pith. machine review for the scientific record. sign in

arxiv: 2605.05329 · v1 · submitted 2026-05-06 · 💻 cs.AI · cs.LG

Recognition: unknown

Understanding Annotator Safety Policy with Interpretability

Alex Oesterling , Donghao Ren , Yannick Assogba , Dominik Moritz , Sunnie S.Y. Kim , Leon Gatys , Fred Hohman
Authors on Pith no claims yet

Pith reviewed 2026-05-08 17:37 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords Annotator Policy Modelssafety policiesinterpretabilitydata annotationAI safetypolicy ambiguityvalue pluralism
0
0 comments X

The pith

Annotator Policy Models recover individual safety rules from labeling data alone without extra questions.

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

The paper establishes that simple interpretable models can extract what safety rules each annotator is following just by observing their yes/no decisions on examples. This matters because annotation disagreements often mix task errors, unclear policy wording, and genuine differences in values, yet current methods cannot tell these apart without costly follow-up interviews. The models reach over 80 percent accuracy in reproducing annotator choices, correctly forecast how the same person would label edited versions of an example, and detect known differences between groups in controlled tests. Applied to both human and LLM annotations, the approach reveals when instructions leave room for interpretation and when demographic groups hold distinct safety priorities.

Core claim

Annotator Policy Models (APMs) are introduced as interpretable models that learn annotators’ internal safety policies from labeling behavior alone. These models accurately capture annotator safety policy with over 80 percent accuracy, faithfully predict responses to counterfactual edits, and recover known policy differences in controlled settings. When applied to real annotations from humans and LLMs, APMs surface how annotators interpret safety instructions differently and uncover systematic differences in safety priorities across demographic groups, supporting more targeted and inclusive policy design.

What carries the argument

Annotator Policy Models (APMs), interpretable models trained solely on an annotator’s labeling decisions to represent their internal safety decision rules.

Load-bearing premise

Labeling behavior alone is enough to recover an annotator’s true internal policy rather than merely fitting patterns in the observed labels.

What would settle it

A controlled experiment in which APMs trained on initial labels fail to match the same annotators’ choices on new examples that require a specific policy distinction or fail to detect deliberately introduced policy differences between groups.

Figures

Figures reproduced from arXiv: 2605.05329 by Alex Oesterling, Dominik Moritz, Donghao Ren, Fred Hohman, Leon Gatys, Sunnie S.Y. Kim, Yannick Assogba.

Figure 1
Figure 1. Figure 1: Annotator Policy Models (APMs) learn interpretable representations of individual annotator safety policies. APMs are view at source ↗
Figure 2
Figure 2. Figure 2: Annotator disagreement on BeaverTails [40]. LLM annotators exhibit high disagreement rates on safety data. view at source ↗
Figure 3
Figure 3. Figure 3: Measuring unsafe annotation percent change and system faithfulness. view at source ↗
Figure 4
Figure 4. Figure 4: Recovery of disagreeing annotator policy with controlled and known annotator behavior. APMs are able to recover view at source ↗
Figure 5
Figure 5. Figure 5: Example findings using APMs on the DICES dataset. view at source ↗
Figure 6
Figure 6. Figure 6: Significance of APM differences across demographic groups. Each cell measures the proportion of disagreeing (where view at source ↗
Figure 7
Figure 7. Figure 7: Significance of group APM unique rules vs majority vote APM. For each demographic group, the figure plots the count view at source ↗
Figure 8
Figure 8. Figure 8: ROC curves for Annotator Policy Models For NNLR, we draw an ROC curve by sweeping over predicted probabilities. view at source ↗
Figure 9
Figure 9. Figure 9: Bootstrap uncertainty of NNLR weights by age. view at source ↗
Figure 10
Figure 10. Figure 10: Bootstrap uncertainty of NNLR weights by gender. view at source ↗
Figure 11
Figure 11. Figure 11: Bootstrap uncertainty of NNLR weights by race. view at source ↗
Figure 12
Figure 12. Figure 12: Bootstrap uncertainty of NNLR weights by education. view at source ↗
Figure 13
Figure 13. Figure 13: Significance of group APM unique rules vs gold-standard APM. For each demographic group, the figure plots the view at source ↗
read the original abstract

