men” and “White W…
Figure 7
Figure 7. Figure 7: Group-based Intrinsic Bias Visualization [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Figure details the specifics of the 8 experiments used to show the propagation of intrinsic bias to extrinsic [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 39 canonical work pages

  1. [1]

    Manuele Barraco, Marcella Cornia, Silvia Cascianelli, Lorenzo Baraldi, and Rita Cucchiara. 2022. The unreasonable effectiveness of clip features for image captioning: an experimental analysis. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4662--4670

    work page 2022

  2. [2]

    Hugo Berg, Siobhan Hall, Yash Bhalgat, Hannah Kirk, Aleksandar Shtedritski, and Max Bain. 2022. https://doi.org/10.18653/v1/2022.aacl-main.61 A prompt array keeps the bias away: Debiasing vision-language models with adversarial learning . In Proceedings of the 2nd Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and ...

    doi:10.18653/v1/2022.aacl-main.61 2022

  3. [3]

    Su Lin Blodgett, Solon Barocas, Hal Daum \'e III, and Hanna Wallach. 2020. https://doi.org/10.18653/v1/2020.acl-main.485 Language (technology) is power: A critical survey of bias in NLP . In Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5454--5476, Online. Association for Computational Linguistics

    doi:10.18653/v1/2020.acl-main.485 2020

  4. [4]

    James Briggs and Laura Laura. 2022. https://www.pinecone.io/learn/series/image-search/zero-shot-object-detection-clip/ Zero shot object detection with openai’s clip

    work page 2022

  5. [5]

    Huy-Giap Bui, Minh-Huy Trinh, Canh-Toan Le, Quoc-Lam Vu, and Khac-Trieu Vo. 2023. Zero-shot video retrieval using clip with temporally ordered multi-query scoring. In Proceedings of the 12th International Symposium on Information and Communication Technology, pages 938--944

    work page 2023

  6. [6]

    Laura Cabello, Emanuele Bugliarello, Stephanie Brandl, and Desmond Elliott. 2023. https://doi.org/10.18653/v1/2023.emnlp-main.525 Evaluating bias and fairness in gender-neutral pretrained vision-and-language models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pages 8465--8483, Singapore. Association for Comp...

    work page doi:10.18653/v1/2023.emnlp-main.525 2023

  7. [7]

    Aylin Caliskan. 2023. Artificial intelligence, bias, and ethics. In Proceedings of the Thirty-Second International Joint Conference on Artificial Intelligence (IJCAI)

    work page 2023

  8. [8]

    Aylin Caliskan, Pimparkar Parth Ajay, Tessa Charlesworth, Robert Wolfe, and Mahzarin R Banaji. 2022. Gender bias in word embeddings: a comprehensive analysis of frequency, syntax, and semantics. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society, pages 156--170

    2022

Show all 57 references
  1. [9]

    Aylin Caliskan, Joanna J Bryson, and Arvind Narayanan. 2017. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183--186

    2017

  2. [10]

    Yang Trista Cao, Yada Pruksachatkun, Kai-Wei Chang, Rahul Gupta, Varun Kumar, Jwala Dhamala, and Aram Galstyan. 2022. https://doi.org/10.18653/v1/2022.acl-short.62 On the intrinsic and extrinsic fairness evaluation metrics for contextualized language representations . In Proce...

    2022 doi

  3. [11]

    Tessa ES Charlesworth and Mahzarin R Banaji. 2022. Word embeddings reveal social group attitudes and stereotypes in large language corpora. Handbook of language analysis in psychology, pages 494--508

    2022

  4. [12]

    Tessa ES Charlesworth, Aylin Caliskan, and Mahzarin R Banaji. 2022. Historical representations of social groups across 200 years of word embeddings from google books. Proceedings of the National Academy of Sciences, 119(28):e2121798119

    2022

  5. [13]

    Tessa ES Charlesworth, Kshitish Ghate, Aylin Caliskan, and Mahzarin R Banaji. 2024. Extracting intersectional stereotypes from embeddings: Developing and validating the flexible intersectional stereotype extraction procedure. PNAS nexus, 3(3):pgae089

    2024

  6. [14]

    Tessa ES Charlesworth, Mayan Navon, Yoav Rabinovich, Nicole Lofaro, and Benedek Kurdi. 2023. The project implicit international dataset: Measuring implicit and explicit social group attitudes and stereotypes across 34 countries (2009--2019). Behavior Research Methods, 55(3):1413--1440

    2023

  7. [15]

    Jacob Cohen. 2013. Statistical power analysis for the behavioral sciences. Academic press

    2013

  8. [16]

    Thierry Devos and Mahzarin R Banaji. 2005. American= white? Journal of personality and social psychology, 88(3):447

    2005

  9. [17]

    A Eagly and Shelly Chaiken. 1998. Attitude structure. Handbook of social psychology, 1:269--322

    1998

  10. [18]

    Shangbin Feng, Chan Young Park, Yuhan Liu, and Yulia Tsvetkov. 2023. https://doi.org/10.18653/v1/2023.acl-long.656 From pretraining data to language models to downstream tasks: Tracking the trails of political biases leading to unfair NLP models . In Proceedings of the 61st An...

