On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜
Open accessThe DOI matches its Crossref record; 156 of the 159 references it cites were found in a registry or an index, and 3 are not checked yet.
Abstract
The past 3 years of work in NLP have been characterized by the development and deployment of ever larger language models, especially for English. BERT, its variants, GPT-2/3, and others, most recently Switch-C, have pushed the boundaries of the possible both through architectural innovations and through sheer size. Using these pretrained models and the methodology of fine-tuning them for specific tasks, researchers have extended the state of the art on a wide array of tasks as measured by leaderboards on specific benchmarks for English. In this paper, we take a step back and ask: How big is too big? What are the possible risks associated with this technology and what paths are available for mitigating those risks? We provide recommendations including weighing the environmental and financial costs first, investing resources into curating and carefully documenting datasets rather than ingesting everything on the web, carrying out pre-development exercises evaluating how the planned approach fits into research and development goals and supports stakeholder values, and encouraging research directions beyond ever larger language models.
Checked 2026-11-22 · 3 references not checked yet
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
The text we found for this paper looks like another document, so the self-check wasn’t run.
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
BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding
Devlin, J., Chang, M.-W., Lee, K., & Toutanova, K. (2019) · Proceedings of the 2019 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long and Short Papers) · doi:10.18653/v1/n19-1423 · Cited by 54,325 · cited 5×
“[39] in particular noted that training on a large dataset and fine-tuning for specific tasks leads to strictly increasing results on the GLUE tasks [138] for English as the hyperparameters of the model were increased.”
BackgroundMethodResultInfluentialGlove: Global Vectors for Word Representation
Pennington, J., Socher, R., & Manning, C. (2014) · Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP) · doi:10.3115/v1/d14-1162 · Cited by 34,071
Language 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×
“documents linked to from Reddit [25], plus Wikipedia and a collection”
BackgroundMethodInfluentialRoBERTa: A Robustly Optimized BERT Pretraining Approach
Liu, Y., Ott, M., Goyal, N., Du, J., Joshi, M., Chen, D., Levy, O., Lewis, M., Zettlemoyer, L., & Stoyanov, V. (2019) · arXiv · doi:10.48550/arxiv.1907.11692 · Cited by 16,515
Distilling the Knowledge in a Neural Network
Hinton, G., Vinyals, O., & Dean, J. (2015) · arXiv · doi:10.48550/arxiv.1503.02531 · Cited by 14,204 · cited 1×
“[26, 58], quantization [118, 153], factorized embedding parameterization and cross-layer parameter sharing [70], and progressive module replacing [146].”
Deep Contextualized Word Representations
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) · doi:10.18653/v1/n18-1202 · Cited by 11,983
Proceedings of the 24th international conference on Machine learning
(2007) · doi:10.1145/1273496 · Cited by 11,630
Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer
Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., & Liu, P. J. (2019) · arXiv · doi:10.48550/arxiv.1910.10683 · Cited by 10,929
The Mathematical Theory of Communication
Shannon, C. E., Weaver, W., & Wiener, N. (1950) · Physics Today · doi:10.1063/1.3067010 · Cited by 9,801
Research works printed but not linked: 48
- [4] American Psychological Association. 2019. Discrimination: What it is, and how to cope. https://www.apa.org/topics/discrimination (2019). Google Scholar
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- [15] Ruha Benjamin. 2019. Race After Technology: Abolitionist Tools for the New Jim Code. Polity Press, Cambridge, UK. Crossref Google Scholar
- [16] Elettra Bietti and Roxana Vatanparast. 2020. Data Waste. Harvard International Law Journal 61 (2020). Google Scholar
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- 10.1016/0010-0277(86) · doi:10.1016/0010-0277(86
- [38] Terrance de Vries, Ishan Misra, Changhan Wang, and Laurens van der Maaten. 2019. Does object recognition work for everyone?. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops. 52--59. Google Scholar
- 10.18653/v1 · DOI
- [42] Ethan Fast, Tina Vachovsky, and Michael Bernstein. 2016. Shirtless and Dangerous: Quantifying Linguistic Signals of Gender Bias in an Online Fiction Writing Community. In Proceedings of the International AAAI Conference on Web and Social Media, Vol. 10. Google Scholar
- 10.18653/v1 · DOI
- [45] Darja Fišer, Ruihong Huang, Vinodkumar Prabhakaran, Rob Voigt, Zeerak Waseem, and Jacqueline Wernimont (Eds.). 2018. Proceedings of the 2nd Workshop on Abusive Language Online (ALW2). Association for Computational Linguistics, Brussels, Belgium. https://www.aclweb.org/anthology/W18-5100 Google Scholar
- [50] Batya Friedman, Peter H. Kahn, Jr., and Alan Borning. 2006. Value sensitive design and information systems. In Human-Computer Interaction in Management Information Systems: Foundations, P Zhang and D Galletta (Eds.). M. E. Sharpe, Armonk NY, 348--372. Google Scholar
