Grounded Chess Reasoning in Language Models via Master Distillation

Ashton Anderson; Blair Yang; Honghua Dong; Qianfeng Wen; Seth Grief-Albert; Yahya Elgabra; Zhenwei Tang

REVIEW 2 major objections 2 minor 61 references

Distilling an expert chess engine’s full reasoning into natural-language explanations lets a 4B model reach 48.1% accuracy, beat its teacher, and explain its moves.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-13 21:25 UTC pith:5IK5BHX6

load-bearing objection Only the abstract of the chess distillation paper is present; the attached full text is an unrelated HCI manuscript, so the headline claims cannot be checked. the 2 major comments →

arxiv 2603.20510 v2 pith:5IK5BHX6 submitted 2026-03-20 cs.AI

Grounded Chess Reasoning in Language Models via Master Distillation

Zhenwei Tang , Qianfeng Wen , Seth Grief-Albert , Yahya Elgabra , Blair Yang , Honghua Dong , Ashton Anderson This is my paper

classification cs.AI

keywords master distillationchesslanguage modelschain-of-thoughtknowledge distillationreinforcement learningexplainable AIcompact models

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

Language models still fail at grounded reasoning in domains like chess, where labeled training data is scarce but strong expert systems already exist. This paper introduces Master Distillation: convert an opaque expert’s computations into natural-language chain-of-thought traces, then train a compact model on those full reasoning processes rather than only final answers. Their 4B-parameter model C1 rises from a near-zero baseline to 48.1% accuracy, outperforming open-source models and most frontier proprietary systems, surpassing its own distillation teacher, and solving problems in two orders of magnitude fewer tokens. Unlike earlier neural chess systems that only emit best moves, C1 produces step-by-step strategic explanations. The pipeline pairs supervised fine-tuning with reinforcement learning and theme-balanced tactical sampling, and is offered as a general recipe for injecting expert knowledge into small models so reinforcement learning can work in domains where base LLMs start too weak.

Core claim

Capturing an expert chess system’s full reasoning as natural-language chain-of-thought—not merely its final moves—and training a 4B model on those traces with supervised fine-tuning plus theme-balanced reinforcement learning yields C1, which reaches 48.1% accuracy, surpasses its distillation teacher and most frontier systems, and generates faithful strategic explanations in far fewer tokens than baselines.

What carries the argument

Master Distillation: a framework that transforms opaque expert-system computations into transparent natural-language chain-of-thought explanations, then trains a compact language model via supervised fine-tuning and reinforcement learning with theme-balanced data sampling so it acquires both domain skill and grounded, inspectable reasoning.

Load-bearing premise

The natural-language explanations pulled from the expert engine are assumed to be faithful and complete enough that training on them produces real grounded chess reasoning rather than surface patterns or teacher-specific artifacts.

What would settle it

Run the same training pipeline on deliberately incomplete or corrupted reasoning traces and check whether accuracy and explanation quality collapse; if they stay high, the gains are not coming from faithful process distillation.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

Compact models can reach expert-level chess performance without massive scale if they train on full expert reasoning traces rather than move labels alone.
Process distillation can produce student models that beat their teachers on accuracy while using far fewer tokens.
Chess LMs can output inspectable strategic explanations instead of bare best-move predictions.
The same pipeline is proposed as a way to unlock reinforcement learning in other under-optimized domains where base LLMs lack starting competence.
Theme-balanced sampling can deliver broader tactical coverage than uncurated problem sets.

Where Pith is reading between the lines

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

If process distillation generalizes, similar pipelines could bootstrap compact models wherever strong solvers exist but LLM reasoning is weak (e.g., formal methods, specialized medical decision support).
Large token savings imply lower serving cost for interactive tutoring or analysis compared with long chain-of-thought frontier models.
Beating the teacher suggests natural-language traces may regularize or filter noise in the original expert signal—an effect worth measuring directly.
Independent fidelity audits of generated explanations against engine internals would be required before treating the CoT as a true account of computation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Desk Editor's Note

Only the abstract of the chess distillation paper is present; the attached full text is an unrelated HCI manuscript, so the headline claims cannot be checked.

read the letter

The cache is broken for this one. What we actually have is the abstract of Tang et al. on Master Distillation for chess, plus the complete body of an entirely different paper (CARE, arXiv 2603.20511) on reproductive-equity HCI. So every number, baseline, and method detail is currently unverifiable.