Safety policies define what constitutes safe and unsafe AI outputs, guiding data annotation and model development. However, annotation disagreement is pervasive and can stem from multiple sources such as operational failures (annotators misunderstand or misexecute the task), policy ambiguity (policy wording leaves room for interpretation), or value pluralism (different annotators hold different perspectives on safety). Distinguishing these sources matters. For example, operational failures call for quality control, ambiguity calls for policy clarification, and pluralism calls for deliberation about incorporating diverse perspectives. Yet understanding why annotators disagree is difficult. Directly asking annotators for their reasoning is costly, substantially increasing annotation burden, and can be unreliable for both human and LLM annotators as self-reported reasoning often fails to reflect actual decision processes. We introduce Annotator Policy Models (APMs), interpretable models that learn annotators' internal safety policies from labeling behavior alone, making annotator reasoning visible and comparable without additional annotation effort. We validate that APMs accurately model annotator safety policy (>80% accuracy), faithfully predict responses to counterfactual edits, and recover known policy differences in controlled settings. Applying APMs to LLM and human annotations, we demonstrate two core applications: (1) surfacing policy ambiguity by revealing how annotators interpret safety instructions differently, and (2) surfacing value pluralism by uncovering systematic differences in safety priorities across demographic groups. Together, these capabilities support more targeted, transparent, and inclusive safety policy design.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

1 major / 2 minor

Summary. The paper introduces Annotator Policy Models (APMs), interpretable models that infer annotators' internal safety policies directly from labeling behavior without extra annotations. It claims APMs achieve >80% accuracy in modeling policies, faithfully predict responses to counterfactual edits, and recover known policy differences in controlled settings. Applications include surfacing policy ambiguity (differing interpretations of instructions) and value pluralism (systematic differences across demographic groups) for both human and LLM annotators, to better distinguish operational failures, ambiguity, and pluralism in safety annotation.

Significance. If the empirical validations hold under scrutiny, this provides a practical, low-burden method to make annotator reasoning visible and comparable, supporting more targeted safety policy design. Strengths include the use of controlled simulations with known ground-truth policies and the focus on interpretability to avoid reliance on self-reported reasoning.

major comments (1)
  1. [Abstract and §4] Abstract and §4 (Experiments): The central validation claims (>80% accuracy, counterfactual faithfulness, recovery of known differences) are presented without reported baselines, data splits, statistical tests, or the precise metric and procedure used to assess counterfactual faithfulness. These details are load-bearing for evaluating whether APMs recover internal policies rather than surface label patterns.
minor comments (2)
  1. [§3] §3 (Method): Clarify the exact form of the interpretable models (e.g., decision trees, linear probes, or other) and how policy parameters are extracted and compared across annotators.
  2. [Figure captions and §5] Figure captions and §5 (Applications): Ensure all figures include error bars or confidence intervals when reporting accuracy or difference recovery, and explicitly state the number of annotators and items in each experiment.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive review and recommendation for minor revision. We address the single major comment below and will revise the manuscript accordingly to strengthen the empirical presentation.

read point-by-point responses

  1. Referee: [Abstract and §4] Abstract and §4 (Experiments): The central validation claims (>80% accuracy, counterfactual faithfulness, recovery of known differences) are presented without reported baselines, data splits, statistical tests, or the precise metric and procedure used to assess counterfactual faithfulness. These details are load-bearing for evaluating whether APMs recover internal policies rather than surface label patterns.

    Authors: We agree these details are necessary for rigorous assessment. In the revised manuscript we will add: (1) explicit baselines (majority-class predictor, logistic regression on surface features, and a non-interpretable neural model) with accuracy comparisons; (2) data-split procedures (per-annotator 5-fold cross-validation with held-out counterfactual test sets); (3) statistical tests (McNemar’s test for accuracy differences and bootstrap confidence intervals); and (4) the precise counterfactual faithfulness metric (exact-match accuracy of APM predictions on synthetically edited inputs against the annotator’s actual labels on those edits). These additions will clarify that performance exceeds surface-pattern baselines and that controlled simulations recover known policy differences, supporting the claim that APMs capture internal policy structure. revision: yes