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- [67] Brendan Kennedy, Drew Kogon, Kris Coombs, Joseph Hoover, Christina Park, Gwenyth Portillo-Wightman, Aida Mostafazadeh Davani, Mohammad Atari, and Morteza Dehghani. 2018. A typology and coding manual for the study of hate-based rhetoric. PsyArXiv. July 18 (2018). Google Scholar
- [77] Marianna Martindale and Marine Carpuat. 2018. Fluency Over Adequacy: A Pilot Study in Measuring User Trust in Imperfect MT. In Proceedings of the 13th Conference of the Association for Machine Translation in the Americas (Volume 1: Research Track). Association for Machine Translation in the Americas, Boston, MA, 13--25. https://www.aclweb.org/anthology/W18-1803 Google Scholar
- [81] Douglas M McLeod. 2007. News coverage and social protest: How the media's protect paradigm exacerbates social conflict. Journal of Dispute Resolution (2007), 185. Google Scholar
- [87] Robert C. Moore and William Lewis. 2010. Intelligent Selection of Language Model Training Data. In Proceedings of the ACL 2010 Conference Short Papers. Association for Computational Linguistics, Uppsala, Sweden, 220--224. https://www.aclweb.org/anthology/P10-2041 Google Scholar
- 10.18653/v1 · DOI
- [92] Maggie Nelson. 2015. The Argonauts. Graywolf Press, Minneapolis. Google Scholar
- 10.18653/v1 · DOI
- [96] David Ortiz, Daniel Myers, Eugene Walls, and Maria-Elena Diaz. 2005. Where do we stand with newspaper data? Mobilization: An International Quarterly 10, 3 (2005), 397--419. Google Scholar
- 10.3115/v1 · doi:10.3115/v1
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- [106] Alec Radford, Jeffrey Wu, Rewon Child, David Luan, Dario Amodei, and Ilya Sutskever. 2019. Language models are unsupervised multitask learners. OpenAI Blog 1, 8 (2019), 9. Google Scholar
- 10.18653/v1 · DOI
- [109] Sarah T. Roberts, Joel Tetreault, Vinodkumar Prabhakaran, and Zeerak Waseem (Eds.). 2019. Proceedings of the Third Workshop on Abusive Language Online. Association for Computational Linguistics, Florence, Italy. https://www.aclweb.org/anthology/W19-3500 Google Scholar
- [112] Corby Rosset. 2020. Turing-NLG: A 17-billion-parameter language model by Microsoft. Microsoft Blog (2020). Google Scholar
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- [121] Joonbo Shin, Yoonhyung Lee, and Kyomin Jung. 2019. Effective Sentence Scoring Method Using BERT for Speech Recognition. In Asian Conference on Machine Learning. 1081--1093. Google Scholar
- [124] Karen Spärck Jones. 2004. Language modelling's generative model: Is it rational? Technical Report. Computer Laboratory, University of Cambridge. Google Scholar
- [127] Katrina Srigley and Lorraine Sutherland. 2019. Decolonizing, Indigenizing, and Learning Biskaaybiiyang in the Field: Our Oral History Journey1. The Oral History Review (2019). Google Scholar
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- [150] Zhilin Yang, Zihang Dai, Yiming Yang, Jaime Carbonell, Russ R Salakhutdinov, and Quoc V Le. 2019. Xlnet: Generalized autoregressive pretraining for language understanding. In Advances in neural information processing systems. 5753--5763. Google Scholar
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
- 10.18653/v1 · DOI
Cited by
Items 1,701–1,800 of 6,931
A powerful potion for a potent problem: transformative justice for generative AI in healthcare
Gross, N. (2024) · AI and Ethics · doi:10.1007/s43681-024-00519-1 · Cited by 13
Dinçer, S. (2024) · Educational Research & Implementation · doi:10.14527/edure.2024.10 · Cited by 15
Large Language Models (LLMs) as a Tool to Facilitate Information Seeking Behavior
Mannuru, N. R., Mannuru, A., & Lund, B. D. (2024) · InfoScience Trends · doi:10.61186/ist.202401.01.15 · Cited by 17
AI’s Impact on Sustainability Targets: A Cross-Country NCA and fsQCA Study
Vasist, P. N., & Krishnan, S. (2024) · Information Systems Frontiers · doi:10.1007/s10796-024-10543-5 · Cited by 13
“This growth ranges from DistilBERT, with 66 million parameters, to Switch-C, with over 1,500 billion parameters (Bender et al., 2021; Jiang et al., 2024).”
Gorrieri, L. (2024) · Journal of Ethics and Emerging Technologies · doi:10.55613/jeet.v34i1.149 · Cited by 17
The use of generative AI for coding in academia
Wills, S., Poon, S. T. S., Salili‐James, A., & Scott, B. (2024) · Methods in Ecology and Evolution · doi:10.1111/2041-210x.14454 · Cited by 11
“The environmental costs of generative AI cannot be ignored (Bender et al., 2021) with the International Energy Agency estimating that data centers and data transmission networks account for 2%–3% of global electricity usage (International Energy Agency, 2024) and the World Economic Forum estimating…”
Leveraging Large Language Models for Enhanced Classification and Analysis: Fire Incidents Case Study
Alkhammash, E. H. (2024) · Fire · doi:10.3390/fire8010007 · Cited by 12
Algorithmic bias: sexualized violence against women in GPT-3 models
Wyer, S., & Black, S. (2025) · AI and Ethics · doi:10.1007/s43681-024-00641-0 · Cited by 19
“The industry’s focus on quantity rather than quality of data has led to concern over diversity and lack of repre - sentation [ 7, 39 , 66], leading to problematic text generation with social harms relating to downstream tasks [ 11].”
