arxiv: 2605.13113 · v1 · submitted 2026-05-13 · 💻 cs.CY · cs.AI

Recognition: 2 theorem links

· Lean Theorem

Context Matters: Auditing Gender Bias in T2I Generation through Risk-Tiered Use-Case Profiles

Jose Luna , Yankun Wu , Xiaofei Xie , Noa Garcia

Authors on Pith no claims yet

Pith reviewed 2026-05-14 18:31 UTC · model grok-4.3

classification 💻 cs.CY cs.AI

keywords gender biastext-to-image generationauditing frameworkrisk assessmentharm typologyT2I modelsEU AI Actbias metrics

0 comments

The pith

Text-to-image models require gender bias audits that align with the risk level of their specific use cases.

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

The paper develops a framework to make audits for gender bias in text-to-image generation more useful by tying them to how the models are deployed. It creates risk-based profiles for different applications, groups bias measurement methods into three categories, and links them to context-specific harms. This approach aims to move beyond fragmented metrics that ignore whether the system is used in low-stakes or high-stakes settings. By connecting these elements, the framework supports more targeted evaluations that reflect real exposure to harms in areas like education and media.

Core claim

The authors propose a risk-aligned auditing framework for gender bias in T2I models that consists of three parts: risk-tiered use-case profiles based on regulatory categories, a catalog of metrics organized into gender prediction, embedding similarity, and downstream task types, and a harm typology that maps representational and other harms to specific scenarios. This is operationalized through THUMB cards that incorporate context, bias manifestations, harm hypotheses, and audit strategies to guide systematic evaluation.

What carries the argument

THUMB cards that combine context, scenario details, bias manifestations, harm hypotheses, and chosen audit strategies into a single planning tool.

If this is right

Auditors can choose evaluation metrics according to the risk tier of the deployment context.
Harm assessments become more precise by mapping bias types to specific use-case risks.
Evaluations gain consistency through the standardized metric catalog across different studies.
Regulatory compliance efforts can reference the aligned risk categories and harm typologies.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If adopted, this could lead to more nuanced regulations that differentiate bias auditing requirements based on intended use.
Testing the framework on current T2I models might reveal gaps in how well the three metric categories cover emerging bias issues.
Extending the approach to other biases like race or age could build on the same risk-harms-metrics structure.
Developers might use the harm typology to prioritize fixes in high-risk applications first.

Load-bearing premise

Existing gender bias metrics can be grouped into three measurement categories and linked to context-specific harms without losing important information or validity.

What would settle it

An experiment applying the framework to multiple T2I models where the consolidated metrics miss a known bias that individual metrics detect in a high-risk use case.

Figures

Figures reproduced from arXiv: 2605.13113 by Jose Luna, Noa Garcia, Xiaofei Xie, Yankun Wu.

**Figure 1.** Figure 1: Overview of the T2I auditing framework, which is formed by three constituents and operationalized through the [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

**Figure 2.** Figure 2: THUMB Card — High-Risk Use-Case Profile: Law Enforcement Forensic Imagery. [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗

**Figure 3.** Figure 3: THUMB Card — Limited-Risk Use-Case Profile: Public-Facing Communication Imagery [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

**Figure 4.** Figure 4: THUMB Card — Minimal-Risk Use-Case Profile: Educational and Creative Imagery [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗

read the original abstract

Text-to-image (T2I) generative models are increasingly used to produce content for education, media, and public-facing communication, and are starting to be integrated into higher-impact pipelines. Since generated images tend to reinforce stereotypes, producing representational erasure via "default" depictions and shaping perceptions of who belongs in certain roles, a growing body of work has proposed metrics to quantify gender bias in T2I outputs. Yet existing evaluations remain fragmented. Metrics are often reported without a shared view of what they measure, what assumptions they entail, or how their results should be interpreted under different deployment contexts. This limits the usefulness of gender bias measurement for both technical auditing and emerging governance discussions. We propose a risk-aligned auditing framework for gender bias in T2I models composed of three constituents that connects risk categories, evaluation metrics, and harms. First, we identify risk-tiered use-case profiles aligned with the EU AI Act's risk categories to motivate why auditing expectations may vary with deployment contexts and stakeholder exposure. Second, we construct a metric catalog that consolidates gender-bias evaluation methods and organizes them in three measurement categories: gender prediction, embedding similarity, and downstream task. Third, we introduce a harm typology that maps context-dependent harm categories (e.g., representational, quality-of-service) to specific risk-tired scenarios. Finally, we introduce THUMB cards (Text-to-image Harms-informed Use-case-aligned Metrics of gender Bias) that help formulate auditing systematically by the incorporation of context, scenario and bias manifestation, harm hypotheses, and audit strategy.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper organizes gender bias metrics for T2I models into a risk-tiered framework tied to the EU AI Act with THUMB cards for auditing, which helps with consistency but stays conceptual without tests.

