Persistent Anti-Muslim Bias in Large Language Models
Open accessThe DOI matches its Crossref record; 13 of the 13 references it cites were found in a registry or an index.
Abstract
It has been observed that large-scale language models capture undesirable societal biases, e.g. relating to race and gender; yet religious bias has been relatively unexplored. We demonstrate that GPT-3, a state-of-the-art contextual language model, captures persistent Muslim-violence bias. We probe GPT-3 in various ways, including prompt completion, analogical reasoning, and story generation, to understand this anti-Muslim bias, demonstrating that it appears consistently and creatively in different uses of the model and that it is severe even compared to biases about other religious groups. For instance, Muslim is analogized to terrorist in 23% of test cases, while Jewish is mapped to its most common stereotype, money, in 5% of test cases. We quantify the positive distraction needed to overcome this bias with adversarial text prompts, and find that use of the most positive 6 adjectives reduces violent completions for Muslims from 66% to 20%, but which is still higher than for other religious groups.
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.
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
The source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“Pre-trained on massive amounts of text on the Internet, LMs are known to produce linguistically fawed, factually incorrect, or even ethically problematic text at times [1,24].”
From CoAuthor: Designing a Human-AI Collaborative Writing Dataset for Exploring Language Model Capabilities (2022) · cites this paper as [1]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“LLMs and other AI tools may have a number of biases that could arise from various factors, such as unbalanced training datasets74–77.”
From On opportunities and challenges of large multimodal foundation models in education (2025) · cites this paper as [75]
Experimental — accuracy not yet measuredThe source backs part of the claim, not all of it — or backs it for a different population, condition or comparison.
In the citing paper
“it is also a liability; all AI systems might inherit the same problematic biases of a few foundation models [Bolukbasi et al. 2016; Caliskan et al. 2017; Abid et al. 2021, inter alia]”
From On the Opportunities and Risks of Foundation Models (2021) · cites this paper as Abid et al. 2021
Experimental — accuracy not yet measuredThe cited source says what the claim attributes to it.
In the citing paper
“Previous work has shown that language models tend to exhibit human-like biases with respect to specific protected groups such as gender, ethnicity or religious orientation3–6.”
From Generative language models exhibit social identity biases (2024) · cites this paper as [4]
In the citing paper
“If a medical LLM has been trained on data that contains misinformation or biased content, LLMs deployed for clinical decision-making are at risk of reproducing these problematic contents [120,124–126].”
From Bias in medical AI: Implications for clinical decision-making (2024) · cites this paper as [126]
1 more claim about this paper
Create a free account to see every claim other papers make about this one.
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
Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2017) · arXiv · doi:10.48550/arxiv.1706.03762 · Cited by 120,404 · cited 1×
“These languagemodels have evolved, from learning individual word vectors with single-layer models [8], to more complex language generation architectures such as recurrent neural networks [5] and most recently transformers [4, 6, 14].”
BackgroundEfficient 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 · cited 1×
“These languagemodels have evolved, from learning individual word vectors with single-layer models [8], to more complex language generation architectures such as recurrent neural networks [5] and most recently transformers [4, 6, 14].”
BackgroundLanguage Models are Few-Shot Learners
Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. (2020) · arXiv · doi:10.48550/arxiv.2005.14165 · Cited by 23,372 · cited 4×
“These languagemodels have evolved, from learning individual word vectors with single-layer models [8], to more complex language generation architectures such as recurrent neural networks [5] and most recently transformers [4, 6, 14].”
BackgroundMethodInfluentialSelf-Critical Sequence Training for Image Captioning
Rennie, S. J., Marcheret, E., Mroueh, Y., Ross, J., & Goel, V. (2017) · 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) · doi:10.1109/cvpr.2017.131 · Cited by 2,112 · cited 1×
“These descriptions are generated by passing images through a pretrained image captioning model [11], and fine-tuning GPT-3 to transform the short captions into long descriptions, as shown in Fig.”
MethodReformer: The Efficient Transformer
Kitaev, N., Kaiser, Ł., & Levskaya, A. (2020) · arXiv · doi:10.48550/arxiv.2001.04451 · Cited by 1,435 · cited 1×
“These languagemodels have evolved, from learning individual word vectors with single-layer models [8], to more complex language generation architectures such as recurrent neural networks [5] and most recently transformers [4, 6, 14].”
BackgroundSemi-supervised Sequence Learning
Dai, A. M., & Le, Q. V. (2015) · arXiv · doi:10.48550/arxiv.1511.01432 · Cited by 994
StereoSet: Measuring stereotypical bias in pretrained language models
Nadeem, M., Bethke, A., & Reddy, S. (2020) · arXiv · doi:10.48550/arxiv.2004.09456 · Cited by 643 · cited 2×
“As a result, even though the various language models have different architectures, since they are trained on similar corpora of text, they often learn similar biases [9].”