    2023 doi

  11. [19]

    you don't look mexican!

    Justin D Garc \' a. 2013. “you don't look mexican!” my life in ethnic ambiguity and what it says about the construction of race in america. Multicultural Perspectives, 15(4):234--238

    2013

  12. [20]

    Diab, and Aylin Caliskan

    Kshitish Ghate, Isaac Slaughter, Kyra Wilson, Mona T. Diab, and Aylin Caliskan. 2025. https://aclanthology.org/2025.naacl-long.148/ Intrinsic bias is predicted by pretraining data and correlates with downstream performance in vision-language encoders . In Proceedings of the 20...

    2025

  13. [21]

    Negin Ghavami and Letitia Anne Peplau. 2013. An intersectional analysis of gender and ethnic stereotypes: Testing three hypotheses. Psychology of Women Quarterly, 37(1):113--127

    2013

  14. [22]

    Sourojit Ghosh and Aylin Caliskan. 2023. https://doi.org/10.18653/v1/2023.findings-emnlp.465 person' == light-skinned, western man, and sexualization of women of color: Stereotypes in stable diffusion . In Findings of the Association for Computational Linguistics: EMNLP 2023, ...

    2023 doi

  15. [23]

    Anthony G Greenwald, Debbie E McGhee, and Jordan LK Schwartz. 1998. Measuring individual differences in implicit cognition: the implicit association test. Journal of personality and social psychology, 74(6):1464

    1998

  16. [24]

    Arnold K Ho, Jim Sidanius, Daniel T Levin, and Mahzarin R Banaji. 2011. Evidence for hypodescent and racial hierarchy in the categorization and perception of biracial individuals. Journal of personality and social psychology, 100(3):492

    2011

  17. [25]

    Rachel Hong, William Agnew, Tadayoshi Kohno, and Jamie Morgenstern. 2024. Who's in and who's out? a case study of multimodal clip-filtering in datacomp. In Proceedings of the 4th ACM Conference on Equity and Access in Algorithms, Mechanisms, and Optimization, pages 1--17

    2024

  18. [26]

    Kimmo Karkkainen and Jungseock Joo. 2021. Fairface: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1548--1558

    2021

  19. [27]

    Fanjie Kong, Shuai Yuan, Weituo Hao, and Ricardo Henao. 2024. Mitigating test-time bias for fair image retrieval. Advances in Neural Information Processing Systems, 36

    2024

  20. [28]

    Benedek Kurdi, Shayn Lozano, and Mahzarin R Banaji. 2017. Introducing the open affective standardized image set (oasis). Behavior research methods, 49:457--470

    2017

  21. [29]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In International conference on machine learning, pages 19730--19742. PMLR

    2023

  22. [30]

    Shayne Longpre, Gregory Yauney, Emily Reif, Katherine Lee, Adam Roberts, Barret Zoph, Denny Zhou, Jason Wei, Kevin Robinson, David Mimno, and Daphne Ippolito. 2024. https://doi.org/10.18653/v1/2024.naacl-long.179 A pretrainer ' s guide to training data: Measuring the effects o...

    2024 doi

  23. [31]

    Debbie S Ma, Joshua Correll, and Bernd Wittenbrink. 2015. The chicago face database: A free stimulus set of faces and norming data. Behavior research methods, 47:1122--1135

    2015

  24. [32]

    Bowman, and Rachel Rudinger

    Chandler May, Alex Wang, Shikha Bordia, Samuel R. Bowman, and Rachel Rudinger. 2019. https://doi.org/10.18653/v1/N19-1063 On measuring social biases in sentence encoders . In Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational...