read the letter

The main takeaway is that this paper pulls scattered gender bias metrics for text-to-image models into one structure: risk profiles based on EU AI Act categories, a three-way metric catalog covering gender prediction, embedding similarity, and downstream tasks, plus a harm typology and THUMB cards to guide context-specific audits. The consolidation step is the clearest addition, since prior work often reports metrics in isolation without shared assumptions or deployment links. Aligning the whole thing with regulatory risk levels makes it more usable for people who need to justify evaluations to stakeholders or policymakers. The THUMB card format looks like a practical way to document scenario, harm hypothesis, and audit choices in one place. The limitation is that the paper offers no new data, model runs, or validation to check whether the category mappings hold up or miss key cases in practice. Everything rests on logical grouping rather than evidence that the framework improves audit outcomes or covers real harms better than existing ad-hoc approaches. This is aimed at fairness researchers and auditors who want an organizing tool for bias work, especially those dealing with governance or standards. Readers focused on frameworks will find the structure helpful even if they have to fill in the implementation details themselves. It should go to peer review because the problem of fragmented metrics is real and the proposed connections are clear enough to get useful feedback on the details.

Referee Report

2 major / 2 minor

Summary. The paper proposes a risk-aligned auditing framework for gender bias in text-to-image (T2I) generative models. It consists of three constituents: (1) risk-tiered use-case profiles aligned with the EU AI Act's risk categories to differentiate auditing expectations by deployment context and stakeholder exposure; (2) a metric catalog that consolidates existing gender-bias evaluation methods into three measurement categories (gender prediction, embedding similarity, and downstream task); and (3) a harm typology that maps context-dependent harm categories (e.g., representational, quality-of-service) to specific risk-tiered scenarios. The framework is operationalized via THUMB cards (Text-to-image Harms-informed Use-case-aligned Metrics of gender Bias) that incorporate context, scenario, bias manifestation, harm hypotheses, and audit strategy.

Significance. If the mappings hold, the framework offers a structured synthesis that connects fragmented gender-bias metrics to deployment contexts and harms, improving the interpretability of audits for both technical and governance purposes. The explicit alignment with the EU AI Act and the introduction of THUMB cards as an organizational tool are concrete strengths that could facilitate more consistent auditing practices. The contribution is primarily conceptual and definitional rather than empirical, so its significance will depend on adoption and validation in follow-on work.

major comments (2)

[Metric Catalog] § Metric Catalog: the consolidation of prior metrics into exactly three categories (gender prediction, embedding similarity, downstream task) is presented as a definitional step, but without an explicit coverage table or boundary-case analysis it is unclear whether the organization avoids gaps or overlaps that would reduce validity across T2I models and deployment scenarios.
[Harm Typology] § Harm Typology: the mapping of harm categories to risk-tiered scenarios is offered without concrete worked examples or a check that original metric assumptions are preserved, which is load-bearing for the central claim that the framework enables context-dependent auditing without significant loss of coverage.

minor comments (2)

[Abstract] The abstract contains the typo 'risk-tired scenarios' (should be 'risk-tiered').
[THUMB cards] THUMB acronym expansion is given only in the abstract; repeat the expansion on first use in the main text for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback and positive assessment of the THUMB framework. We have revised the manuscript to address the two major comments by adding the requested coverage table and worked examples.

read point-by-point responses

Referee: [Metric Catalog] § Metric Catalog: the consolidation of prior metrics into exactly three categories (gender prediction, embedding similarity, downstream task) is presented as a definitional step, but without an explicit coverage table or boundary-case analysis it is unclear whether the organization avoids gaps or overlaps that would reduce validity across T2I models and deployment scenarios.