Circularity Check

0 steps flagged

No significant circularity identified

full rationale

The paper introduces Annotator Policy Models (APMs) as interpretable models fitted directly to observed labeling behavior, then validates them via accuracy on held-out data (>80%), counterfactual edit predictions, and recovery of known policy differences in controlled simulations where ground-truth policies are available. No equations, derivations, or first-principles results are shown that reduce by construction to the fitted inputs or to self-citations. The central claims rest on standard empirical model fitting plus external validation benchmarks rather than any self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation chain. The derivation is therefore self-contained.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only review prevents detailed audit. The central claim rests on the unstated assumption that annotator labels are generated by an internal policy that can be approximated by an interpretable model; no free parameters, axioms, or invented entities are explicitly listed.

pith-pipeline@v0.9.0 · 5573 in / 1093 out tokens · 45014 ms · 2026-05-08T17:37:33.941233+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

86 extracted references · 38 canonical work pages · 8 internal anchors

  1. [1]

    International conference on machine learning , pages=

    Concept bottleneck models , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  2. [2]

    Label-free concept bottleneck models.arXiv preprint arXiv:2304.06129, 2023

    Label-free concept bottleneck models , author=. arXiv preprint arXiv:2304.06129 , year=

    work page arXiv

  3. [3]

    Aegis: Online adaptive ai content safety moderation with ensemble of llm experts, 2024 , author=

    2024

  4. [4]

    International conference on machine learning , pages=

    From softmax to sparsemax: A sparse model of attention and multi-label classification , author=. International conference on machine learning , pages=. 2016 , organization=

    2016

  5. [5]

    Advances in neural information processing systems , volume=

    Boolean decision rules via column generation , author=. Advances in neural information processing systems , volume=

  6. [6]

    International ai safety report

    International ai safety report , author=. arXiv preprint arXiv:2501.17805 , year=

    work page arXiv

  7. [7]

    arXiv preprint arXiv:2305.06626 , year=

    When the majority is wrong: Modeling annotator disagreement for subjective tasks , author=. arXiv preprint arXiv:2305.06626 , year=

    work page arXiv

  8. [8]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    You are what you annotate: Towards better models through annotator representations , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  9. [9]

    arXiv preprint arXiv:2509.01186 , year=

    Statutory Construction and Interpretation for Artificial Intelligence , author=. arXiv preprint arXiv:2509.01186 , year=

    work page arXiv

  10. [10]

    Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency , pages=

    AI Alignment at Your Discretion , author=. Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2025

  11. [11]

    Advances in Neural Information Processing Systems , volume=

    Beavertails: Towards improved safety alignment of llm via a human-preference dataset , author=. Advances in Neural Information Processing Systems , volume=

  12. [12]

    2024 , eprint=

    WildGuard: Open One-Stop Moderation Tools for Safety Risks, Jailbreaks, and Refusals of LLMs , author=. 2024 , eprint=

    2024

  13. [13]

    Advances in Neural Information Processing Systems , volume=

    Causal abstractions of neural networks , author=. Advances in Neural Information Processing Systems , volume=

  14. [14]

    Ohio , booktitle=

    Jacobellis v. Ohio , booktitle=. 1964 , publisher=

    1964

  15. [15]

    Ginette Vincendeau , title =

  16. [16]

    Yale LJ , volume=

    On I know it when I see it , author=. Yale LJ , volume=. 1995 , publisher=

    1995

  17. [17]

    Constitutional AI: Harmlessness from AI Feedback

    Constitutional ai: Harmlessness from ai feedback , author=. arXiv preprint arXiv:2212.08073 , year=

    work page internal anchor Pith review arXiv

  18. [18]

    Advances in Neural Information Processing Systems , volume=

    Dices dataset: Diversity in conversational ai evaluation for safety , author=. Advances in Neural Information Processing Systems , volume=

  19. [19]

    , author=

    Telling more than we can know: Verbal reports on mental processes. , author=. Psychological review , volume=. 1977 , publisher=

    1977

  20. [20]

    Advances in neural information processing systems , volume=

    Training language models to follow instructions with human feedback , author=. Advances in neural information processing systems , volume=

  21. [21]

    Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency , pages=

    Collective constitutional ai: Aligning a language model with public input , author=. Proceedings of the 2024 ACM Conference on Fairness, Accountability, and Transparency , pages=

    2024

  22. [22]

    On the hardness of faithful chain-of-thought reasoning in large language models.arXiv preprint arXiv:2406.10625, 2024

    On the hardness of faithful chain-of-thought reasoning in large language models , author=. arXiv preprint arXiv:2406.10625 , year=

    work page arXiv

  23. [23]