BackgroundUniversal Adversarial Triggers for Attacking and Analyzing NLP
Wallace, E., Feng, S., Kandpal, N., Gardner, M., & Singh, S. (2019) · Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) · doi:10.18653/v1/d19-1221 · Cited by 642
Identifying and Reducing Gender Bias in Word-Level Language Models
Bordia, S., & Bowman, S. R. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Student Research Workshop · doi:10.18653/v1/n19-3002 · Cited by 219 · cited 1×
“Previous work has explored the biases that are ultimately learned by language models, with the primary focus being racial and gender stereotypes [2, 3, 7, 12].”
BackgroundThe Woman Worked as a Babysitter: On Biases in Language GenerationPreprint
Sheng, E., Chang, K.-W., Natarajan, P., & Peng, N. (2019) · arXiv (Cornell University) · doi:10.48550/arxiv.1909.01326 · Cited by 205 · cited 2×
“In [12], the authors made similar observations about language generation models.”
Background
Research works printed but not linked: 6
- 7. T. Bolukbasi, K.-W. Chang, J. Y. Zou, V. Saligrama, A. T. Kalai, Advances in neural information processing systems 29 , 4349 (2016).
- 9. S. Bordia, S. R. Bowman, arXiv preprint arXiv:1904.03035 (2019). · arXiv
- 2. A. M. Dai, Q. V. Le, Advances in neural information processing systems 28 , 3079 (2015).
- 10. K. Lu, P. Mardziel, F. Wu, P. Amancharla, A. Datta, Logic, Language, and Security (Springer, 2020), pp. 189-202.
- 3. A. Vaswani, et al. , Advances in neural information processing systems (2017), pp. 59986008.
- 14. E. Wallace, S. Feng, N. Kandpal, M. Gardner, S. Singh, arXiv preprint arXiv:1908.07125 (2019). · arXiv
Cited by
Items 701–800 of 986
Geise, S., & Xu, Y. (2026) · Digital Journalism · doi:10.1080/21670811.2026.2700594 · Cited by 2
“As shown here and elsewhere, generative systems tend to reproduce iconic templates and stylistic ste-reotypes rooted in the training data instead of context-rich or metaphorical imagery (Abid, Farooqi, and Zou 2021; Olier and Spadavecchia 2022; Thomson 2025).”
Exploring occupational prestige through large language models: A multi-dimensional approach
de Vries, R., Hill, M., & Ruis, L. (2026) · Research in Social Stratification and Mobility · doi:10.1016/j.rssm.2026.101185 · Cited by 2
Ethical AI: Designing Responsible And Trustworthy Systems
Mana, S. C., Kalaiarasi, G., Selvi, M., Velvizhi, R., & Nair, M. C. (2023) · Jupiter Publications Consortium eBooks · doi:10.47715/jpc.978-93-91303-66-2 · Cited by 2
Silvestro, J. J. (2023) · College English · doi:10.58680/ce202332661 · Cited by 2
Making the Case for Establishing a South African AI Innovation Hub
Althoff-Thomson, S., & Belle, J.-P. V. (2024) · Lecture notes in networks and systems · doi:10.1007/978-3-031-54235-0_21 · Cited by 2
Evaluating Stereotypical Biases and Implications for Fairness in Large Language Models
Cao, C., & Bandara, D. (2024) · doi:10.18260/1-2--45767 · Cited by 2
It is Time to Develop an Auditing Framework to Promote Value Aware Chatbots
Wang, Y., & Singh, L. (2024) · International Conference on Data Science, Technology and Applications (DATA) · doi:10.5220/0012806800003756 · Cited by 2
A Method for Implicit Risk Detection and Evaluation of Chinese LLMs through TextCNN
Liu, Z., Zhang, X., Ge, J., Li, G., Han, J., & Shi, B. (2024) · Asia-Pacific Conference on Image Processing, Electronics and Computers (IPEC) · doi:10.1109/ipec61310.2024.00083 · Cited by 2
Corpus Development Based on Conflict Structures in the Security Field and LLM Bias Verification
Inoshita, K. (2024) · International Conference on Natural Language Processing for Digital Humanities (NLP4DH) · doi:10.18653/v1/2024.nlp4dh-1.49 · Cited by 2
“However, LLMs may inherit biases from their training data, reflecting prejudices related to race, gender, religion, and nationality (Abid et al., 2021; Venkit et al., 2023).”