Authors: We agree that an explicit coverage table and boundary-case analysis would strengthen the presentation. In the revised manuscript we have added Table 2, which maps representative metrics from the literature to the three categories, flags boundary cases (e.g., hybrid prediction-similarity metrics), and discusses potential overlaps. The table confirms that the categorization covers the dominant evaluation approaches without material gaps, thereby preserving validity across T2I models and scenarios. revision: yes
Referee: [Harm Typology] § Harm Typology: the mapping of harm categories to risk-tiered scenarios is offered without concrete worked examples or a check that original metric assumptions are preserved, which is load-bearing for the central claim that the framework enables context-dependent auditing without significant loss of coverage.

Authors: We accept this observation. The revised manuscript now includes a new subsection (5.3) containing worked examples for each risk tier. Each example specifies the use-case profile, selected metric, harm hypothesis, and an explicit verification that the metric's original assumptions (e.g., label definitions or embedding spaces) are left unchanged. These examples demonstrate that context-dependent auditing is possible without loss of coverage or metric integrity. revision: yes

Circularity Check

0 steps flagged

No significant circularity in conceptual framework proposal

full rationale

The paper advances a risk-aligned auditing framework for gender bias in T2I models by synthesizing existing metrics into three measurement categories, aligning them with EU AI Act risk tiers, and introducing a harm typology plus THUMB cards. This is presented as an organizational and definitional contribution rather than a derivation from equations or fitted parameters. No load-bearing steps reduce to self-definition, fitted inputs renamed as predictions, or self-citation chains; the consolidation of prior work is explicitly framed as a new organizational layer without quantitative claims that would be tautological. The central proposal remains self-contained against external benchmarks and does not rely on uniqueness theorems or ansatzes imported from the authors' prior work.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

The framework rests on the assumption that EU AI Act risk categories transfer appropriately to T2I generation contexts and that existing metrics retain their meaning when reorganized into the proposed categories.

axioms (1)

domain assumption EU AI Act risk categories are suitable for classifying T2I deployment contexts and associated gender bias harms
The paper aligns its use-case profiles directly with these categories without additional justification in the abstract.

invented entities (1)

THUMB cards no independent evidence
purpose: To formulate auditing systematically by incorporating context, scenario, bias manifestation, harm hypotheses, and audit strategy
New artifact introduced by the paper to operationalize the framework.

pith-pipeline@v0.9.0 · 5589 in / 1314 out tokens · 36719 ms · 2026-05-14T18:31:59.472267+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Foundation/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

We propose a risk-aligned auditing framework for gender bias in T2I models composed of three constituents that connects risk categories, evaluation metrics, and harms.
IndisputableMonolith/Foundation/AbsoluteFloorClosure.lean reality_from_one_distinction unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

THUMB cards (Text-to-image Harms-informed Use-case-aligned Metrics of gender Bias)

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

114 extracted references · 114 canonical work pages · 3 internal anchors

[1]

https://standards.ieee.org/ieee/7003/11357/

2024.IEEE Standard for Algorithmic Bias Considerations. https://standards.ieee.org/ieee/7003/11357/

work page 2024
[2]

InstructBlip-2

2025. InstructBlip-2. https://huggingface.co/nnpy/Instruct-blip-v2

work page 2025
[3]

2025.General Purpose AI (GPAI): High-Level Summary of the AI Act

ArtificialIntelligenceAct.eu. 2025.General Purpose AI (GPAI): High-Level Summary of the AI Act. https://artificialintelligenceact.eu/high- level-summary/ Accessed: 2025-11-19

work page 2025
[4]

Anusha Asim. 2026. Through the AI looking glass: measuring gendered objectification in user-generated AI images.AI and Ethics6, 1 (2026), 19

work page 2026
[5]

Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. 2023. Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond.arXiv preprint arXiv:2308.12966(2023)

work page internal anchor Pith review Pith/arXiv arXiv 2023
[6]

Eslam Mohamed Bakr, Pengzhan Sun, Xiaogian Shen, Faizan Farooq Khan, Li Erran Li, and Mohamed Elhoseiny. 2023. HRS-Bench: Holistic, reliable and scalable benchmark for text-to-image models. InICCV

work page 2023
[7]

2023.Fairness and machine learning: Limitations and opportunities

Solon Barocas, Moritz Hardt, and Arvind Narayanan. 2023.Fairness and machine learning: Limitations and opportunities. MIT press

work page 2023
[8]