What the abstract itself claims is coherent and potentially useful. The idea is to turn an opaque expert engine’s full search/reasoning process into natural-language chain-of-thought, then SFT + theme-balanced RL a 4B model (C1) on those traces. They report 48.1 % accuracy from a near-zero baseline, beating open-source models and most proprietary ones, surpassing the teacher, and using two orders of magnitude fewer tokens while producing strategic explanations rather than bare moves. Distilling full reasoning traces instead of only final moves, plus the theme-balancing step, is a clear methodological step beyond classic engine-to-net move prediction. If the missing body later supplies fidelity checks, open data, and clean evaluation, this would be a practical recipe for any domain that already has a strong solver.

Right now the soft spots are total: no tables, no accuracy definition, no teacher architecture, no fidelity metric for the engine-to-NL conversion, no ablations, no confirmation that the test set is clean. The weakest assumption is exactly the one the reader flagged—whether those natural-language traces are faithful and complete enough to produce genuine grounded reasoning rather than surface patterns or teacher artifacts. We simply cannot assess it.

This is for people working on distillation, tool-use, or domain-specialized LMs who already care about chess or other solver-rich domains. It is not yet for anyone who needs to inspect the evidence. A serious editor should desk-reject or hold until the correct manuscript is supplied; there is nothing for a referee to evaluate. Once the real paper appears, it would almost certainly clear the bar for review. Until then, do not cite and do not schedule for reading group.

Referee Report

2 major / 2 minor

Summary. The abstract of arXiv:2603.20510 claims a general Master Distillation framework that converts opaque expert-system computations into natural-language chain-of-thought traces, then trains compact LMs via SFT plus theme-balanced RL. In chess, a 4B model (C1) is reported to rise from near-zero to 48.1% accuracy, outperforming open-source models and most frontier proprietary systems, surpassing its own distillation teacher, using two orders of magnitude fewer tokens, and producing explainable strategic solutions rather than bare moves. The pipeline is presented as a recipe for injecting expert knowledge into under-optimized domains and unlocking RLVR where base LLM capabilities are insufficient.

Significance. If the empirical claims hold under standard evaluation (held-out tactical/strategic problems, clear accuracy definition, non-overlapping distillation corpus, and demonstrated fidelity of engine-to-NL traces), the work would be a meaningful contribution: it would show that full-process distillation from a strong external expert can produce a compact, explainable chess reasoner that is both more accurate and far more token-efficient than much larger general models. That would be of interest for specialized-domain LM training and for RLVR bootstrapping. However, none of the supporting evidence (tables, baselines, fidelity metrics, data construction, ablations) can be inspected from the material supplied for review.

major comments (2)

The full manuscript text provided under paper_id 2603.20510 is an unrelated work (CARE, Capability Approach for Reproductive Equity in Human-AI Interaction, arXiv:2603.20511). Only the abstract of the chess/Master Distillation paper is available. Consequently every load-bearing claim—48.1% accuracy definition and protocol, teacher architecture and strength, construction and fidelity of natural-language CoT traces, theme-balanced sampling, SFT+RL pipeline, baseline comparisons, token counts, and the claim that C1 surpasses its teacher—cannot be verified against sections, tables, or equations. A technical referee report on the central claims is not possible until the correct manuscript body is supplied.
Abstract-level risk that remains uncheckable: the weakest assumption is that engine-derived NL CoT traces are faithful to the expert computation and sufficiently complete that SFT+RL yields grounded chess reasoning rather than surface pattern matching or teacher-specific artifacts. Without fidelity metrics, overlap analysis between distillation corpus and evaluation set, or ablations on trace quality, the headline accuracy and “surpasses teacher” results cannot be assessed for circularity or leakage.

minor comments (2)

Abstract alone does not define the accuracy metric (move match, puzzle solve rate, mate-in-N, etc.), the evaluation set size/composition, or the precise teacher system; these must appear in the methods and results of the correct manuscript.
The abstract asserts a “general framework” but demonstrates only chess; the correct paper should clarify the scope of the generality claim and any domain-transfer evidence.