BackgroundPujari, T. D., Goel, A., & Sharma, A. (2024) · International Journal Science and Technology · doi:10.56127/ijst.v3i3.1968 · Cited by 2
A Hypothesis-Driven Framework for the Analysis of Self-Rationalising Models
Braun, M., & Kunz, J. (2024) · Conference of the European Chapter of the Association for Computational Linguistics: Student Research Workshop (EACL SRW) · doi:10.18653/v1/2024.eacl-srw.11 · Cited by 2
Brink, L., & Grittmann, E. (2025) · Medien & Kommunikationswissenschaft · doi:10.5771/1615-634x-2025-3-442 · Cited by 2
Clustering Discourses: Racial Biases in Short Stories about Women Generated by Large Language Models
Bonil, G., Gondim, J., dos Santos, M., Hashiguti, S. T., Maia, H., da Silva, N. F. F., Pedrini, H., & Avila, S. (2025) · Simpósio Brasileiro de Tecnologia da Informação e da Linguagem Humana (STIL) · doi:10.5753/stil.2025.37814 · Cited by 2
“For instance, [Abid et al. 2021] and [Venkit et al. 2023] contend that certain models generate negative outputs for underrepresented nationalities , while [Bordia and Bowman 2019] argue gender bias can be documented as operating in a lexical dimension, and [Lucy and Bamman 2021]demonstrate that…”
From n-gram to Attention: How Model Architectures Learn and Propagate Bias in Language Modeling
Kabir, M., Tahsin, T., & Ananiadou, S. (2025) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2025.findings-emnlp.1003 · Cited by 2
“…of training data (heterogeneous data) create substantial pathways for social biases to manifest in language models, whether in relatively simple word embed-dings (Bolukbasi et al., 2016; Basta et al., 2019) or in far more complex large language models (Abid et al., 2021; Abrar et al., 2025).”
2025 年にむけた生成AI への大学の対応 : 国立大学指針の分析と情報セキュリティの標示方法の提案
福島, 宙. (2024) · 大學教育研究 · doi:10.24546/0100488362 · Cited by 2
Maj, A. (2023) · Ethos Kwartalnik Instytutu Jana Pawła II KUL · doi:10.12887/36-2023-3-143-11 · Cited by 2
Apolitical Intelligence? Auditing Delphi's responses on controversial political issues in the USPreprint
Rystrøm, J. H. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2306.13000 · Cited by 2
“In addition, extensive case studies have been conducted to examine different types of bias, such as those within gender [50], anti-Muslim sentiment [1], and disabilities [39].”
BackgroundTIDE: Textual Identity Detection for Evaluating and Augmenting Classification and Language ModelsPreprint
Klu, E., & Sethi, S. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2309.04027 · Cited by 2
“When models contain biases, they can play an active role in perpetuating societal inequities and unfair outcomes for under-represented groups (Sweeney, 2013; Abid et al., 2021).”
BackgroundThe Ethical Concerns Surrounding GPT in EducationPreprint
Tajik, E., & Alijani, F. (2024) · SSRN Electronic Journal · doi:10.36227/techrxiv.170421413.32906559/v1 · Cited by 2
Dewitte, P. (2024) · SSRN Electronic Journal · doi:10.2139/ssrn.4773907 · Cited by 2
Aligning language models with human preferencesPreprint
Korbak, T. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2404.12150 · Cited by 2
Toward Efficient Influence Function: Dropout as a Compression Tool
Zhang, Y., & Amiri, M. M. (2025) · Transactions on machine learning research · doi:10.48550/arxiv.2509.15651 · Cited by 2
“Despite their success, even the state-of-the-art models face persistent challenges, including hallucination (Lin et al., 2021; Huang et al., 2025) and generating toxic or biased content (Abid et al., 2021; Wang et al., 2023).”
nDNA -- the Semantic Helix of Artificial CognitionPreprint
Das, A. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2509.18216 · Cited by 2
“…aligned (Bai et al., 2022), merged (Ilharco et al., 2023), distilled (Mirzadeh et al., 2020), and deployed across diverse cultural and linguistic domains (Abid et al., 2021; Arora et al., 2023), we lack a principled framework to discern what is preserved, what mutates, and what is silently erased.”
Human Preferences for Constructive Interactions in Language Model AlignmentPreprint
Kyrychenko, Y., Roozenbeek, J., Davidson, B., van der Linden, S., & Debnath, R. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2503.16480 · Cited by 2
Affective Polarization Amongst Swedish PoliticiansPreprint
t'Serstevens, F., Cerina, R., & Peper, G. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2503.16193 · Cited by 2
“While LLMs offer powerful tools for political analysis (T¨ornberg, 2024), they are not neutral observers; research has shown that their political classifications often reflect biases inherent in their training data or reinforcement learning processes (Abid et al., 2021, Hartmann et al., 2023).”
