Self-Diagnosis and Self-Debiasing: A Proposal for Reducing Corpus-Based Bias in NLP
Open accessThe DOI matches its Crossref record; 59 of the 59 references it cites were found in a registry or an index.
Abstract
⚠ This paper contains prompts and model outputs that are offensive in nature. When trained on large, unfiltered crawls from the Internet, language models pick up and reproduce all kinds of undesirable biases that can be found in the data: They often generate racist, sexist, violent, or otherwise toxic language. As large models require millions of training examples to achieve good performance, it is difficult to completely prevent them from being exposed to such content. In this paper, we first demonstrate a surprising finding: Pretrained language models recognize, to a considerable degree, their undesirable biases and the toxicity of the content they produce. We refer to this capability as self-diagnosis. Based on this finding, we then propose a decoding algorithm that, given only a textual description of the undesired behavior, reduces the probability of a language model producing problematic text. We refer to this approach as self-debiasing. Self-debiasing does not rely on manually curated word lists, nor does it require any training data or changes to the model’s parameters. While we by no means eliminate the issue of language models generating biased text, we believe our approach to be an important step in this direction.1
Overview
Integrity signals
Registry
Cites
Statistics
Verification
Claims the paper makes
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?
In the citing paper
“However, it has limitations as it cannot be applied to downstream tasks.”
From Towards trustworthy LLMs: a review on debiasing and dehallucinating in large language models (2024) · cites this paper as Schick et al. (2021)
In the citing paper
“Importantly, this method does not compromise the language model and requires no additional training.”
From Towards trustworthy LLMs: a review on debiasing and dehallucinating in large language models (2024) · cites this paper as Schick et al. (2021)
Citations
Cites
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 16×
“…illustrating that both static (e.g., Mikolov et al., 2013; Bojanowski et al., 2017) and contextualized word embeddings (e.g., Peters et al., 2018; Devlin et al., 2019) pretrained in a self-supervised fashion exhibit all kinds of unfair and discriminative biases (Bolukbasi et al., 2016; Caliskan…”
BackgroundMethodInfluentialEfficient 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×
“There is a large body of work illustrating that both static (e.g., Mikolov et al., 2013; Bojanowski et al., 2017) and contextualized word embeddings (e.g., Peters et al., 2018; Devlin et al., 2019) pretrained in a self-supervised fashion exhibit all kinds of unfair and discriminative biases…”
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
RoBERTa: 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
Advances in neural information processing systems 7
(1996) · Computers & Mathematics with Applications · doi:10.1016/s0898-1221(96)90463-0 · Cited by 14,281
Deep Contextualized Word Representations
Peters, M., Neumann, M., Iyyer, M., Gardner, M., Clark, C., Lee, K., & Zettlemoyer, L. (2018) · Proceedings of the 2018 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, Volume 1 (Long Papers) · doi:10.18653/v1/n18-1202 · Cited by 11,983 · cited 1×
“…a large body of work illustrating that both static (e.g., Mikolov et al., 2013; Bojanowski et al., 2017) and contextualized word embeddings (e.g., Peters et al., 2018; Devlin et al., 2019) pretrained in a self-supervised fashion exhibit all kinds of unfair and discriminative biases (Bolukbasi et…”
BackgroundExploring 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 · cited 13×
“Figure 1: Most probable continuations according to T5-3B (Raffel et al., 2020) and GPT2-XL (Radford et al.”
BackgroundMethodInfluentialEnriching Word Vectors with Subword Information
Bojanowski, P., Grave, E., Joulin, A., & Mikolov, T. (2017) · Transactions of the Association for Computational Linguistics · doi:10.1162/tacl_a_00051 · Cited by 9,645 · cited 1×
“There is a large body of work illustrating that both static (e.g., Mikolov et al., 2013; Bojanowski et al., 2017) and contextualized word embeddings (e.g., Peters et al., 2018; Devlin et al., 2019) pretrained in a self-supervised fashion exhibit all kinds of unfair and discriminative biases…”
BackgroundTransformers: State-of-the-Art Natural Language Processing
Wolf, T., Debut, L., Sanh, V., Chaumond, J., Delangue, C., Moi, A., Cistac, P., Rault, T., Louf, R., Funtowicz, M., et al. (2020) · Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations · doi:10.18653/v1/2020.emnlp-demos.6 · Cited by 8,741
On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? 🦜
Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021) · Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency · doi:10.1145/3442188.3445922 · Cited by 6,988 · cited 1×
“Building training datasets with more care and deliberation, an alternative solution discussed by Bender et al. (2021), is important, especially for improving linguistic and cultural diversity in online and other forms of communication.”