    Measuring Faithfulness in Chain-of-Thought Reasoning

    Measuring faithfulness in chain-of-thought reasoning , author=. arXiv preprint arXiv:2307.13702 , year=

    work page Pith review arXiv

  24. [24]

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

    Training a helpful and harmless assistant with reinforcement learning from human feedback , author=. arXiv preprint arXiv:2204.05862 , year=

    work page internal anchor Pith review arXiv

  25. [25]

    Concrete Problems in AI Safety

    Concrete problems in AI safety , author=. arXiv preprint arXiv:1606.06565 , year=

    work page internal anchor Pith review arXiv

  26. [26]

    A General Language Assistant as a Laboratory for Alignment

    A general language assistant as a laboratory for alignment , author=. arXiv preprint arXiv:2112.00861 , year=

    work page internal anchor Pith review arXiv

  27. [27]

    Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations

    Llama guard: Llm-based input-output safeguard for human-ai conversations , author=. arXiv preprint arXiv:2312.06674 , year=

    work page internal anchor Pith review arXiv

  28. [28]

    Proceedings of the 31st International Conference on Computational Linguistics , pages=

    Persona: A reproducible testbed for pluralistic alignment , author=. Proceedings of the 31st International Conference on Computational Linguistics , pages=

  29. [29]

    Neha Srikanth, Jordan Boyd-Graber, and Rachel Rudinger

    A roadmap to pluralistic alignment , author=. arXiv preprint arXiv:2402.05070 , year=

    work page arXiv

  30. [30]

    Available at SSRN 5319936 , year=

    Can Large Language Models Capture Human Annotator Disagreements? , author=. Available at SSRN 5319936 , year=

  31. [31]

    arXiv preprint arXiv:2408.14141 , year=

    Crowd-Calibrator: Can Annotator Disagreement Inform Calibration in Subjective Tasks? , author=. arXiv preprint arXiv:2408.14141 , year=

    work page arXiv

  32. [32]

    arXiv preprint arXiv:2502.08266 , year=

    Dealing with annotator disagreement in hate speech classification , author=. arXiv preprint arXiv:2502.08266 , year=

    work page arXiv

  33. [33]

    arXiv preprint arXiv:2409.17577 , year=

    Leveraging annotator disagreement for text classification , author=. arXiv preprint arXiv:2409.17577 , year=

    work page arXiv

  34. [34]

    arXiv preprint arXiv:2109.13563 , year=

    Agreeing to disagree: Annotating offensive language datasets with annotators' disagreement , author=. arXiv preprint arXiv:2109.13563 , year=

    work page arXiv

  35. [35]

    Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , year=

    Beyond black & white: Leveraging annotator disagreement via soft-label multi-task learning , author=. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , year=

    2021

  36. [36]

    arXiv preprint arXiv:2502.06207 , year=

    Is LLM an Overconfident Judge? Unveiling the Capabilities of LLMs in Detecting Offensive Language with Annotation Disagreement , author=. arXiv preprint arXiv:2502.06207 , year=

    work page arXiv

  37. [37]

    arXiv preprint arXiv:2410.14632 , year=

    Diverging Preferences: When do Annotators Disagree and do Models Know? , author=. arXiv preprint arXiv:2410.14632 , year=

    work page arXiv

  38. [38]

    Machine learning , volume=

    Induction of decision trees , author=. Machine learning , volume=. 1986 , publisher=

    1986

  39. [39]

    2017 , publisher=

    Classification and regression trees , author=. 2017 , publisher=

    2017

  40. [40]

    Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining , pages=

    Interpretable decision sets: A joint framework for description and prediction , author=. Proceedings of the 22nd ACM SIGKDD international conference on knowledge discovery and data mining , pages=

  41. [41]

    arXiv preprint arXiv:2412.07992 , year=

    Concept bottleneck large language models , author=. arXiv preprint arXiv:2412.07992 , year=

    work page arXiv

  42. [42]

    The Twelfth International Conference on Learning Representations , year=

    DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines , author=. The Twelfth International Conference on Learning Representations , year=

  43. [43]

    What's In My Human Feedback? Learning Interpretable Descriptions of Preference Data

    What's In My Human Feedback? Learning Interpretable Descriptions of Preference Data , author=. arXiv preprint arXiv:2510.26202 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  44. [44]

    Proceedings of the 2022 conference of the North American chapter of the association for computational linguistics: human language technologies , pages=