Biased by Design: Leveraging AI Biases to Enhance Critical Thinking of News ReadersPreprint
Zavolokina, L., Sprenkamp, K., Katashinskaya, Z., & Jones, D. G. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2504.14522 · Cited by 2
Addressing Misalignment in Language Model Deployments through Context-Specific Evaluations
Soni, P. (2024) · DSpace@MIT (Massachusetts Institute of Technology) · Cited by 2
Banyas, P., Sharma, S., Simmons, A., & Vispute, A. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2510.13852 · Cited by 2
“Although it can be argued that LLMs will never be ideologically neutral because they are trained on information produced by biased humans [1, 43, 46], a factual consistency benchmark could be an indicator of reduced ideological partiality.”
Hamer, J., Kool, L., Hijstek, B., van Eeden, Q., & Das, D. (2023) · KNAW Research Portal (The Royal Netherlands Academy of Arts and Sciences) · Cited by 2
Adversarial Alignment: Ensuring Value Consistency in Large Language Models for Sensitive DomainsPreprint
Gao, Y., Liu, Z., Yao, X., Chen, B., & Zhao, X. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2601.13137 · Cited by 2
“For example, Abid et al. [9] find that GPT-3 demonstrates significant anti-Muslim bias, while Salecha et al. [16] show that LLMs tend to overrepresent overly positive personality traits when simulating human characteristics.”
Style Transfer as Bias Mitigation: Diffusion Models for Synthetic Mental Health Text for ArabicPreprint
Mankarious, S., & Zirikly, A. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2601.14124 · Cited by 2
“Additional work demonstrates that LLMs frequently portray women as less powerful than men and associate Muslim groups with violence Abid et al. (2021); Lucy and Bamman (2021).”
YAQIN: Culturally Sensitive, Agentic AI for Mental Healthcare Support Among Muslim Women in the UKPreprint
Zaraket, Y., & Mougenot, C. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.07709 · Cited by 2
“General-purpose AI models are predominantly trained on Western datasets, limiting their cultural nuance [76, 77], RAG enables YAQIN to respond from a curated, context-specific knowledge base.”
ProText: A benchmark dataset for measuring (mis)gendering in long-form textsPreprint
Kotek, H., Bowler, M., Sonnenberg, P., & Yang, Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2603.27838 · Cited by 2
“Bias extends beyond gender to other social categories such as religion, race, nationality, disability, and occupation, as well [1, 16, 18, 24, 36, 37, 42, among many others].”
Meta-Evaluation Methodology and Benchmark for Automatic Story Generation
Chhun, C. (2024) · theses.fr (ABES) · doi:10.70675/ec296eb8z87a1z4fdbzbbbdz12bb27449b8f · Cited by 2
Keough, P. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.17359 · Cited by 2
Do Models Share Safety Representations? Cross-Model Steering for Safe Visual GenerationPreprint
Poppi, T., Cappelletti, S., Sarto, S., Schiffers, F., Kessler, G. N., Cornia, M., Baraldi, L., & Cucchiara, R. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2606.05290 · Cited by 2
“…text-to-image systems [14, 26, 44], and video generators [51] are trained on massive, only partially curated corpora, and inevitably internalize unsafe or policy-sensitive concepts, including sexual content, hate, violence, illegal activity, and copyrighted or trademarked entities [1, 8, 9, 52].”
Boullineau, E., & Arciniegas, J. D. M. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7369759 · Cited by 2
Adaptive Deployment of Untrusted LLMs Reduces Distributed Threats
Wen, J., Hebbar, V., Larson, C., Bhatt, A., Radhakrishnan, A., Sharma, M., Sleight, H., Feng, S., He, H., Perez, E., et al. (2024) · arXiv · doi:10.48550/arxiv.2411.17693 · Cited by 1
“LLMs have been shown to exhibit various safety risks, such as generating offensive content (Gehman et al., 2020) and reinforcing unfairness (Abid et al., 2021).”
Robustly Improving LLM Fairness in Realistic Settings via Interpretability
Karvonen, A., & Marks, S. (2025) · arXiv · doi:10.48550/arxiv.2506.10922 · Cited by 1
“LLMs exhibit similar issues, where studies have uncovered gender bias (Nangia et al., 2020; Vig et al., 2020), religious bias (Abid et al., 2021), and ethnic bias (Ahn & Oh, 2021).”