James Betker, Gabriel Goh, Li Jing, Tim Brooks, Jianfeng Wang, Linjie Li, Long Ouyang, Juntang Zhuang, Joyce Lee, Yufei Guo, et al

work page
[9]

https://cdn.openai.com/papers/dall-e-3.pdf2, 3 (2023), 8

Improving image generation with better captions.Computer Science. https://cdn.openai.com/papers/dall-e-3.pdf2, 3 (2023), 8

work page 2023
[10]

Federico Bianchi, Pratyusha Kalluri, Esin Durmus, Faisal Ladhak, Myra Cheng, Debora Nozza, Tatsunori Hashimoto, Dan Jurafsky, James Zou, and Aylin Caliskan. 2023. Easily accessible text-to-image generation amplifies demographic stereotypes at large scale. In Proceedings of the 2023 ACM conference on fairness, accountability, and transparency. 1493–1504

work page 2023
[11]

Charlotte Bird, Eddie Ungless, and Atoosa Kasirzadeh. 2023. Typology of risks of generative text-to-image models. InProceedings of the 2023 AAAI/ACM Conference on AI, Ethics, and Society. 396–410

work page 2023
[12]

Joy Buolamwini and Timnit Gebru. 2018. Gender shades: Intersectional accuracy disparities in commercial gender classification. In FAccT. PMLR, 77–91

work page 2018
[13]

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

work page 2017
[14]

Nicolas Carlini, Jamie Hayes, Milad Nasr, Matthew Jagielski, Vikash Sehwag, Florian Tramer, Borja Balle, Daphne Ippolito, and Eric Wallace. 2023. Extracting training data from diffusion models. InUSENIX Security Symposium

work page 2023
[15]

Mathilde Caron, Hugo Touvron, Ishan Misra, Hervé Jégou, Julien Mairal, Piotr Bojanowski, and Armand Joulin. 2021. Emerging properties in self-supervised vision transformers. InICCV

work page 2021
[16]

Jane Castleman and Aleksandra Korolova. 2025. Adultification Bias in LLMs and Text-to-Image Models. InProceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency. 2751–2767

work page 2025
[17]

Jun Chen, Deyao Zhu, Xiaoqian Shen, Xiang Li, Zechun Liu, Pengchuan Zhang, Raghuraman Krishnamoorthi, Vikas Chandra, Yunyang Xiong, and Mohamed Elhoseiny. 2023. Minigpt-v2: large language model as a unified interface for vision-language multi-task learning. arXiv preprint arXiv:2310.09478(2023)

work page arXiv 2023
[18]

Aditya Chinchure, Pushkar Shukla, Gaurav Bhatt, Kiri Salij, Kartik Hosanagar, Leonid Sigal, and Matthew Turk. 2023. TIBET: Identifying and Evaluating Biases in Text-to-Image Generative Models.arXiv preprint arXiv:2312.01261(2023)

work page arXiv 2023
[19]

Chloe Xiang. 2023. Developers Created AI to Generate Police Sketches. Experts Are Horrified.VICE(2023). https://www.vice.com/en/ article/ai-police-sketches/ Accessed 2025-12-15

work page 2023
[20]

Jaemin Cho, Abhay Zala, and Mohit Bansal. 2023. Dall-Eval: Probing the reasoning skills and social biases of text-to-image generation models. InICCV

work page 2023
[21]

CompVis. 2022. Model card for Stable diffusion safety checker. https://huggingface.co/CompVis/stable-diffusion-safety-checker

work page 2022
[22]

Sepehr Dehdashtian, Gautam Sreekumar, and Vishnu Naresh Boddeti. 2025. Oasis uncovers: High-quality t2i models, same old stereotypes. InICLR

work page 2025
[23]

Jiankang Deng, Jia Guo, Evangelos Ververas, Irene Kotsia, and Stefanos Zafeiriou. 2020. Retinaface: Single-shot multi-level face localisation in the wild. InCVPR

work page 2020
[24]

Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. 2019. Arcface: Additive angular margin loss for deep face recognition. In CVPR

work page 2019
[25]

Moreno D’Incà, Elia Peruzzo, Massimiliano Mancini, Dejia Xu, Vidit Goel, Xingqian Xu, Zhangyang Wang, Humphrey Shi, and Nicu Sebe. 2024. OpenBias: Open-set Bias Detection in Text-to-Image Generative Models. InCVPR

work page 2024
[26]