Background
Research works printed but not linked: 1
- Wang, 2019, BERT has a mouth, and it must speak: BERT as a Markov random field language model, 30, 10.18653/v1/W19-2304 · doi:10.18653/v1/w19-2304
Cited by
Items 201–237 of 237
Schema-tune: noise-driven bias mitigation in transformer-based language models
Shokrollahi, O., Penumatcha, R., Ensan, F., & Noorian, Z. (2025) · Machine Learning · doi:10.1007/s10994-024-06670-4
“Another approach to bias mitigation involve self-debiasing methods, i.e. those that do not require extensive labeled datasets (Guo et al. 2022; Schick et al. 2021; Gallegos et al. 2024).”
InfluentialDETCP: Self-Detoxifying Language Models With Contrastive Pairs
Liu, D., Liu, Y., Guo, J., & Mao, Z. (2025) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp49660.2025.10887947 · Cited by 1
“Compared with previous works that utilized prompt learning to detoxify PLMs [12,14], this work does not perform gradient backpropagation during decoding and considers the toxicity of the original context during sampling.”
InfluentialMalevolence Attacks Against Pretrained Dialogue Models
Ren, P., Li, R., Ren, Z., Chen, Z., de Rijke, M., & Zhang, Y. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-88708-6_24
“These models are pretrained on large-scale datasets that may contain negative views, bias, and other unsafe content [4,31].”
Debiasing Large Language Models with Structured Knowledge
Ma, C., Tianyu, Z. H. A. O., & Okumura, M. (2025) · IEICE Transactions on Information and Systems · doi:10.1587/transinf.2024edp7326
MGSMCAN: MultiGranularity Semantic Margin Cross-Annotation Network
Liao, G., & Ji, X. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-92747-8_18 · Cited by 1
Berlincioni, L., Cultrera, L., Becattini, F., Bertini, M., & Bimbo, A. D. (2025) · Lecture notes in computer science · doi:10.1007/978-3-031-92648-8_7 · Cited by 2
Expert-Guided Toxicity Filtration for Debiased Generation
Sun, X., Shi, K., Tang, H., Xu, G., & Li, Q. (2025) · Lecture notes in computer science · doi:10.1007/978-981-96-8183-9_10
FairTriplet: Balancing Fairness and Accuracy in Contextual Pre-Trained Models Through Prefix Tuning
Long, Z., Zheng, W., Zhou, D., Zhou, Y., Lin, N., & Yang, A. (2025) · International Conference on Computer Supported Cooperative Work in Design (CSCWD) · doi:10.1109/cscwd64889.2025.11033465
“These methods involve a variety of techniques, including projection-based methods such as IGBP; post-hoc text generation technologies [13], which train models using unbiased text; and fine-tuning PLMs with prompts that contain biased viewpoints [14], intended to directly adjust the model’s semantic…”
Examining and Mitigating Ability-bias in LLMs via Self-Reflection
Iyer, N., Jha, A., & Pradhan, A. (2025) · International Web for All Conference (W4A) · doi:10.1145/3744257.3744268
“Others have explored instruction-based prompting [10, 16, 26] to mitigate stereotypical responses or used self-debiasing to decrease the probability of predicting harmful next words for bias mitigation [24].”
Bias Beyond English: Evaluating Social Bias and Debiasing Methods in a Low-Resource Setting
Zhou, E., & Lü, W. (2025) · Lecture notes in computer science · doi:10.1007/978-981-95-3349-7_17 · Cited by 3
Uncovering Cultural Biases and Stereotypes in Large Language Models
Dorleon, G., & Shujaa, S. (2025) · Lecture notes in computer science · doi:10.1007/978-981-95-4861-3_5 · Cited by 2
LMDMI: A Lightweight Multilevel Defense against Malicious Inputs for Generative Language Models
Yan, S., & Wan, C. (2025) · International Conference on Artificial Intelligence and Foundation Model · doi:10.1145/3786709.3786720
“However, because safety alignment relies heavily on training samples from malicious corpora, which are scarce and lack diversity [18], it struggles to effectively counter various forms of malicious inputs, such as sensitive content, non-compliant content, prompt injection, and prohibited…”
PRISM: Principled Reasoning for Identifying and Suppressing Model Biases at Scale
Zhu, X., Li, W., & Lin, J. (2026) · Lecture notes in computer science · doi:10.1007/978-981-95-7081-2_3
Debias Your Large Multi-Modal Model at Test-Time via Non-Contrastive Visual Attribute Steering
Ratzlaff, N., Olson, M., Hinck, M., Tseng, S.-Y., Lal, V., & Howard, P. (2025) · IEEE International Conference on Computer Vision Workshops · doi:10.1109/iccvw69036.2025.00651 · Cited by 3
“Intervening on outputs has been explored via token-level block lists [13], domain classifiers [20] to trigger regeneration, or decoding strategies that decrease probabilities of generating certain tokens [27].”