Aus Versehen verzerrt: Wie identifiziert man einen Bias in Machine Learning Modellen?
Pohlink, C., & Fischer, S. (2021) · Springer eBooks · doi:10.1007/978-3-658-34670-6_4 · Cited by 1
Between This and That is It: Embodied Semantic Space at the Edge
Rayzhekov, A., & Murer, M. (2024) · International Conference on Tangible, Embedded, and Embodied Interaction (TEI) · doi:10.1145/3623509.3635326 · Cited by 1
“Therefore, they naturally inherit all built-in biases, stereotypes and ideological beliefs [1, 22].”
BackgroundDeceptively simple: An outsider's perspective on natural language processing
KhudaBukhsh, A. R. (2024) · AI Magazine · doi:10.1002/aaai.12204 · Cited by 1
Chistruga, A. (2024) · ANALELE ȘTIINŢIFICE ALE UNIVERSITĂŢII „ALEXANDRU IOAN CUZA” DIN IAȘI (SERIE NOUĂ) ȘTIINŢE JURIDICE · doi:10.47743/jss-2024-70-2-3 · Cited by 2
Kern, R. (2025) · Cambridge University Press eBooks · doi:10.1017/9781009294850.025 · Cited by 1
Biasbarrier a Fairness and Equity Filter for LLM Responses Under Algorithmic Accountability Acts
Chinnachamy, T. (2025) · International Journal of Scientific Research and Modern Technology. · doi:10.38124/ijsrmt.v4i8.754 · Cited by 1
The potential benefits of establishing an ai-arbitration legal framework in Malaysia
Labanieh, M. F., Alshamsi, A. A. S. S., Ayub, Z. A., & Wahab, H. A. (2025) · Multidisciplinary Science Journal · doi:10.31893/multiscience.2026171 · Cited by 1
Fereidooni, S. (2025) · Proceedings of the AAAI/ACM Conference on AI Ethics and Society · doi:10.1609/aies.v8i3.36773 · Cited by 1
SESGO: Spanish Evaluation of Stereotypical Generative Outputs
Robles, M., Bernal, C., Raigoso, D., & Rubio, M. D. (2025) · Proceedings of the AAAI/ACM Conference on AI Ethics and Society · doi:10.1609/aies.v8i3.36707 · Cited by 1
Unmaking AI: A Framework for Critical Investigation
Munn, L., Magee, L., Arora, V., & Khan, A. H. (2025) · Critical AI · doi:10.1215/2834703x-12095973 · Cited by 1
Conclusion: Faith, Identity, and Power in Networked Societies
Yılmaz, İ. (2026) · Springer eBooks · doi:10.1007/978-981-95-4660-2_8 · Cited by 1
Laywine, N., & Simon, V. (2026) · Springer eBooks · doi:10.1007/978-3-032-05155-4_12 · Cited by 1
Do Large Language Models Understand Mansplaining? Well, Actually...
Perez-Almendros, C., & Camacho-Collados, J. (2024) · Language Resources and Evaluation Conference (LREC) · doi:10.63317/35kuk7bfhi7t · Cited by 1
“On the one hand, LLMs, as well as their predecessors, contain biases, as they are trained on vast amounts of (biased) data online (Abid et al., 2021; Schramowski et al., 2022; Gallegos et al., 2023).”
BackgroundDutta, A., Fayyazi, R., Yang, S. J., & KhudaBukhsh, A. R. (2026) · Proceedings of the AAAI Conference on Artificial Intelligence · doi:10.1609/aaai.v40i45.41181 · Cited by 2
Council of LLMs: Evaluating Capability of Large Language Models to Annotate Propaganda
Sharma, V., Jain, S., Shokri, M., Levitan, S. I., & Filatova, E. (2026) · Workshop on Computational Approaches to Subjectivity, Sentiment and Social Media Analysis (WASSA) · doi:10.18653/v1/2026.wassa-1.1 · Cited by 1
“Societal risks persist: removing human annotators can erase interpretive insight and reinforce systemic bias (Das et al., 2024; Abid et al., 2021; Wang et al., 2023), leading to unfair outcomes and broader disparities (Dillion et al., 2023).”
Kayani, A. I., Anwar, B., Brashi, A., El‐Dakhs, D. A. S., & Younas, M. S. (2026) · Asian Women · doi:10.64446/aw.2026.3.42.1.41 · Cited by 1
Jeon, S., Lee, M. H. J., Montgomery, J., & Lai, C. K. (2026) · Political Analysis · doi:10.1017/pan.2026.10038 · Cited by 1
“Extensive research documents that FMs perpetuate gender (e.g., Hofmann and Mõttus 2025; Rudinger et al. 2018; Zhao et al. 2018) and group-based stereotypes (e.g., Abid, Farooqi, and Zou 2021; Lee, Montgomery, and Lai 2024; Lucy and Bamman 2021).”