Wala Elsharif, Mahmood Alzubaidi, and Marco Agus. 2025. Cultural Bias in Text-to-Image Models: A Systematic Review of Bias Identification, Evaluation and Mitigation Strategies.IEEE Access(2025)

work page 2025
[27]

European Commission. 2024. AI Act. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai Regulation (EU) 2024/1689. Retrieved January 5, 2026

work page 2024
[28]

European Parliament and Council of the European Union. 2024. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 laying down harmonised rules on artificial intelligence and amending Regulations (EC) No 300/2008, (EU) No 167/2013, (EU) No 168/2013, (EU) 2018/858, (EU) 2018/1139 and (EU) 2019/2144 and Directives 2014/90/E...

work page 2024
[29]

European Union. 2024. Artificial Intelligence Act: Annex III - High-Risk AI Systems Referred to in Article 6(2). Official Journal of the European Union. https://artificialintelligenceact.eu/annex/3/ Regulation (EU) 2024/1689, Annex III

work page 2024
[30]

European Union. 2024. Artificial Intelligence Act: Article 5 - Prohibited AI Practices. Official Journal of the European Union. https://artificialintelligenceact.eu/article/5/ Regulation (EU) 2024/1689, Article 5

work page 2024
[31]

European Union. 2024. Recital 48 - Fundamental rights and risk assessment of AI systems. https://ai-act-law.eu/recital/48/

work page 2024
[32]

Artur Fortunato and Filipe Reynaud. 2022. Forensic Sketch AIrtist. LabLab.ai Hackathon Project. https://lablab.ai/ai-hackathons/openai- whisper-gpt3-codex-dalle2-hackathon/eagleai/forensic-sketch-airtist

work page 2022
[33]

Felix Friedrich, Katharina Hämmerl, Patrick Schramowski, Manuel Brack, Jindřich Libovick`y, Alexander Fraser, and Kristian Kersting

work page
[34]

Multilingual text-to-image generation magnifies gender stereotypes. InACL

work page
[35]

Zihao Fu, Ryan Brown, Shun Shao, Kai Rawal, Eoin D Delaney, and Chris Russell. 2025. FairImagen: Post-Processing for Bias Mitigation in Text-to-Image Models. InNeurIPS

work page 2025
[36]

Future of Life Institute. 2024. Implementation Timeline | EU Artificial Intelligence Act. https://artificialintelligenceact.eu/implementation- timeline/

work page 2024
[37]

Noa Garcia, Yusuke Hirota, Yankun Wu, and Yuta Nakashima. 2023. Uncurated Image-Text Datasets: Shedding Light on Demographic Bias. InCVPR

work page 2023
[38]

2024.Accuracy in criminal statistics matters

Richard Garside. 2024.Accuracy in criminal statistics matters. https://www.crimeandjustice.org.uk/accuracy-criminal-statistics-matters Accessed: 2026-01-03

work page 2024
[39]

Sourojit Ghosh and Aylin Caliskan. 2023. ‘Person’== Light-skinned, Western Man, and Sexualization of Women of Color: Stereotypes in Stable Diffusion. InEMNLP Findings

work page 2023
[40]

I Don’t See Myself Represented Here at All

Sourojit Ghosh, Nina Lutz, and Aylin Caliskan. 2024. “I Don’t See Myself Represented Here at All”: User Experiences of Stable Diffusion Outputs Containing Representational Harms across Gender Identities and Nationalities. InAIES, Vol. 7. 463–475

work page 2024
[41]

Sourojit Ghosh, Pranav Narayanan Venkit, Sanjana Gautam, Shomir Wilson, and Aylin Caliskan. 2024. Do generative AI models output harm while representing non-Western cultures: Evidence from a community-centered approach. InAIES, Vol. 7. 476–489

work page 2024
[42]

Leander Girrbach, Stephan Alaniz, Genevieve Smith, and Zeynep Akata. 2025. A Large Scale Analysis of Gender Biases in Text-to-Image Generative Models.arXiv preprint arXiv:2503.23398(2025)

work page arXiv 2025
[43]

Douglas Guilbeault, Solène Delecourt, Tasker Hull, Bhargav Srinivasa Desikan, Mark Chu, and Ethan Nadler. 2024. Online images amplify gender bias.Nature626, 8001 (2024), 1049–1055

work page 2024
[44]

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep residual learning for image recognition. InCVPR

work page 2016
[45]