Reducing Social Biases in Language Models: A Comparative Evaluation of Debiasing Strategies
Kamboj, P., Kumar, S., & Goyal, V. (2025) · International Conference on Advanced Computing and Emerging Technologies (ACET) · doi:10.1109/acet67282.2025.11430255 · Cited by 1
“…(CDS) [6], Counterfactual Data Augmentation (CDA) [5], debiased model (aieng-lab/roberta-large-gradiend-gender-debiased) [12], adversarial training [13], fine-tuned RoBERTa-large with CDA, Finetuned RoBERTa-large with CDA using LoRA [14], post-hoc prompting [15], post-hoc filtering [16].”
Towards Reliable Reasoning in Vision-Language Models
Hoang, C., Nassef, Y., & Hasan, M. R. (2025) · International Conference on Machine Learning and Applications · doi:10.1109/icmla66185.2025.00124
“Debiasing efforts in LLMs, such as Counterfactual Data Augmentation and Self-Debias techniques [22], [23], primarily address linguistic biases but often involve trade-offs with model performance.”
On Subjectivity, Bias and Fairness in Language Model Learning
Gaci, Y. (2023) · theses.fr (ABES) · doi:10.70675/dca82074za0c4z46d1z8a6bz96ce30bea3b6 · Cited by 3
Exploring attribution bias in LLMs: Social influences and prompt-based mitigation
Jiang, L., Cao, J., Zhu, G., & Wang, Y. (2026) · Technology in Society · doi:10.1016/j.techsoc.2026.103361 · Cited by 1
Wang, R., Zhu, F., & Zhang, X. (2026) · IEEE International Conference on Acoustics Speech and Signal Processing · doi:10.1109/icassp55912.2026.11460558
Calibrating Stereotype Preference with Multi-Agent Graph Reasoning in Large Language Models
Amin, M. N., & Jesser, A. (2025) · International Conference on Computer and Information Technology (ICCIT) · doi:10.1109/iccit68739.2025.11491743
“…outputs post-training [33,34], with trade-offs between helpfulness and harmlessness; and (iii) Inference/post-processing: prompt-time self-diagnosis/self-debiasing [40], contextual calibration [35] and label-bias calibration for in-context learning [36] adjust predictions without weight updates.”
Privacy-Aware Prompt Masking Interface for Secure Large Language Model Interactions
Sankar, H., Godwin, M. P., Vijay, A., Cladis, F. S. H., Prakasam, S., & Harshavarth, M. S. (2026) · 2026 International Conference on Data Science, Agents and Artificial Intelligence (ICDSAAI) · doi:10.1109/icdsaai69492.2026.11505205
TCG-CTG: A Thinking-based Constraint Guidance Dataset for Controllable Text Generation
He, Y., Zhuang, Y., & Sun, Y. (2026) · International Conference on Generative Artificial Intelligence and Information Security (GAIIS) · doi:10.1109/gaiis69281.2026.11519175
Don't be Tricked by Iterative Masking
Crothers, E., Viktor, H. L., & Japkowicz, N. (2026) · Computational Intelligence · doi:10.1111/coin.70247
“Note that there are also other post hoc meth-ods such as “Self-Debias” [62], which has strong performance in benchmarks but which cannot be used when fine-tuning on classification tasks [55].”
Expert Consensus on LLM Safety in Clinical Decision Support
Gayathri, K. V., Kanshiya, S., Saranraj, P., Lokeshwari, B., Vilmasen, P., & Wins, J. A. (2026) · Advances in computational intelligence and robotics book series · doi:10.4018/979-8-3373-7862-6.ch011
Counterfactual Consistency LoRA for Bias Mitigation on BBQ
Feng, D., Chen, H., & Zhou, Z. (2026) · International Seminar on Artificial Intelligence, Networking and Information Technology (AINIT) · doi:10.1109/ainit70033.2026.11557867
“Prompt-side debiasing is easy to deploy but often wording-sensitive [5], [9]; training-time methods based on counterfactual augmentation and parameter-efficient tuning are usually more stable under limited compute [2], [4], [10].”