InfluentialSummon a demon and bind it: A grounded theory of LLM red teaming
Inie, N., Stray, J., & Derczynski, L. (2025) · PLoS ONE · doi:10.1371/journal.pone.0314658 · Cited by 46
Machine Learning, Synthetic Data, and the Politics of Difference
Jacobsen, B. N. (2025) · Theory Culture & Society · doi:10.1177/02632764241304687 · Cited by 12
“…of all the different ethnic groups’ and have ‘gender balance and age balance’ (Kaliouby cited in Ford, 2018: 222) through the generation of synthetic data attributes on race, gender, and other protected characteristics (Bender et al., 2021; Buolamwini and Gebru, 2019; Crawford and Paglen, 2019).”
Schwitter, N. (2025) · Methodological Innovations · doi:10.1177/20597991251313876 · Cited by 17
“The deep neural networks used in LLMs are widely recognised as black boxes with opaque decision-making processes (Dobson, 2023) which retain the biases inherent in the data that they had been trained on (Bender et al., 2021; Navigli et al., 2023).”
Magnus, P. D. (2025) · Episteme · doi:10.1017/epi.2024.29 · Cited by 15
“It is, in the vivid metaphor of Emily Bender and collaborators, a stochastic parrot (Bender et al. 2021).”
Kirpalani, A., Grimmer, J., & Wang, P. Z. T. (2025) · Medical Science Educator · doi:10.1007/s40670-025-02289-9 · Cited by 14
“However, the use of artificial intelligence (AI) and NLP in education also introduces the potential for misinformation and hallucination, where the AI might generate plausible but incorrect or misleading information [2].”
Training Green AI Models Using Elite Samples
Alswaitti, M., Verdecchia, R., Danoy, G., Bouvry, P., & Pecero, J. E. (2025) · IEEE Transactions on Sustainable Computing · doi:10.1109/tsusc.2025.3544430 · Cited by 21
“As documented in the literature, training an AI modelcanemitasmuchcarbonasfivecarsintheirlifetimes[11], anequivalentto125round-tripflightsbetweenNewYorkand Beijing[12],morethan626,000poundsofcarbondioxide[13], [14],andenergyconsumptionequivalenttoatrans-American flight[15].”
Zhu, M. (2025) · Journal of Computer Assisted Learning · doi:10.1111/jcal.70013 · Cited by 13
Gen AI and research integrity: Where to now?
Vasconcelos, S. M. R., & Marušić, A. (2025) · EMBO Reports · doi:10.1038/s44319-025-00424-6 · Cited by 15
“Even if one regards LLMs as “ stochastic parrots ” , a term adopted by Bender et al (2021) to emphasize that these systems remix patterns in their training data without true understanding, alignment remains a critical problem.”
Between Innovation and Caution: How Consumers’ Risk Perception Shapes AI Product Decisions
Moon, W.-K., Wei, X., Overton, H. K., & Kim, J. K. (2025) · Journal of Current Issues & Research in Advertising · doi:10.1080/10641734.2025.2468474 · Cited by 13
“Consumers may have more concerns about AI’s lower-than-expected performance than given explanations by makers (Bender et al. 2021).”
Cyber–Physical Resilience: Evolution of Concept, Indicators, and Legal Frameworks
Longo, A., Ardebili, A. A., Lazari, A., & Ficarella, A. (2025) · Electronics · doi:10.3390/electronics14081684 · Cited by 17
Diversification in the World of Data and AI
Lanzolla, G., & Markides, C. (2025) · Cambridge University Press eBooks · doi:10.1017/9781009515801 · Cited by 10
Evaluating the performance of large language & visual-language models in cervical cytology screening
Hong, Q., Liu, S., Wu, L., Lu, Q., Yang, P., Chen, D., Rao, G., Liu, X., Ye, H., Zhuang, P., et al. (2025) · npj Precision Oncology · doi:10.1038/s41698-025-00916-7 · Cited by 15
“GPT-4V can accurately describe smear characteristics, while other LVLMs exhibited some hallucinations 34 , introducing fabricated information that could potentially mislead clinical decision-making (Fig.”
Large language models in climate and sustainability policy: limits and opportunities
Larosa, F., Hoyas, S., Conejero, J. A., García‐Martínez, J., Nerini, F. F., & Vinuesa, R. (2025) · Environmental Research Letters · doi:10.1088/1748-9326/addd36 · Cited by 28
“The strongest fronts call for more transparent, open and explainable models [17, 18], but there are growing concerns about AI’s environmental footprint and costs [19, 20].”