Circularity Check

0 steps flagged

No circularity: abstract describes standard expert-to-student distillation evaluated on external chess accuracy; full text supplied is an unrelated HCI paper.

full rationale

Only the abstract of arXiv:2603.20510 is available. It claims an external expert chess system is converted into natural-language chain-of-thought traces, a 4B student is trained via SFT+RL on those traces, and accuracy is measured on chess problems (48.1%). This is ordinary knowledge distillation: the teacher is an independent engine, the student is optimized against its traces, and the reported metric is held-out task accuracy, not a quantity defined by the training objective. No equations, fitted parameters renamed as predictions, uniqueness theorems, or self-citation chains appear. The CACHEABLE PAPER SOURCE CONTEXT instead contains the full text of an unrelated manuscript (CARE, arXiv:2603.20511), a qualitative capability-framework paper with no mathematical derivation of predictions from inputs. Neither document exhibits self-definitional loops, fitted-input-as-prediction, load-bearing self-citation, or renaming of known results as first-principles claims. Per the analyzer rules, honest non-finding is required; score is therefore 0 with empty steps.

Axiom & Free-Parameter Ledger

0 free parameters · 3 axioms · 2 invented entities

With only the abstract, the ledger is necessarily thin. The central claim rests on three unexamined premises: (1) that an external chess engine’s internal computation can be faithfully verbalized as natural-language CoT, (2) that theme-balanced sampling yields comprehensive tactical coverage, and (3) that the reported accuracy metric is a valid proxy for grounded chess reasoning. No free parameters or invented physical entities appear; the invented methodological entities are the distillation pipeline and the C1 model itself.

axioms (3)

domain assumption An expert chess engine’s full search/evaluation process can be converted into natural-language chain-of-thought traces that remain faithful to the original computation.
Load-bearing premise of Master Distillation; never justified in the abstract.
ad hoc to paper Theme-balanced sampling of tactical puzzles produces training data that covers the space of chess reasoning sufficiently for generalization.
Stated as part of the pipeline; no evidence of coverage completeness is given.
domain assumption The accuracy metric used to claim 48.1% and superiority over baselines is a valid measure of grounded chess reasoning.
Central quantitative claim depends on this metric; its definition is absent from the abstract.

invented entities (2)

Master Distillation framework no independent evidence
purpose: General procedure that turns opaque expert-system computations into natural-language CoT for LM training.
Core methodological contribution; independent evidence would require successful application outside chess, which is only promised.
C1 (4B-parameter chess LM) no independent evidence
purpose: Concrete artifact that demonstrates the framework’s efficacy.
New model whose reported performance is the paper’s main empirical claim; no public weights or evaluation harness supplied in the available text.

pith-pipeline@v1.1.0-grok45 · 19536 in / 2653 out tokens · 33582 ms · 2026-07-13T21:25:38.011771+00:00 · methodology

0 comments

read the original abstract

Language models often lack grounded reasoning capabilities in specialized domains where training data is scarce but bespoke systems excel. We introduce a general framework for distilling expert system reasoning into natural language chain-of-thought explanations, enabling compact models to acquire domain expertise and the ability to generate faithful, grounded explanations. Rather than distilling only final outputs, we capture the full reasoning process, transforming opaque expert computations into transparent, step-by-step explanations. We demonstrate this approach in chess, a canonical reasoning domain where language models continue to underperform. Our 4B parameter model, C1, advances from a near-zero baseline to 48.1\% accuracy, outperforming all open-source models and most frontier proprietary systems. Notably, C1 surpasses its distillation teacher and generates solutions in two orders of magnitude fewer tokens than baselines. Unlike prior neural chess approaches that predict only best moves, C1 generates explainable solutions revealing strategic reasoning. Our pipeline combines supervised fine-tuning and reinforcement learning with theme-balanced data sampling for comprehensive tactical coverage. Master Distillation demonstrates how to inject expert-level knowledge into compact models for under-optimized domains, offering a recipe for unlocking RLVR where LLMs lack sufficient base capabilities.