What Stays and What Goes: Auditing the Impact of LLM Summarization on News Partisanship
Savgira, P., Kreiss, E., & Hosseinmardi, H. (2026) · CHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA) · doi:10.1145/3772363.3799057 · Cited by 1
“Instead, they come to actively shape discourse and risk amplifying social biases that are already prevalent in LLMs [1, 2, 5, 16, 17, 21].”
Challenges and Implications of Generative AI for Anti-Discrimination Law and Policy
Towfigh, E. V. (2025) · German Law Journal · doi:10.1017/glj.2026.10189 · Cited by 1
ALTAY, R. (2026) · Etkileşim · doi:10.32739/etkilesim.2026.9.17.345 · Cited by 1
Hegberg, A., Setor, T. K., & Joseph, D. (2026) · ACM SIGCPR Conference on Computer Personnel Research (SIGCPR) · doi:10.1145/3768310.3807790 · Cited by 1
“This means scanning demographic narratives for recurring scripts and revising personas or regenerating them with counter-stereotypical constraints so that requirements are not anchored to a narrow social storyline (Abid et al., 2021; Blodgett et al., 2020; Gallegos et al., 2024).”
Samad, H., Lam, T., Mügge-Durum, C., & Akinwumi, M. (2026) · Proceedings of the AAAI/ACM Conference on AI Ethics and Society · doi:10.1609/aies.v9i4.43303 · Cited by 1
Do Language Models Pass the Bechdel Test? Auditing Gender Biases in LLM-Generated Screenplays
Govindu, M. N., Wang, S., Friedler, S. A., & Metaxa, D. (2026) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3805689.3812208 · Cited by 1
Reflexive Research with LLMs: Considering the Positionality of Users and Systems
Smith, E. L. T., da Costa, L. M., & Fokkens, A. S. (2026) · Language Resources and Evaluation Conference (LREC) · doi:10.63317/3ycn8grfuwtx · Cited by 1
“Abid et al. (2021) found strong and consistent associations between Muslims and violence in the representations of GPT-3.”
Toward Dialect-Aware Safety Evaluation for Arabic Large Language Models
Zaghouani, W. (2026) · Workshop on Trustworthy NLP (TrustNLP) · doi:10.18653/v1/2026.trustnlp-main.37 · Cited by 1
“…provide principled diagnostic tools: Dwork et al. (2012) introduce fairness through awareness, while Barocas and Selbst (2016) (Davidson et al., 2017; Waseem and Hovy, 2016) and that biased annotation data replicates harm against the communities systems purport to protect (Abid et al., 2021).”
Decolonizing the literary AI in the age of LLMs and digital neocolonialism
Shabanpour, M. B. (2026) · PhilPapers (PhilPapers Foundation) · Cited by 1
Chen, Z., Jiao, W.-C., Tsang, Y. P., Wu, C. H., & Wang, X.-J. (2026) · Group Decision and Negotiation · doi:10.1007/s10726-026-10017-w · Cited by 1
“While nascent research primarily addressed sociodemographic prejudices, such as gender and racial bias (Abid et al. 2021).”
McMaster, N., Morrison, R., Jordan, R., & O’Chin, H. (2026) · Australasian Journal of Technology Education · doi:10.15663/ajte.v11.122 · Cited by 1
Goudarzi, A., & Zonouz, A. N. (2026) · Digital studies in language and literature · doi:10.1515/dsll-2026-0052 · Cited by 1
Hayoune, K. (2026) · Islamic Reports · doi:10.66638/fsavt586 · Cited by 1
“For Muslim learners the bias can be pointed: Abid et al. (2021) showed that GPT-3 captured a persistent bias linking Muslims to violence, including analogies that paired the word Muslim with terrorist in 23 per cent of test cases.”