Validity and Reliability Analysis of the Artificial Intelligence–Digital Life Balance Scale
Erdemir, N., & Atik, S. (2025) · Psychiatric Quarterly · doi:10.1007/s11126-025-10167-1 · Cited by 15
Farjam, M., Meyer, H., & Lohkamp, M. (2025) · Social Science Computer Review · doi:10.1177/08944393251349541 · Cited by 34
“By overseeing the reasoning and interactions between the agents, human intermediaries can identify and rectify issues such as misinterpretations, ambiguities, or biases — problems that LLMs can (re)produce according to Bender et al. (2021).”
Nan, L., Siriaraya, P., She, W. J., & Kuwahara, N. (2025) · ACM Conference on User Modeling, Adaptation and Personalization (UMAP) · doi:10.1145/3708319.3733691 · Cited by 11
Transphobia Is in the Eye of the Prompter: Trans-Centered Perspectives on Large Language Models
Scheuerman, M. K., Weathington, K., Petterson, A., Doyle, D. T., Das, D., DeVito, M. A., & Brubaker, J. R. (2025) · ACM Transactions on Computer-Human Interaction · doi:10.1145/3743676 · Cited by 14
“Given the fiscal, computational, and environmental cost of training LLMs 52:27 [85, 186], re-training LLMs is considered an undesirable approach to dealing with concept drift or problematic outputs.”
Rethinking Error: “Hallucinations” and Epistemological Indifference
Fredrikzon, J. (2025) · Critical AI · doi:10.1215/2834703x-11700255 · Cited by 15
Empathetic Large Language Models, the social capacities and human flourishing
Sung, L. U., & Ferdman, A. (2025) · Inquiry · doi:10.1080/0020174x.2025.2518448 · Cited by 15
From meaning to emotions: LLMs as artificial communication partners
Gutiérrez, J. L. M. (2025) · AI & Society · doi:10.1007/s00146-025-02481-w · Cited by 18
“These models generate fluent, contex-tually appropriate responses by probabilistically assembling linguistic patterns drawn from vast training data, yet they do so without any grasp of meaning—hence the apt, now-famous description of them as “stochastic parrots” (Bender et al. 2021).”
Do Artifacts Have Political Economy?
Birch, K. (2025) · Science Technology & Human Values · doi:10.1177/01622439251352167 · Cited by 21
Technosolutionism and the empathetic medical chatbot
Sharon, T. (2025) · AI & Society · doi:10.1007/s00146-025-02441-4 · Cited by 17
“As Bender and her co-authors have famously argued, LLMs should be understood as “stochastic parrots” (Bender et al. 2021 As a computational linguist, Bender’s claims pertain to the absence of intentional states and reference to meaning in LLMs in general, not to empathy in particular.”
How to Use Generative AI in Educational Research
Roe, J. (2025) · Cambridge University Press eBooks · doi:10.1017/9781009675338 · Cited by 22
Abrar, A., Oeshy, N. T., Kabir, M., & Ananiadou, S. (2025) · AI & Society · doi:10.1007/s00146-025-02721-z · Cited by 32
“When trained on extensive uncurated internet data, LLMs often perpetuate biases that disproportionately a ff ect vulnerable and marginalized communities(Bender et al., 2021; Dodge et al., 2021; Sheng et al., 2021).”
MMChat: Multi-Modal Chat Dataset on Social Media
Zheng, Y., Chen, G., Liu, X., & Sun, J. (2022) · Language Resources and Evaluation Conference (LREC) · doi:10.63317/2adki5cd6ucv · Cited by 48
“Regarding the potential biases, except those from the dataset itself (Henderson et al., 2018) (which always exists in dialogue datasets), biases might be introduced by the pre-trained language model (Bender et al., 2021) and the pre-trained image encoder (Steed and Caliskan, 2020) used in this work.”
The five Is: Key principles for interpretable and safe conversational AI
Wahde, M., & Virgolin, M. (2021) · doi:10.1145/3507623.3507632 · Cited by 16
Connecting the Dots in Self-Supervised Learning: A Brief Survey for Beginners
Fang, P.-F., Li, X., Yan, Y., Zhang, S., Kang, Q., Li, X.-F., & Lan, Z. (2022) · Journal of Computational Science and Technology · doi:10.1007/s11390-022-2158-x · Cited by 9
“Because SSL with Trans-former can continuously bene(cid:12)t from larger architectures and larger quantities of data, one of the biggest trends for SSL in NLP has been the ever-increasing model size [83] .”
BackgroundAI algorithmic oversight: new frontiers in regulation
Busuioc, M. (2022) · Edward Elgar Publishing eBooks · doi:10.4337/9781839108990.00043 · Cited by 37
Cheng, T., Abowd, G. D., Oh, H., & Hester, J. D. (2023) · Workshop on Sustainable Computer Systems (HotCarbon) · doi:10.1145/3604930.3605723 · Cited by 14
“Data centers host innumerable applications, including cutting-edge large language models (LLMs), e-commerce websites, video streaming, sensor data fusion and search engines [6].”
Training Computational Social Science PhD Students for Academic and Non-Academic Careers
Kesari, A., Kim, J. Y., Shah, S., Brown, T., Ventura, T., & Law, T. L. (2023) · PS Political Science & Politics · doi:10.1017/s1049096523000732 · Cited by 12
“For example, the fairly recent development of large language models such as ChatGPT raises ethical concerns about everything from plagiarism to data rights to the safety of open-access code (Bender et al. 2021).”