discussion (0)

Reference graph

Works this paper leans on

61 extracted references · 3 canonical work pages

[1]

Saleema Amershi, Dan Weld, Mihaela Vorvoreanu, Adam Fourney, Besmira Nushi, Penny Collisson, Jina Suh, Shamsi Iqbal, Paul N Bennett, Kori Inkpen, et al. 2019. Guidelines for human-AI interaction. InProceedings of the 2019 chi conference on human factors in computing systems. 1–13

2019
[2]

Deborah J Anderson, Jonathan M Bearak, Frances W Grimstad, Thesla Palanee- Phillips, and Ariane van der Straten. 2025. Biomedical innovations in contra- ception: gaps, obstacles, and solutions for sexual and reproductive health.The Lancet406, 10515 (2025), 2119–2132

2025
[3]

Anthropic. 2024. Claude 4.5. Large language model. Available at: https: //www.anthropic.com/claude

2024
[4]

Erika Bonnevie, Tiffany D Lloyd, Sarah D Rosenberg, Kara Williams, Jaclyn Gold- barg, and Joe Smyser. 2021. Layla’s got you: developing a tailored contraception chatbot for Black and Hispanic young women.Health Education Journal80, 4 (2021), 413–424

2021
[5]

Anjun Chen and Drake O Chen. 2023. Accuracy of chatbots in citing journal articles.JAMA Network Open6, 8 (2023), e2327647–e2327647

2023
[6]

Catherine L Clelland, Stuart Moss, and James D Clelland. 2024. Warning: Artificial intelligence chatbots can generate inaccurate medical and scientific information and references.Exploration of Digital Health Technologies2, 1 (2024), 1–6

2024
[7]

Kya family planning after marriage hoti hai?

Roshini Deva, Dhruv Ramani, Tanvi Divate, Suhani Jalota, and Azra Ismail. 2025. "Kya family planning after marriage hoti hai?": Integrating Cultural Sensitivity in an LLM Chatbot for Reproductive Health. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article...

arXiv 2025
[8]

Google. 2026. Gemini 3 Flash. https://gemini.google.com/. Accessed: 2026-01-21

2026
[9]

Google. 2026. Google AI Mode (January 2026 version). https://www.google.com/

2026
[10]

Google. 2026. Google AI Overview (January 2026 version). https://www.google. com/

2026
[11]

Girl, I’m so Serious

Connie Guan, Anya Bouzida, Ramzy M Oncy-Avila, Sanika Moharana, and Laurel D Riek. 2021. Taking an (embodied) cue from community health: Designing 6 “Girl, I’m so Serious”: CARE, a Capability Framework for Reproductive Equity in Human-AI Interaction dementia caregiver support technology to advance health equity. InProceedings of the 2021 CHI Conference on...

2021
[12]

Calum Handforth and Kecia Bertermann. 2018. How Girl Effect built a chatbot. Girl Effect(2018)

2018
[13]

It’s kind of like code-switching

Christina N Harrington, Radhika Garg, Amanda Woodward, and Dimitri Williams. 2022. “It’s kind of like code-switching”: Black older adults’ expe- riences with a voice assistant for health information seeking. InProceedings of the 2022 CHI Conference on Human Factors in Computing Systems. 1–15

2022
[14]

Emma Harvey, Emily Sheng, Su Lin Blodgett, Alexandra Chouldechova, Jean Garcia-Gathright, Alexandra Olteanu, and Hanna Wallach. 2025. Understand- ing and Meeting Practitioner Needs When Measuring Representational Harms Caused by LLM-Based Systems.arXiv preprint arXiv:2506.04482(2025)