    2019 doi

  25. [33]

    Saif Mohammad. 2018. Obtaining reliable human ratings of valence, arousal, and dominance for 20,000 english words. In Proceedings of the 56th annual meeting of the association for computational linguistics (volume 1: Long papers), pages 174--184

    2018

  26. [34]

    Veronica A Newton. 2023. Hypervisibility and invisibility: Black women’s experiences with gendered racial microaggressions on a white campus. Sociology of Race and Ethnicity, 9(2):164--178

    2023

  27. [35]

    Gandalf Nicolas and Susan T Fiske. 2023. Valence biases and emergence in the stereotype content of intersecting social categories. Journal of Experimental Psychology: General

    2023

  28. [36]

    Charles Egerton Osgood, George J Suci, and Percy H Tannenbaum. 1957. The measurement of meaning. 47. University of Illinois press

    1957

  29. [37]

    Wasin Pirom. 2022. Object detection and position using clip with thai voice command for thai visually impaired. In 2022 37th International Technical Conference on Circuits/Systems, Computers and Communications (ITC-CSCC), pages 391--394. IEEE

    2022

  30. [38]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. In International conference on machine learni...

    2021

  31. [39]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj\"orn Ommer. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684--10695

    2022

  32. [40]

    Charles Spearman. 1961. The proof and measurement of association between two things

    1961

  33. [41]

    Tejas Srinivasan and Yonatan Bisk. 2022. https://doi.org/10.18653/v1/2022.gebnlp-1.10 Worst of both worlds: Biases compound in pre-trained vision-and-language models . In Proceedings of the 4th Workshop on Gender Bias in Natural Language Processing (GeBNLP), pages 77--85, Seat...

    2022 doi

  34. [42]

    Ryan Steed and Aylin Caliskan. 2021. Image representations learned with unsupervised pre-training contain human-like biases. In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 701--713

    2021

  35. [43]

    Ryan Steed, Swetasudha Panda, Ari Kobren, and Michael Wick. 2022. Upstream mitigation is not all you need: Testing the bias transfer hypothesis in pre-trained language models. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1:...

    2022

  36. [44]

    Weijie Su, Xizhou Zhu, Yue Cao, Bin Li, Lewei Lu, Furu Wei, and Jifeng Dai. 2019. Vl-bert: Pre-training of generic visual-linguistic representations. arXiv preprint arXiv:1908.08530

    2019 arXiv

  37. [45]

    Mohammad Reza Taesiri, Finlay Macklon, and Cor-Paul Bezemer. 2022. Clip meets gamephysics: Towards bug identification in gameplay videos using zero-shot transfer learning. In Proceedings of the 19th International Conference on Mining Software Repositories, pages 270--281

    2022

  38. [46]

    Yi Chern Tan and L Elisa Celis. 2019. Assessing social and intersectional biases in contextualized word representations. Advances in neural information processing systems, 32

    2019

  39. [47]

    Autumn Toney and Aylin Caliskan. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.574 V al N orm quantifies semantics to reveal consistent valence biases across languages and over centuries . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Proce...

    2021 doi

  40. [48]

    Madalina Vlasceanu and David M Amodio. 2022. Propagation of societal gender inequality by internet search algorithms. Proceedings of the National Academy of Sciences, 119(29):e2204529119

    2022

  41. [49]

    Robert Wolfe, Mahzarin R Banaji, and Aylin Caliskan. 2022. Evidence for hypodescent in visual semantic ai. In 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1293--1304

    2022

  42. [50]

    Robert Wolfe and Aylin Caliskan. 2022 a . American== white in multimodal language-and-image ai. In Proceedings of the 2022 AAAI/ACM Conference on AI, Ethics, and Society, pages 800--812

    2022

  43. [51]

    Robert Wolfe and Aylin Caliskan. 2022 b . Contrastive visual semantic pretraining magnifies the semantics of natural language representations. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3050--3061

    2022

  44. [52]

    Robert Wolfe and Aylin Caliskan. 2022 c . Markedness in visual semantic ai. In Proceedings of the 2022 ACM Conference on Fairness, Accountability, and Transparency, pages 1269--1279

    2022

  45. [53]

    Robert Wolfe and Aylin Caliskan. 2022 d . Vast: The valence-assessing semantics test for contextualizing language models. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 36, pages 11477--11485

    2022

  46. [54]

    Robert Wolfe, Yiwei Yang, Bill Howe, and Aylin Caliskan. 2023. Contrastive language-vision ai models pretrained on web-scraped multimodal data exhibit sexual objectification bias. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Transparency, pages 1174--1185

    2023

  47. [55]

    Kankan Zhou, Yibin LAI, and Jing Jiang. 2022. Vlstereoset: A study of stereotypical bias in pre-trained vision-language models. Association for Computational Linguistics

    2022

  48. [56]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  49. [57]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.