BackgroundHandling the hype: Implications of AI hype for public interest tech projects
Züger, T., Kuper, F., Faßbender, J., Katzy-Reinshagen, A., & Kühnlein, I. (2023) · TATuP - Zeitschrift für Technikfolgenabschätzung in Theorie und Praxis · doi:10.14512/tatup.32.3.34 · Cited by 15
Thiel, S., Bernhardt, J., & Medien, B. D. B. F. K. U. (2023) · Edition Museum · doi:10.1515/9783839467107 · Cited by 36
The harms of terminology: why we should reject so-called “frontier AI”
Helfrich, G. (2024) · AI and Ethics · doi:10.1007/s43681-024-00438-1 · Cited by 24
Baer, M. F., Wartmann, F. M., Fagerholm, N., & Purves, R. S. (2024) · Ecosystems and People · doi:10.1080/26395916.2024.2331761 · Cited by 13
“…essential to many machine learning or computational linguistics efforts, especially in domains not likely to be common in large-scale training data, which are dominated by easily available online collections such as Wikipedia, news stories and bulletin boards such as Reddit (Bender et al. 2021).”
BackgroundFairframe: a fairness framework for bias detection and mitigation in news
Sallami, D., & Aïmeur, E. (2024) · AI and Ethics · doi:10.1007/s43681-024-00568-6 · Cited by 19
“Identifying and addressing these biases is essential for the development of fairer and more inclusive NLP technologies, as biases can lead to social harm by fostering preju-dices and perpetuating stereotypes [32, 45, 46].”
Enabling Collaborative Data Science Development with the Ballet Framework
Smith, M. J., Cito, J., Lu, K., & Veeramachaneni, K. K. (2021) · Proceedings of the ACM on Human-Computer Interaction · doi:10.1145/3479575 · Cited by 13
“As the field of machine learning rapidly advances, more and more ethical considerations are being raised, including of recent models [7, 11, 83].”
BackgroundInvestigating Memorization of Conspiracy Theories in Text Generation
Levy, S., Saxon, M., & Wang, W. Y. (2021) · Findings of the Association for Computational Linguistics: ACL (Findings of ACL) · doi:10.18653/v1/2021.findings-acl.416 · Cited by 25
“As NLG models are being utilized for various tasks such as chatbots and recommendations systems (Gatt and Krahmer, 2018), cases arise in which these conspiracy theories and other biases can propagate unintentionally (Bender et al., 2021).”
MethodHaider, P., Ellenberger, B., Kriener, L., Jordan, J., Senn, W., & Petrovici, M. A. (2021) · BORIS (University Library Bern) · doi:10.12751/nncn.bc2022.016 · Cited by 22
Graham, S., Yates, D., & El-Roby, A. (2023) · Open Research Europe · doi:10.12688/openreseurope.16003.1 · Cited by 14
“(Shanahan, 2022) LLMs have famously been called ‘stochastic parrots’ (Bender et al., 2021) meaning they have no access to meaning.”
InfluentialStabilized In-Context Learning with Pre-trained Language Models for Few Shot Dialogue State Tracking
Chen, D., Qian, K., & Zhou, Y. (2023) · Findings of the Association for Computational Linguistics: EACL · doi:10.18653/v1/2023.findings-eacl.115 · Cited by 21
BERT Goes Shopping: Comparing Distributional Models for Product Representations
Tagliabue, J., Bianchi, F., & Yu, B. (2021) · Workshop on e-Commerce and NLP (ECNLP) · doi:10.18653/v1/2021.ecnlp-1.1 · Cited by 20
“Even if the debate on the social impact of large-scale models often feels very complex (Thompson et al., 2020; Bender et al., 2021) – and, sometimes, removed from our day-to-day duties – Prod2BERT gave us a glimpse of what unequal access to resources may mean in more meaningful contexts.”
BackgroundScatterbrain: Unifying Sparse and Low-rank Attention Approximation
Chen, B., Dao, T., Winsor, E., Song, Z., Rudra, A., & Ré, C. (2021) · neural information processing systems · doi:10.48550/arxiv.2110.15343 · Cited by 40
“Specially, for language models (LMs), while our work partially addresses the issue of environmental cost of LMs raised in [6], it does not address other issues such as unfathomable training data [6].”