Pith/arXiv arXiv 2025
[15]

Nchebe-Jah Iloanusi and Soon Ae Chun. 2024. AI impact on health equity for marginalized, racial, and ethnic minorities. InProceedings of the 25th annual international conference on digital government research. 841–848

2024
[16]

Minal Jain and Pradeep Yammiyavar. 2015. Game based learning tool seeking peer support for empowering adolescent girls in rural Assam. InProceedings of the 14th International Conference on Interaction Design and Children(Boston, Massachusetts)(IDC ’15). Association for Computing Machinery, New York, NY, USA, 275–278. doi:10.1145/2771839.2771895

doi:10.1145/2771839.2771895 2015
[17]

Eunkyung Jo, Yuin Jeong, SoHyun Park, Daniel A Epstein, and Young-Ho Kim
[18]

InPro- ceedings of the 2024 CHI Conference on Human Factors in Computing Systems

Understanding the impact of long-term memory on self-disclosure with large language model-driven chatbots for public health intervention. InPro- ceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–21

2024
[19]

Aman Khullar, Nikhil Nalin, Abhishek Prasad, Ann John Mampilli, and Neha Kumar. 2025. Nurturing Capabilities: Unpacking the Gap in Human-Centered Evaluations of AI-Based Systems. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–18

2025
[20]

Duri Long and Brian Magerko. 2020. What is AI Literacy? Competencies and Design Considerations. InProceedings of the 2020 CHI Conference on Human Factors in Computing Systems(Honolulu, HI, USA)(CHI ’20). Association for Computing Machinery, New York, NY, USA, 1–16. doi:10.1145/3313831.3376727

doi:10.1145/3313831.3376727 2020
[21]

Gabriel Edgar Villagomez Martinez, Clara del Carmen Flores-Acosta, Juan Anto- nio Soria López, Abel Guzman Lopez, Nubia Alejandra Wong Arce, Fernanda Isabel Velazquez Vega, et al. 2025. 105. Sexual and Reproductive Health Education for Adolescents: Opportunities and Challenges in the Social Media Era.Journal of Pediatric and Adolescent Gynecology38, 2 (2025), 281

2025
[22]

Giselle Lorrane Nobre Melo, Nicoly Da Silva Menezes, Ingrid Moreira Miranda Da Silva, Luciano Arruda Teran, and Marcelle Pereira Mota. 2024. Inspecting the Accessibility of Instant Payment Systems from the Perspective of Low Literacy People. InProceedings of the XXII Brazilian Symposium on Human Factors in Com- puting Systems(Maceió, Brazil)(IHC ’23). Ass...

work page doi:10.1145/3638067.3638127 2024
[23]

Microsoft. 2025. Copilot (Smart Mode). https://copilot.microsoft.com

2025
[24]

Rhiana Mills, Emily Rose Mangone, Neal Lesh, Gayatri Jayal, Diwakar Mohan, and Paula Baraitser. 2024. Chatbots that deliver contraceptive support: systematic review.Journal of medical Internet research26 (2024), e46758

2024
[25]

But I Won’t Say That It Was Bad Seeing a Real Vagina

Sara Moin, Manshul Belani, Pragya Singh, Nishtha Phutela, and Pushpendra Singh. 2025. "But I Won’t Say That It Was Bad Seeing a Real Vagina": Understand- ing Perspectives toward Learning Sensitive-Critical Health Topic. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, ...

doi:10.1145/3706598.3713807 2025
[26]

Jakob Nielsen. 10. Usability heuristics for user interface design
[27]

2000.Women and human development: The capabilities approach

Martha C Nussbaum. 2000.Women and human development: The capabilities approach. Vol. 3. Cambridge university press

2000
[28]

OpenAI. 2025. Preparedness Framework. Version 2. Last updated: April 15, 2025

2025
[29]

OpenAI. 2026. ChatGPT (GPT-5.2). https://chat.openai.com/. Accessed: January 21, 2026