Jack Hessel, Ari Holtzman, Maxwell Forbes, Ronan Le Bras, and Yejin Choi. 2021. CLIPScore: A Reference-free Evaluation Metric for Image Captioning. InEMNLP

work page 2021
[46]

Zhang, Qingwen Bu, Xiaofei Xie, Junjie Chen, and Heming Cui

Dong Huang, Jie M. Zhang, Qingwen Bu, Xiaofei Xie, Junjie Chen, and Heming Cui. 2025. Bias testing and mitigation in llm-based code generation.ACM Transactions on Software Engineering and Methodology35, 1 (2025), 1–31

work page 2025
[47]

2018.ISO 19011:2018 Guidelines for auditing management systems(3 ed.)

International Organization for Standardization. 2018.ISO 19011:2018 Guidelines for auditing management systems(3 ed.). Technical Report. ISO. https://www.iso.org/standard/70017.html Third edition 2018-07; Accessed: 2025-12-17

work page 2018
[48]

2021.Bias in AI systems and AI aided decision making

International Organization for Standardization and International Electrotechnical Commission. 2021.Bias in AI systems and AI aided decision making. Technical Report. ISO/IEC. https://www.iso.org/standard/77607.html

work page 2021
[49]

2021.Information technology — Artificial intelligence (AI) — Bias in AI systems and AI aided decision making

ISO/IEC. 2021.Information technology — Artificial intelligence (AI) — Bias in AI systems and AI aided decision making. Technical Report ISO/IEC TR 24027:2021. International Organization for Standardization. https://www.iso.org/standard/77607.html

work page 2021
[50]

2023.ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system

ISO/IEC. 2023.ISO/IEC 42001:2023 Information technology — Artificial intelligence — Management system. Technical Report ISO/IEC 42001:2023. International Organization for Standardization. https://www.iso.org/standard/42001

work page 2023
[51]

Joint Research Centre. 2025. Auditing the Algorithms: New JRC review categorises risk investigations methodologies through the lens of the EU Digital Services Act. https://algorithmic-transparency.ec.europa.eu/news/auditing-algorithms-new-jrc-review-categorises- risk-investigations-methodologies-through-lens-eu-2025-06-30_en

work page 2025
[52]

Sangwon Jung, Alex Oesterling, Claudio Mayrink Verdun, Sajani Vithana, Taesup Moon, and Flavio P Calmon. 2025. Multi-Group Proportional Representations for Text-to-Image Models. InCVPR

work page 2025
[53]

Mintong Kang, Vinayshekhar Bannihatti Kumar, Shamik Roy, Abhishek Kumar, Sopan Khosla, Balakrishnan Murali Narayanaswamy, and Rashmi Gangadharaiah. 2025. FairGen: Controlling sensitive attributes for fair generations in diffusion models via adaptive latent guidance. InProceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. ...

work page 2025
[54]

Kimmo Karkkainen and Jungseock Joo. 2021. FairFace: Face attribute dataset for balanced race, gender, and age for bias measurement and mitigation. InW ACV

work page 2021
[55]

Amelia Katirai, Noa Garcia, Kazuki Ide, Yuta Nakashima, and Atsuo Kishimoto. 2024. Situating the social issues of image generation models in the model life cycle: a sociotechnical approach.AI and Ethics(2024)

work page 2024
[56]

Eunji Kim, Siwon Kim, Minjun Park, Rahim Entezari, and Sungroh Yoon. 2025. Rethinking Training for De-biasing Text-to-Image Generation: Unlocking the Potential of Stable Diffusion. InCVPR. 17

work page 2025
[57]

Hyung-Kwon Ko, Gwanmo Park, Hyeon Jeon, Jaemin Jo, Juho Kim, and Jinwook Seo. 2023. Large-scale text-to-image generation models for visual artists’ creative works. InProceedings of the 28th international conference on intelligent user interfaces. 919–933

work page 2023
[58]

Gant Laborde. [n. d.]. Deep NN for NSFW Detection. https://github.com/GantMan/nsfw_model

work page
[59]

Tony Lee, Michihiro Yasunaga, Chenlin Meng, Yifan Mai, Joon Sung Park, Agrim Gupta, Yunzhi Zhang, Deepak Narayanan, Han- nah Benita Teufel, Marco Bellagente, et al. 2023. Holistic Evaluation of Text-to-Image Models. InNeurIPS Datasets and Benchmarks Track

work page 2023
[60]