BackgroundChatGPT and Finetuned BERT: A Comparative Study for Developing Intelligent Design Support SystemsPreprint
Qiu, Y., & Jin, Y. (2023) · SSRN Electronic Journal · doi:10.2139/ssrn.4516782 · Cited by 9
The Robots are Here: Navigating the Generative AI Revolution in Computing Education
Prather, J., Denny, P., Leinonen, J., Becker, B. A., Albluwi, I., Craig, M., Keuning, H., Kiesler, N., Kohn, T., Luxton-Reilly, A., et al. (2023) · arXiv · doi:10.48550/arxiv.2310.00658 · Cited by 9
Building the Epistemic Community of AI SafetyPreprint
Ahmed, S., Jaźwińska, K., Ahlawat, A., Winecoff, A., & Wang, M. (2023) · SSRN Electronic Journal · doi:10.2139/ssrn.4641526 · Cited by 9
AI Personal Assistants and Sustainability: Risks and Opportunities
Rillig, M. C., & Kasirzadeh, A. (2024) · Environmental Science & Technology · doi:10.1021/acs.est.4c03300 · Cited by 10
Reasons in the Loop: The Role of Large Language Models in Medical Co-Reasoning
Mann, S. P., Earp, B. D., Liu, P., & Savulescu, J. (2024) · The American Journal of Bioethics · doi:10.1080/15265161.2024.2383121 · Cited by 11
“One might object that LLMs merely generate statistically likely responses without genuine understanding (Bender et al. 2021), or that they lack access to relevant feelings and contextual nuances because they are unlikely to be conscious (Chalmers 2023).”
Peddinti, S. R., Katragadda, S., Pandey, B. K., & Tanikonda, A. (2025) · SSRN Electronic Journal · doi:10.2139/ssrn.5119925 · Cited by 24
4E cognition and the coevolution of human–AI interaction
Noller, J. (2025) · Discover Artificial Intelligence · doi:10.1007/s44163-025-00595-0 · Cited by 15
RL with KL penalties is better viewed as Bayesian inference
Korbak, T., Perez, E., & Buckley, C. (2022) · Findings of the Association for Computational Linguistics: EMNLP 2022 · doi:10.18653/v1/2022.findings-emnlp.77 · Cited by 8
GhostCite: A Large-Scale Analysis of Citation Validity in the Age of Large Language Models
Xu, Z., Qiu, Y., Sun, L., Miao, F., Wu, F., Li, X., Wang, X., Lu, H., Zhang, Z., Hu, Y., et al. (2026) · arXiv · doi:10.48550/arxiv.2602.06718 · Cited by 3
“As Bender et al. [23] noted, LLMs function as stochastic parrots that prioritize the structure of language over its truth.”
Writing and Thinking: What Changes with Digital Technologies?
Kruse, O., & Anson, C. M. (2023) · Springer eBooks · doi:10.1007/978-3-031-36033-6_29 · Cited by 17
Foster, J. G. (2023) · Public Culture · doi:10.1215/08992363-10742593 · Cited by 14
Bai, J. Y. H., Marín, V. I., Jung, I., & Zawacki‐Richter, O. (2023) · Ubiquity Proceedings · doi:10.5334/uproc.74 · Cited by 8
“However, the prohibitive costs to developing cutting-edge AI tools, combined with resource disparities between stakeholders, can result in AI applications that further entrench existing inequalities (see Bender et al., 2021; Facer & Selwyn, 2021; Hu et al., 2019, for critical discussions).”
BackgroundMitigating Biases in Hate Speech Detection from A Causal Perspective
Zhang, Z., Chen, J., & Yang, D. (2023) · Findings of the Association for Computational Linguistics: EMNLP (Findings of EMNLP) · doi:10.18653/v1/2023.findings-emnlp.440 · Cited by 22
From sublime awe to abject cringe: on the embodied processing of AI art
Denson, S. (2023) · Journal of Visual Culture · doi:10.1177/14704129231194136 · Cited by 19
“A central statement, focused on large language models (LLMs) but with implications for other forms of AI as well, can be found in Bender et al. (2021: 610–623).”
BackgroundPerspectives on the promise and perils of generative AI in academia
Yawson, R. M. (2024) · Human Resource Development International · doi:10.1080/13678868.2024.2334983 · Cited by 15
“Large language models can have carbon footprints equivalent to entire human lifetimes from training alone – an environmental impact that is difficult to justify unless very selectively deployed (Bender et al. 2021).”
Bridging the Gap: Exploring the Revolutionary Application of GenAI in Language Teaching and Learning
Jadhav, S. (2024) · Journal of Electrical Systems · doi:10.52783/jes.2325 · Cited by 11
Opening a conversation on responsible environmental data science in the age of large language models
Oliver, R. Y., Chapman, M., Emery, N. C., Gillespie, L., Gownaris, N. J., Leiker, S., Nisi, A. C., Ayers, D., Breckheimer, I., Blondin, H. E., et al. (2024) · Environmental Data Science · doi:10.1017/eds.2024.12 · Cited by 13
“While intended to remove objectionable content such as violence or hate speech, data curation can have unintended consequences and create temporal homogenization of viewpoints (Bender et al., 2021).”
BackgroundInfluentialLeverage Large Language Models For Enhanced Aviation Safety
Fox, K. L., Niewoehner, K. R., Rahmes, M. D., Wong, J., & Razdan, R. (2024) · Integrated Communications, Navigation and Surveillance Conference (ICNS) · doi:10.1109/icns60906.2024.10550651 · Cited by 8
“Language Models (LMs) are representations of natural human language and are typically used to predict the next word in a sequence [4] [5] (see Fig.”
Backgroundvan Hoof, M., Trilling, D., Meppelink, C. S., Möller, J. E., & Löecherbach, F. (2024) · Communication Methods and Measures · doi:10.1080/19312458.2024.2363776 · Cited by 12
“Moreover, if the pre-training data is biased toward certain topics or perspectives (Bender et al., 2021), the classifier may be less accurate at identifying PNR searches outside of those biases.”