CoPAMoE: Cognitive Process-Aware Mixture of Experts for Bias Understanding and MitigationPreprint
Liu, X., Wang, F., Wang, F., Jia, Z., Wang, H., & Ran, C. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6987086
CoPAMoE: Cognitive Process-Aware Mixture of Experts for Bias Understanding and MitigationPreprint
Liu, X., Wang, F., Wang, F., Jia, Z., Wang, H., & Ran, C. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.6997071
Rothe‐Wulf, A., Sekar, S. D., & Marsden, N. (2026) · Lecture notes in computer science · doi:10.1007/978-3-032-30846-7_34
RWKV-TS+: Adaptive Long-Range Prompt for Multiple Time Series Tasks
Pang, H., Xie, X., Li, Y., Jiang, Z., Zhang, X., Zhang, H., Wei, S., Zhao, Y., & Zhao, Y. (2026) · IEEE Transactions on Knowledge and Data Engineering · doi:10.1109/tkde.2026.3717305
Cohen, E., Adler, Y., & Gannot, R. N. (2026) · Healthcare · doi:10.3390/healthcare14162642
“While technical interventions—such as automated self-diagnosis and self-debiasing mechanisms embedded directly into the AI architecture [10]—are essential to reduce the cognitive burden on medical staff, they must be supported by a multidisciplinary governance structure to ensure long-term…”
The Migration of Bias in Law: From Human Cognition to Predictive AI System and Large Language Models
Zhong, H. (2026) · Springer eBooks · doi:10.1007/978-3-031-87993-7_185-1 · Cited by 1
CARP: Constraint-Aware Re-Prompting for Zero-Shot Bias Mitigation in Compact LLMs
Jagtap, P. A., & Ugale, T. (2026) · International Conference on Computing Communication Control and automation · doi:10.1109/iccubea69821.2026.11702098
“Chen et al. went in a slightly different direction with ZDDR, a detection-and-restore mechanism for adversarial inputs at inference time [2], and Schick et al. built an early self-diagnosis loop where the model tries to identify and then patch its own biases with no external validator [14].”
Beyond Static Guardrails: Auditable Surgical Output Control for Agentic LLM Workflows
Mauri, L., & Damiani, E. (2026) · IEEE International Conference on Web Services (ICWS) · doi:10.1109/icws72778.2026.11718616
“In recent years, several popular bias benchmarks have emerged [1]–[3], typically evaluating model behavior at the sequence level or through aggregated probability scores across demographic groups [4].”
Stereotype Bias Detection and Mitigation in Slovak and Multilingual Language Models
Sokolová, Z., Harahus, M., Sokol, M., & Kupcová, E. (2026) · International Conference Radioelektronika · doi:10.1109/radioelektronika71189.2026.11726461
“Self-Debias modifies the inference process to reduce biased outputs without retraining the model [11].”
FAIRITHM: Unlearning Bias and Rewriting for Fairness
Nam, Y., Lim, D., Im, S., Xiong, Z., & Park, J. (2026) · International Journal of Human-Computer Interaction · doi:10.1080/10447318.2026.2744493
Serçe, A., Kaçıkan, E., & Onan, A. (2026) · International Artificial Intelligence and Data Processing Symposium (IDAP) · doi:10.1109/idap71869.2026.11727894
“Schick et al. [13] propose self-diagnosis and self-debiasing, showing that pretrained models can recognize and reduce undesirable behavior to some extent using textual descriptions.”
Behavioral Calibration for Context-Adaptive Debiasing in Large Language ModelsPreprint
Bai, Y., Gao, M., Wang, Z., Gu, B., Xiong, Q., Wang, X., & Yang, L. (2026) · SSRN Electronic Journal · doi:10.2139/ssrn.7675451
Metadata
- Authors
- Schick, T., Udupa, S., & Schütze, H.
- Affiliations
- Ludwig-Maximilians-Universität München (DE)
- Corresponding author
- Schick, Timo
- Udupa, Sahana
- Schütze, Hinrich
- Venue
- Transactions of the Association for Computational Linguistics
- Publisher
- MIT Press
- Published
- 2021
- Volume
- 9
- Pages
- 1408-1424
- ISSN
- 2307-387X (ISSN-L)
- OpenAlex
- W3134354193
- Cited by
- 227
- References
- 58
- Citations by year
- 2026: 29 · 2025: 40 · 2024: 36 · 2023: 70 · 2022: 36 · 2021: 16
- FWCI
- 13.96
- Type
- journal article
- Language
- en
- Open access
- diamond
- doi.org
- Locations
- Transactions of the Association for Computational Linguistics (publishedVersion, cc-by)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Open access LMU (Ludwid Maxmilian's Universitat Munchen) (submittedVersion, cc-by)
- Universitätsbibliothek der LMU
- Topics
- Physical Sciences › Computer Science › Artificial Intelligence › Hate Speech and Cyberbullying Detection
- Physical Sciences › Computer Science › Artificial Intelligence › Ethics and Social Impacts of AI
- Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
- Keywords
- self-diagnosis, bias in language models, toxicity, bias detection, decoding algorithms, large language models
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.