2026
[30]

2025.Health Equity

World Health Organization. 2025.Health Equity. Retrieved October 25, 2025 from https://www.who.int/health-topics/health-equity#tab=tab_1

2025
[31]

Andrea G Parker, Laura M Vardoulakis, Jatin Alla, and Christina N Harrington
[32]

In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems

Participatory AI Considerations for Advancing Racial Health Equity. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–24

2025
[33]

Planned Parenthood Bot. [n. d.]. Planned Parenthood Case Study — work.co. https://work.co/clients/planned-parenthood/
[34]

Rishadur Rahman, Nafis Irtiza Tripto, Mohammed Eunus Ali, Sajid Hasan Apon, and Rifat Shahriyar

Rifat Rahman, Md. Rishadur Rahman, Nafis Irtiza Tripto, Mohammed Eunus Ali, Sajid Hasan Apon, and Rifat Shahriyar. 2021. AdolescentBot: Understanding Opportunities for Chatbots in Combating Adolescent Sexual and Reproductive Health Problems in Bangladesh. InProceedings of the 2021 CHI Conference on Human Factors in Computing Systems(Yokohama, Japan)(CHI ’...

arXiv 2021
[35]

2017.Wellbeing, freedom and social justice

Ingrid Robeyns. 2017.Wellbeing, freedom and social justice. Open Book, Cam- bridge, England

2017
[36]

2017.Wellbeing, freedom and social justice: The capability ap- proach re-examined

Ingrid Robeyns. 2017.Wellbeing, freedom and social justice: The capability ap- proach re-examined. Open book publishers

2017
[37]

Andrew D Selbst, Danah Boyd, Sorelle A Friedler, Suresh Venkatasubramanian, and Janet Vertesi. 2019. Fairness and abstraction in sociotechnical systems. In Proceedings of the conference on fairness, accountability, and transparency. 59–68

2019
[38]

Amartya Sen. 1993. Capability and well-being73.The quality of life30 (1993), 270–293

1993
[39]

1995.Inequality reexamined

Amartya Sen. 1995.Inequality reexamined. Harvard university press

1995
[40]

Amartya Sen. 1999. Commodities and capabilities.OUP Catalogue(1999)

1999
[41]

1999.Development as freedom

Amartya Sen. 1999.Development as freedom. Anchor

1999
[42]

Amartya Sen, S Anand, and F Peter. 2004. Why health equity? (2004)

2004
[43]

Melanie Shackleford, Anna Horvath, Mayra Repetto, Andrea Thi, Rory Twells, Maggie Sanders, Stephanie Fernandez, Dale Netski, Kavita Batra, Nadia Gomez, et al. 2024. An analysis of oral contraceptive related videos on TikTok.AJOG Global Reports4, 3 (2024), 100364

2024
[44]

Anika Sharma, Malavika Mampally, Chidaksh Ravuru, Kandyce Brennan, and Neil Gaikwad. 2025. Can LLMs Understand What We Cannot Say? Measuring Multilevel Alignment Through Abortion Stigma Across Cognitive, Interpersonal, and Structural Levels.arXiv preprint arXiv:2512.13142(2025)

arXiv 2025
[45]

Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. 2023. A long way to go: Investigating length correlations in rlhf.arXiv preprint arXiv:2310.03716 (2023)

Pith/arXiv arXiv 2023
[46]

Piya Sorcar, Benjamin Strauber, Prashant Loyalka, Neha Kumar, and Shelley Goldman. 2017. Sidestepping the Elephant in the Classroom: Using Culturally Localized Technology To Teach Around Taboos. InProceedings of the 2017 CHI Conference on Human Factors in Computing Systems(Denver, Colorado, USA) (CHI ’17). Association for Computing Machinery, New York, NY...

doi:10.1145/3025453.3025958 2017
[47]