Jones, R. M., Bryan, E.-M., & Preston, C. (2024) · Journal of Chemical Education · doi:10.1021/acs.jchemed.4c00271 · Cited by 12
Babu, C. V. S., Anniyappa, C. S. A., & Dharma Sastha, B. (2024) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3693-3860-5.ch008 · Cited by 8
Chen, Y., Wu, C., Sui, R., & Zhang, J. (2024) · Big Data and Cognitive Computing · doi:10.3390/bdcc8080094 · Cited by 14
Membership categorisation, sociological description and role prompt engineering with ChatGPT
Housley, W., & Dahl, P. (2024) · Discourse & Communication · doi:10.1177/17504813241267068 · Cited by 8
“Consequently, the LLM’s that drive the current applications in text generative A.I. have been critically described as ‘stochastic parrots’ (Bender et al., 2021).”
More, A. B. (2024) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3693-8557-9.ch002 · Cited by 9
Attention (to Virtuosity) Is All You Need: Religious Studies Pedagogy and Generative AI
Barlow, J., & Holt, L. (2024) · Religions · doi:10.3390/rel15091059 · Cited by 8
Vom Homo organisans zum Homo digitalis: Anthropologie als Theorie digitaler Organisation
Wendt, T. G. (2024) · MedienPädagogik Zeitschrift für Theorie und Praxis der Medienbildung · doi:10.21240/mpaed/63/2024.09.13.x · Cited by 9
Lotherington, H., Pegrum, M. A., Thumlert, K., Tomin, B., Boreland, T., & Pobuda, T. (2024) · Interactive Technology and Smart Education · doi:10.1108/itse-02-2024-0038 · Cited by 13
Impact of GPT on the Academic Ecosystem
Cai, Y., Deng, Q., Lv, T., Zhang, W., & Zhou, Y. (2024) · Science & Education · doi:10.1007/s11191-024-00561-9 · Cited by 17
“In this way, a seemingly boring scientific discovery can easily be transformed into materials suitable for people of all ages and knowledge levels.”
The data dance: choreographing seamless partnerships between humans, data, and GenAI
Durango, I., Gallud, J. A., & Penichet, V. M. R. (2024) · International Journal of Data Science and Analytics · doi:10.1007/s41060-024-00684-w · Cited by 11
LLMs beyond the lab: the ethics and epistemics of real-world AI research
Mollen, J. (2024) · Ethics and Information Technology · doi:10.1007/s10676-024-09819-w · Cited by 12
“Due to the energy consumption and global resources required during the entire lifespan of an AI system, scholars have increasingly drawn attention to the carbon cost and environmental impact of AI systems (Dhar, 2020; Bender et al., 2021).”
InfluentialPlacing Human Agency in the AI-powered Media Localisation Industry
Orrego-Carmona, D. (2024) · Journal of Audiovisual Translation · doi:10.47476/jat.v7i2.2024.306 · Cited by 13
“These magnified discussions take precedence over more immediate and systemic risks posed by AI, such as potential cultural insensitivity, exacerbating biases or deepening inequality (see, Aka et al., 2021; Bender et al., 2021; Johnson et al., 2022; Luccioni et al., 2023).”
Thieme, A., Rajamohan, A. M., Cooper, B., Groombridge, H., Simister, R. J., Wong, B., Woznitza, N., Pinnock, M. A., Wetscherek, M. T., Morrison, C. P., et al. (2025) · ACM Transactions on Computer-Human Interaction · doi:10.1145/3716500 · Cited by 20
Disclosure Dynamism: TCFD Aligned Climate Claims of UK Corporates
Cortés, D. G., Lodh, S., Akbar, S., & Nandy, M. (2025) · Business Strategy and the Environment · doi:10.1002/bse.4189 · Cited by 10
Multiagent Actor-Critic Generative AI for Query Resolution and Analysis
Ur Rahman, M. W., Nevarez, R., Mim, L. T., & Hariri, S. (2025) · IEEE Transactions on Artificial Intelligence · doi:10.1109/tai.2025.3544173 · Cited by 8
“The pursuit of more resilient and moral AI systems has led to continuous endeavors to enhance the accountability and transparency of data-driven natural language processing models [21], [22].”
A Review of the Challenges with Massive Web-Mined Corpora Used in Large Language Models Pre-training
Perełkiewicz, M., & Poświata, R. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-81596-6_14 · Cited by 24
Towards semantic versioning of open pre-trained language model releases on hugging face
Ajibode, A., Bangash, A. A., Cogo, F. R., Adams, B., & Hassan, A. E. (2025) · Empirical Software Engineering · doi:10.1007/s10664-025-10631-3 · Cited by 24
“This would enhance transparency and build trust in the evolution of PTLMs.”
Intersectional analysis of visual generative AI: the case of stable diffusion
Jääskeläinen, P., Sharma, N. K., Pallett, H., & Åsberg, C. (2025) · AI & Society · doi:10.1007/s00146-025-02207-y · Cited by 26
“A good example is the industry-based research published by Bender et al. (2021) that highlighted the potential discrimination of AI models and consequently led to the firing of one of the authors from Google’s responsible AI team (Guardian 2023).”