Zhang, Y. (2026) · Topoi · doi:10.1007/s11245-026-10504-4 · Cited by 1
Epistemic Framing Tilt in LLMs: Auditing Religious Discourse Across Traditions
Buzelin, A., Aquino, Y., Bento, P., Chagas, A., Dayrell, L., Pappa, G. L., Almeida, V., & Meira, W., Jr. (2026) · Proceedings of the AAAI/ACM Conference on AI Ethics and Society · doi:10.1609/aies.v9i1.43164 · Cited by 1
Large language models for computational social science
Xu, F., Shao, C., Chen, L., Zeng, Q., Chen, Z., Sukiennik, N., Wang, J., Gao, C., Wang, H., Lian, J., et al. (2026) · Nature Human Behaviour · doi:10.1038/s41562-026-02539-9 · Cited by 1
News Consumption Helps Readers Identify Model-Generated News
Kothari, A., Orama, A., Miller, R. L., Peeks, M., Bailey, R., & Alm, C. O. (2023) · IEEE Western New York Image and Signal Processing Workshop (WNYISPW) · doi:10.1109/wnyispw60588.2023.10349588 · Cited by 1
“In addition, persistent biases are encountered in text generation models [4], [6]–[8] and there is potential for malicious use to misinform readers [4], [7], [9], [10].”
BackgroundRibeiro, T. M. R., Siqueira, S. W. M., & de Bayser, M. G. (2024) · Simpósio Brasileiro de Sistemas Colaborativos (SBSC) · doi:10.5753/sbsc.2024.238053 · Cited by 1
The growth outlook of the ICT sector
OECD (2024) · OECD Digital Economy Outlook · doi:10.1787/8635f10a-en · Cited by 1
Improving Rumor Detection Performance by Using Bias Attributes
Shi, D., & Zeng, J. (2025) · IEEE Transactions on Computational Social Systems · doi:10.1109/tcss.2025.3550170 · Cited by 1
“For example, Abid et al. demonstrated that GPT-3 is more likely to generate violent language when prompted with the word “Muslim” [30].”
Distribution Learning with Valid Outputs Beyond the Worst-Case
Rittler, N., & Chaudhuri, K. (2024) · neural information processing systems · doi:10.52202/079017-0719 · Cited by 1
Die (Un)gerechten : ein kritischer Beitrag zu ethischen Risiken von Large Language Models
Felser, J., & Spranger, M. (2025) · MOnAMi (Hochschule Mittweida Hochschulbibliothek) · doi:10.48446/opus-15908 · Cited by 1
The Rise of AI Avatars: Legal Personhood, Rights and Liabilities in an Evolving Metaverse
Cheong, B. C. (2024) · Apollo (University of Cambridge) · doi:10.17863/cam.109175 · Cited by 1
“Similarly, an AI avatar programmed with biased assumptions about gender roles or cultural norms may produce creative works that reinforce these prejudices and limit the diversity of perspectives represented in the metaverse (Abid et al., 2021).”
Multi-Dimensional Bias Analysis in LLMs Using Hierarchical and Interaction Models
Syed, B., Charlebois, D. A., Ezzati‐Jivan, N., Tahmooresnejad, L., & Ayanso, A. (2025) · AIS Electronic Library (AISeL) · Cited by 1
Red teaming with faith leaders: expanding digital safety and accountability to frontiers of care
Lutz, N., & Weyl, E. G. (2026) · ACM Conference on Fairness, Accountability, and Transparency (FAccT) · doi:10.1145/3805689.3812362 · Cited by 1
“Some audit work has contended with religion, finding persistent religious bias in LLMs [2, 61], reflecting a broader WEIRD framing in AI fairness [57].”
Taught by the Internet, Exploring Bias in OpenAIs GPT3Preprint
Ayaz, A. S., Nawalgaria, A., & Yin, R. (2023) · arXiv (Cornell University) · doi:10.48550/arxiv.2306.02428 · Cited by 1
Takács, D. (2023) · Mesterséges intelligencia · doi:10.35406/mestint · Cited by 1
Measuring the harms of personalization through advertising
Ali, M. (2023) · Digital Repository Service (Northeastern University) · doi:10.17760/d20620606 · Cited by 1
Loquacity and Visible Emotion: ChatGPT as a Policy AdvisorPreprint
Biancotti, C., & Camassa, C. (2024) · SSRN Electronic Journal · doi:10.2139/ssrn.4849363 · Cited by 1
Studies in conversational AI : multilingual capabilities, world knowledge, and evaluation strategies
Bruyn, M. D. (2024) · doi:10.63028/10067/2023030151162165141 · Cited by 1
Fairness in Large Language Models in Three HoursPreprint
Viet, T. D., Wang, Z., Nguyen, M. N., & Zhang, W. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2408.00992 · Cited by 1
“Stereotypical association [1, 5, 32] method assesses bias by measuring the disparity in the rates at which different demographic groups are linked to stereotyped terms ( e.g., occupations) in the text generated by the model in response to a given prompt [32].”