    Two contrasting data annotation paradigms for subjective NLP tasks , author=. Proceedings of the 2022 conference of the North American chapter of the association for computational linguistics: human language technologies , pages=

    2022

  45. [45]

    Proceedings of the NAACL student research workshop , pages=

    Hateful symbols or hateful people? predictive features for hate speech detection on twitter , author=. Proceedings of the NAACL student research workshop , pages=

  46. [46]

    doi: 10.18653/v1/2022.naacl-main.431

    Sap, Maarten and Swayamdipta, Swabha and Vianna, Laura and Zhou, Xuhui and Choi, Yejin and Smith, Noah A. Annotators with Attitudes: How Annotator Beliefs And Identities Bias Toxic Language Detection. Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2022. doi:10...

    work page doi:10.18653/v1/2022.naacl-main.431 2022

  47. [47]

    Identifying and Measuring Annotator Bias Based on Annotators ' Demographic Characteristics

    Al Kuwatly, Hala and Wich, Maximilian and Groh, Georg. Identifying and Measuring Annotator Bias Based on Annotators' Demographic Characteristics. Proceedings of the Fourth Workshop on Online Abuse and Harms. 2020. doi:10.18653/v1/2020.alw-1.21

    work page doi:10.18653/v1/2020.alw-1.21 2020

  48. [48]

    Claude’s Constitution , howpublished =

  49. [49]

    OpenAI Model Spec , howpublished =

  50. [50]

    2012 , publisher=

    The concept of law , author=. 2012 , publisher=

    2012

  51. [51]

    2023 , journal=

    Towards Monosemanticity: Decomposing Language Models With Dictionary Learning , author=. 2023 , journal=

    2023

  52. [52]

    Forty-first International Conference on Machine Learning , year=

    Chatbot arena: An open platform for evaluating llms by human preference , author=. Forty-first International Conference on Machine Learning , year=

  53. [53]

    Qwen3 Technical Report

    Qwen3 technical report , author=. arXiv preprint arXiv:2505.09388 , year=

    work page internal anchor Pith review arXiv

  54. [54]

    GPT-4 Technical Report

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    work page internal anchor Pith review arXiv

  55. [55]

    Semeval-2023 task 11: Learning with disagreements (lewidi),

    Leonardelli, Elisa and Abercrombie, Gavin and Almanea, Dina and Basile, Valerio and Fornaciari, Tommaso and Plank, Barbara and Rieser, Verena and Uma, Alexandra and Poesio, Massimo. S em E val-2023 Task 11: Learning with Disagreements ( L e W i D i). Proceedings of the 17th International Workshop on Semantic Evaluation (SemEval-2023). 2023. doi:10.18653/v...

    work page doi:10.18653/v1/2023.semeval-1.314 2023

  56. [56]

    Journal of Artificial Intelligence Research , volume=

    Learning from disagreement: A survey , author=. Journal of Artificial Intelligence Research , volume=

  57. [57]

    The “problem” of human label variation: On ground truth in data, modeling and evaluation.arXiv:2211.02570, 2022

    The'Problem'of Human Label Variation: On Ground Truth in Data, Modeling and Evaluation , author=. arXiv preprint arXiv:2211.02570 , year=

    work page arXiv

  58. [58]

    , author=

    Corpus annotation through crowdsourcing: Towards best practice guidelines. , author=. LREC , pages=

  59. [59]

    Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Learning whom to trust with MACE , author=. Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2013

  60. [60]

    Scientific reports , volume=

    Comparing large Language models and human annotators in latent content analysis of sentiment, political leaning, emotional intensity and sarcasm , author=. Scientific reports , volume=. 2025 , publisher=

    2025

  61. [61]

    arXiv preprint arXiv:2410.17032 , year=

    Insights on disagreement patterns in multimodal safety perception across diverse rater groups , author=. arXiv preprint arXiv:2410.17032 , year=

    work page arXiv

  62. [62]

    Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

    The risk of racial bias in hate speech detection , author=. Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

  63. [63]

    Proceedings of the 2019 conference on human information interaction and retrieval , pages=

    Online hate ratings vary by extremes: A statistical analysis , author=. Proceedings of the 2019 conference on human information interaction and retrieval , pages=

    2019

  64. [64]

    arXiv preprint arXiv:2110.14839 , year=

    Hate speech classifiers learn human-like social stereotypes , author=. arXiv preprint arXiv:2110.14839 , year=

    work page arXiv

  65. [65]