Hang Li, Chengzhi Shen, Philip Torr, Volker Tresp, and Jindong Gu. 2024. Self-discovering interpretable diffusion latent directions for responsible text-to-image generation. InCVPR

work page 2024
[61]

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. InICML

work page 2023
[62]

Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. InICML

work page 2022
[63]

Lijun Li, Zhelun Shi, Xuhao Hu, Bowen Dong, Yiran Qin, Xihui Liu, Lu Sheng, and Jing Shao. 2025. T2Isafety: Benchmark for assessing fairness, toxicity, and privacy in image generation. InCVPR

work page 2025
[64]

Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. 2024. Improved baselines with visual instruction tuning. InCVPR

work page 2024
[65]

Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. 2023. Visual instruction tuning. InNeurIPS

work page 2023
[66]

Zhixuan Liu, Peter Schaldenbrand, Beverley-Claire Okogwu, Wenxuan Peng, Youngsik Yun, Andrew Hundt, Jihie Kim, and Jean Oh

work page
[67]

InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

SCoFT: Self-contrastive fine-tuning for equitable image generation. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10822–10832

work page
[68]

Alexandra Sasha Luccioni, Christopher Akiki, Margaret Mitchell, and Yacine Jernite. 2023. Stable bias: Analyzing societal representations in diffusion models. InNeurIPS

work page 2023
[69]

Jose Luna, Ivan Tan, Xiaofei Xie, and Lingxiao Jiang. 2024. Navigating governance paradigms: A cross-regional comparative study of generative ai governance processes & principles. InAIES, Vol. 7. 917–931

work page 2024
[70]

Fabian Lütz. 2024. The AI Act, gender equality and non-discrimination: what role for the AI office?. InERA Forum, Vol. 25. Springer, 79–95

work page 2024
[71]

Midjourney

Inc. Midjourney. 2025. Midjourney. https://www.midjourney.com/

work page 2025
[72]

Margaret Mitchell, Simone Wu, Andrew Zaldivar, Parker Barnes, Lucy Vasserman, Ben Hutchinson, Elena Spitzer, Inioluwa Deborah Raji, and Timnit Gebru. 2019. Model cards for model reporting. InProceedings of the conference on fairness, accountability, and transparency. 220–229

work page 2019
[73]

Ranjita Naik and Besmira Nushi. 2023. Social Biases through the Text-to-Image Generation Lens. InAIES

work page 2023
[74]

National Institute of Standards and Technology. 2023. Artificial Intelligence Risk Management Framework (AI RMF 1.0). https: //www.nist.gov/itl/ai-risk-management-framework

work page 2023
[75]

2024.How trans identification can dramatically skew crime statistics

Joe Neil. 2024.How trans identification can dramatically skew crime statistics. https://www.telegraph.co.uk/news/2024/03/08/trans- identification-skew-crime-statistics/ Accessed: 2026-01-03

work page 2024
[76]

Sheilla Njoto, Marc Cheong, Reeva Lederman, Aidan McLoughney, Leah Ruppanner, and Anthony Wirth. 2022. Gender bias in AI recruitment systems: A sociological-and data science-based case study. In2022 IEEE international symposium on technology and society (ISTAS), Vol. 1. IEEE, 1–7

work page 2022
[77]

OECD. 2024. Respect for the rule of law, human rights and democratic values, including fairness and privacy (Principle 1.2). https://oecd.ai/en/dashboards/ai-principles/P6

work page 2024
[78]

Fabian Offert and Thao Phan. 2022. A sign that spells: DALL-E 2, invisual images and the racial politics of feature space.arXiv preprint arXiv:2211.06323(2022)

work page arXiv 2022
[79]

Office of Management and Budget. 2024. M-24-10: Advancing Governance, Innovation, and Risk Management for Agency Use of Artificial Intelligence. https://bidenwhitehouse.archives.gov/wp-content/uploads/2024/03/M-24-10-Advancing-Governance-Innovation-and- Risk-Management-for-Agency-Use-of-Artificial-Intelligence.pdf

work page 2024
[80]

Ninell Oldenburg and Gleb Papyshev. 2025. The Stories We Govern By: AI, Risk, and the Power of Imaginaries. InAIES, Vol. 8. 1939–1950

work page 2025

Showing first 80 references.