MethodImplicit Priors Editing in Stable Diffusion via Targeted Token AdjustmentPreprint
He, F., Zhang, C., & Zhao, Z. (2024) · arXiv (Cornell University) · doi:10.48550/arxiv.2412.03400 · Cited by 1
“Previous work has highlighted the presence of biases in Word Token Embedding (WTE) in language models, particularly concerning sensitive attributes like race and gender [1, 7, 11, 15, 30, 47, 58].”
Caste And Occupational Identity In Large Language ModelsPreprint
Zaveri, J., & Shah, A. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5427214 · Cited by 1
Exploring Gender Bias in Large Language Models: An In-depth Dive into the German LanguagePreprint
Gnadt, K., Thulke, D., Kopeinik, S., & Schlüter, R. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2507.16557 · Cited by 1
“For example, GPT-3 has been shown to (re)produce biased outputs concerning re-ligion, specifically showing anti-Muslim sentiment (Abid et al., 2021).”
Deng, Q., Bai, X., Chen, K., Wang, Y., Nie, L., & Zhang, M. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2503.10093 · Cited by 1
“Large Language Models (LLMs) have achieved significant advancements in various domains, accompanied by growing safety concerns (Tan and Celis, 2019; Sheng et al., 2019; Sandbrink, 2023; Abid et al., 2021; Li et al., 2025).”
MEDEQUALQA: Evaluating Biases in LLMs with Counterfactual ReasoningPreprint
Ghosh, R., Gupta, A., McBride, H., Vaidya, A. J., & Mahmood, F. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2510.12818 · Cited by 1
“Broader audits highlight toxicity and religion-linked harms (Sap et al., 2019; Abid et al., 2021), as well as inequities from tokenization, multilingual gaps, and linguistic discrimination (Petrov et al., 2023; Huang et al., 2023; Dong et al., 2024).”
Through the LLM Looking Glass: A Socratic Probing of Donkeys, Elephants, and MarketsPreprint
Kennedy, M., Imani, A., Spinde, T., Aizawa, A., & Schütze, H. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2503.16674 · Cited by 1
“Recent studies have shown that large language models (LLMs) exhibit various forms of bias, including political, gender, and religious biases (Bender et al., 2021; Weidinger et al., 2022; Bommasani et al., 2021; Vig et al., 2020; Felkner et al., 2023; Abid et al., 2021; Reif and Schwartz, 2024).”
Advancing Academic Chatbots: Evaluation of Non Traditional OutputsPreprint
Favero, N., Salute, F., & Hardt, D. (2025) · arXiv (Cornell University) · doi:10.48550/arxiv.2512.00991 · Cited by 1
“However, their rise introduces pedagogical and institutional concerns: reduced human interaction, plagiarism, unreliable detection tools, and algorithmic bias (Cotton et al., 2023; MIT Sloan Teaching & Learning Technologies, 2025; Abid et al., 2021).”
The Role of Automation, Bots and AI in Influencing Knowledge of the Past
Wills-Eve, B. (2024) · Lancaster EPrints (Lancaster University) · doi:10.17635/lancaster/thesis/2483 · Cited by 1
Addressing Stereotypes in Large Language Models: A Critical Examination and Mitigation Approach
Kazi, F. I. (2024) · eScholarship (California Digital Library) · Cited by 1
Veldhuis, A., Kenny, S., Antle, A. N., Kirmiziyesil, Z., Felsinger, D. N., & Murai, Y. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6482397 · Cited by 1
Seo, H., Lee, B. H., Cho, J., Lim, S., & Chun, S. Y. (2026) · arXiv (Cornell University) · doi:10.48550/arxiv.2604.16481 · Cited by 1
Metadata
- Authors
- Abid, A., Farooqi, M., & Zou, J.
- Affiliations
- Stanford University (US)
- McMaster University (CA)
- Venue
- Proceedings of the 2021 AAAI/ACM Conference on AI, Ethics, and Society
- Publisher
- ACM
- Published
- 2021-07-21
- Pages
- 298-306
- ISSN
- 3065-8365 (ISSN-L)
- OpenAlex
- W3184144760
- Cited by
- 476
- References
- 3
- Citations by year
- 2026: 81 · 2025: 124 · 2024: 105 · 2023: 97 · 2022: 36 · 2021: 33
- FWCI
- 61.36
- Type
- conference paper
- Language
- en
- Open access
- green
- arxiv.org
- Locations
- Proceedings of the AAAI/ACM Conference on AI Ethics and Society (publishedVersion)
- arXiv (Cornell University) (submittedVersion, cc-by)
- arXiv (Cornell University) (cc-by)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Hate Speech and Cyberbullying Detection
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- Islamophobia, religious bias, large language models, GPT-3, analogical reasoning, adversarial prompts, story generation, bias mitigation
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.