    Challenges and Strategies in Cross-Cultural

    Hershcovich, Daniel and Frank, Stella and Lent, Heather and de Lhoneux, Miryam and Abdou, Mostafa and Brandl, Stephanie and Bugliarello, Emanuele and Cabello Piqueras, Laura and Chalkidis, Ilias and Cui, Ruixiang and Fierro, Constanza and Margatina, Katerina and Rust, Phillip and S gaard, Anders. Challenges and Strategies in Cross-Cultural NLP. Proceeding...

    work page doi:10.18653/v1/2022.acl-long.482 2022

  66. [66]

    Reconsidering Annotator Disagreement about Racist Language: Noise or Signal?

    Larimore, Savannah and Kennedy, Ian and Haskett, Breon and Arseniev-Koehler, Alina. Reconsidering Annotator Disagreement about Racist Language: Noise or Signal?. Proceedings of the Ninth International Workshop on Natural Language Processing for Social Media. 2021. doi:10.18653/v1/2021.socialnlp-1.7

    work page doi:10.18653/v1/2021.socialnlp-1.7 2021

  67. [67]

    International Migration , volume=

    Value differences between refugees and German citizens: insights from a representative survey , author=. International Migration , volume=. 2021 , publisher=

    2021

  68. [68]

    PsyArXiv

    The gab hate corpus: A collection of 27k posts annotated for hate speech , author=. PsyArXiv. July , volume=

  69. [69]

    Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    Linguistically debatable or just plain wrong? , author=. Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

  70. [70]

    PloS one , volume=

    Understanding international perceptions of the severity of harmful content online , author=. PloS one , volume=. 2021 , publisher=

    2021

  71. [71]

    arXiv preprint arXiv:2109.09483 , year=

    ConvAbuse: Data, analysis, and benchmarks for nuanced abuse detection in conversational AI , author=. arXiv preprint arXiv:2109.09483 , year=

    work page arXiv

  72. [72]

    Scientific Reports , volume=

    STELA: a community-centred approach to norm elicitation for AI alignment , author=. Scientific Reports , volume=. 2024 , publisher=

    2024

  73. [73]

    , title =

    Kirk, Hannah Rose and Whitefield, Alexander and Röttger, Paul and Bean, Andrew and Margatina, Katerina and Ciro, Juan and Mosquera, Rafael and Bartolo, Max and Williams, Adina and He, He and Vidgen, Bertie and Hale, Scott A. , title =. 2024 , url =. doi:10.57967/hf/2113 , publisher =

    work page doi:10.57967/hf/2113 2024

  74. [74]

    Building guardrails for large lan- guage models.arXiv preprint arXiv:2402.01822, 2024

    Building guardrails for large language models , author=. arXiv preprint arXiv:2402.01822 , year=

    work page arXiv

  75. [75]

    arXiv preprint arXiv:2406.06369 , year=

    Annotation alignment: Comparing LLM and human annotations of conversational safety , author=. arXiv preprint arXiv:2406.06369 , year=

    work page arXiv

  76. [76]

    Advances in Neural Information Processing Systems , volume=

    Interpreting clip with sparse linear concept embeddings (splice) , author=. Advances in Neural Information Processing Systems , volume=

  77. [77]

    Help Me Help the AI

    Kim, Sunnie S. Y. and Watkins, Elizabeth Anne and Russakovsky, Olga and Fong, Ruth and Monroy-Hern\'. "Help Me Help the AI": Understanding How Explainability Can Support Human-AI Interaction , year =. doi:10.1145/3544548.3581001 , booktitle =

    work page doi:10.1145/3544548.3581001

  78. [78]

    and Kim, Sunnie S

    Ramaswamy, Vikram V. and Kim, Sunnie S. Y. and Fong, Ruth and Russakovsky, Olga , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , month =. 2023 , pages =

    2023

  79. [79]

    Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems , pages=

    Jury learning: Integrating dissenting voices into machine learning models , author=. Proceedings of the 2022 CHI Conference on Human Factors in Computing Systems , pages=

    2022

  80. [80]

    Proceedings of the 2021 chi conference on human factors in computing systems , pages=

    The disagreement deconvolution: Bringing machine learning performance metrics in line with reality , author=. Proceedings of the 2021 chi conference on human factors in computing systems , pages=

    2021

Showing first 80 references.