“Desired behaviors”: alignment and the emergence of a machine learning ethics
Schwerzmann, K., & Campolo, A. (2025) · AI & Society · doi:10.1007/s00146-025-02272-3 · Cited by 18
“While many critics have focused on limitations of the statistical nature of the language modeling objective (Bender et al. 2021), what Ouyang et al.’s statement shows is that it is not so much the statistical nature of the objective itself that produces alignment problems (although this may of…”
LLM-Based Automation of COSMIC Functional Size Measurement From Use Cases
Vito, G. D., Martino, S. D., Ferrucci, F., Gravino, C., & Palomba, F. (2025) · IEEE Transactions on Software Engineering · doi:10.1109/tse.2025.3554562 · Cited by 14
“LLMs struggle with tasks requiring deep logical reasoning, often producing incorrect outputs even if the input prompt is clear and unambiguous [65].”
Xiong, J., Pan, L., Liu, Y., Zhu, L., Zhang, L., & Tan, S. (2025) · Applied Sciences · doi:10.3390/app15073850 · Cited by 17
What If I Prefer Robot Journalists? Trust and Objectivity in the AI News Ecosystem
Yeste, E., Suau, J., Olivella, M. S., & Comas, E. X. I. (2025) · Journalism and Media · doi:10.3390/journalmedia6020051 · Cited by 18
“Furthermore, a critical examination by Bender et al. (2021) highlights that large language models, such as GPT-3, generate text based on patterns in their training data without genuine understanding, leading to outputs that may inadvertently reflect biases present in the data.”
Kruse, L., Rings, S., & Steinicke, F. (2025) · CHI Conference Extended Abstracts on Human Factors in Computing Systems (CHI EA) · doi:10.1145/3706599.3719808 · Cited by 8
Raza, S., Bamgbose, O., Ghuge, S., Tavakoli, F., Reji, D. J., & Bashir, S. R. (2025) · Machine Learning · doi:10.1007/s10994-025-06767-4 · Cited by 40
Artificial Intimacy: Exploring Normativity and Personalization Through Fine-tuning LLM Chatbots
Jones, M., Griffioen, N., Neumayer, C., & Shklovski, I. (2025) · CHI Conference on Human Factors in Computing Systems · doi:10.1145/3706598.3713728 · Cited by 26
“Even if that programming includes the ability to learn from user input, the worlds of LLMs are circumscribed by the training data and its limited capacity to represent the human world, even in the face of increasingly “big data” [7].”
Inteligencia artificial en la redacción y producción científica
Muñoz, G. F. R., Pindo, A. D. L. Á. O., Delgado, J. C. V., & Obando, K. E. R. (2025) · Revista Social Fronteriza · doi:10.59814/resofro.2025.5(3)705 · Cited by 24
Theorising notions of searching, (re)sources and evaluation in the light of generative AI
Sundin, O. (2025) · Information Research · doi:10.47989/ir30colis52258 · Cited by 24
Deng, W. H., Barocas, S., & Vaughan, J. W. (2025) · Proceedings of the ACM on Human-Computer Interaction · doi:10.1145/3711076 · Cited by 23
“This need is especially urgent in light of the recent rise of generative AI and the risks it poses [16, 76, 147], since much of the research on generative AI takes place in industry contexts.”
Mohammed, S. A., Murad, S. S., Albeyboni, H. J., Soltani, M. D., Ahmed, R. A., Badeel, R., & Chen, P. (2025) · Telecom · doi:10.3390/telecom6020035 · Cited by 17
Which educational approaches predict students’ generative AI confidence and responsibility?
Consoli, T., & Petko, D. (2025) · Computers and Education Artificial Intelligence · doi:10.1016/j.caeai.2025.100431 · Cited by 22
Su, C. C., & Chan, N. K. (2025) · Information Communication & Society · doi:10.1080/1369118x.2025.2513672 · Cited by 12
“The mystification of the magical power of GenAI (Leaver & Srdarov, 2023) overlooks pre-existing biases encoded in large language models like ChatGPT (Bender et al., 2021).”
Influential
Metadata
- Authors
- Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S.
- Affiliations
- University of Washington (US)
- Black in AI, Palo Alto, CA, USA
- The Aether
- Venue
- Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency
- Publisher
- ACM
- Published
- 2021-03
- Pages
- 610-623
- OpenAlex
- W3133702157
- Cited by
- 6,988
- References
- 117
- Citations by year
- 2026: 2484 · 2025: 1887 · 2024: 1153 · 2023: 842 · 2022: 384 · 2021: 229
- FWCI
- 300.72
- Type
- conference paper
- Language
- en
- Open access
- hybrid
- doi.org
- Licence
- Locations
- ACM Conference on Fairness, Accountability, and Transparency (FAccT) (publishedVersion, cc-by)
- Creighton Digital Repository (CDR) (submittedVersion)
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
- Keywords
- large language models, model size, AI risks, financial costs, data curation, dataset documentation, stakeholder value, BERT, GPT, fine-tuning
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.