Who Knows? Maybe it Really Works

Huiyun Tang, Gabriele Lenzini, Samuel Greiff, Björn Rohles, and Anastasia Sergeeva. 2024. “Who Knows? Maybe it Really Works”: Analysing Users’ Per- ceptions of Health Misinformation on Social Media. InProceedings of the 2024 ACM Designing Interactive Systems Conference. 1499–1517

2024
[48]

Bonnie Tran and Lee Na Choi. 2018. Menstrual Maze: A Toy Exploring Public Engagement in Menstrual Health Education. InExtended Abstracts of the 2018 CHI Conference on Human Factors in Computing Systems(Montreal QC, Canada) (CHI EA ’18). Association for Computing Machinery, New York, NY, USA, 1–6. doi:10.1145/3170427.3180649

doi:10.1145/3170427.3180649 2018
[49]

Anupriya Tuli, Shaan Chopra, Neha Kumar, and Pushpendra Singh. 2018. Learn- ing from and with Menstrupedia: Towards Menstrual Health Education in India. Proc. ACM Hum.-Comput. Interact.2, CSCW, Article 174 (Nov. 2018), 20 pages. doi:10.1145/3274443

work page doi:10.1145/3274443 2018
[50]

It’s a girl thing

Anupriya Tuli, Shruti Dalvi, Neha Kumar, and Pushpendra Singh. 2019. “It’s a girl thing”: Examining Challenges and Opportunities around Menstrual Health Education in India.ACM Trans. Comput.-Hum. Interact.26, 5, Article 29 (July 2019), 24 pages. doi:10.1145/3325282

work page doi:10.1145/3325282 2019
[51]

Anupriya Tuli, Azra Ismail, Karthik S Bhat, Pushpendra Singh, and Neha Kumar
[52]

Information-Backward but Sex-Forward

“Information-Backward but Sex-Forward”: Navigating Masculinity towards Intimate Wellbeing and Heterosexual Relationships. InProceedings of the 2023 CHI Conference on Human Factors in Computing Systems(Hamburg, Germany) (CHI ’23). Association for Computing Machinery, New York, NY, USA, Article 39, 16 pages. doi:10.1145/3544548.3581478

doi:10.1145/3544548.3581478 2023
[53]

UNESCO. 2023. Comprehensive sexuality education: For healthy, informed and empowered learners

2023
[54]

UNESCO. 2024. Comprehensive Sexuality Education. https://www.unesco.org/ en/health-education/cse. Accessed January 2026

2024
[55]

Hanna Wallach, Meera Desai, A Feder Cooper, Angelina Wang, Chad Atalla, Solon Barocas, Su Lin Blodgett, Alexandra Chouldechova, Emily Corvi, P Alex Dow, et al. 2025. Position: Evaluating generative ai systems is a social science measurement challenge.arXiv preprint arXiv:2502.00561(2025)

Pith/arXiv arXiv 2025
[56]

William H Walters and Esther Isabelle Wilder. 2023. Fabrication and errors in the bibliographic citations generated by ChatGPT.Scientific Reports13, 1 (2023), 14045

2023
[57]

Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, et al
[58]

InProceedings of the 2022 ACM conference on fairness, accountability, and transparency

Taxonomy of risks posed by language models. InProceedings of the 2022 ACM conference on fairness, accountability, and transparency. 214–229

2022
[59]

Wai Lok Woo, Bin Gao, Raid Rafi Omar Al-Nima, and Wing-Kuen Ling. 2020. Development of conversational artificial intelligence for pandemic healthcare query support.International Journal of Automation, Artificial Intelligence and Machine Learning1, 1 (2020), 54–79

2020
[60]

Jenny Wu, Esmé Trahair, Megan Happ, and Jonas Swartz. 2023. TikTok,# IUD, and user experience with intrauterine devices reported on social media.Obstetrics & Gynecology141, 1 (2023), 215–217

2023
[61]

Girl, I’m so Serious

Wenting Zhao, Xiang Ren, Jack Hessel, Claire Cardie, Yejin Choi, and Yuntian Deng. 2024. Wildchat: 1m chatgpt interaction logs in the wild.arXiv preprint arXiv:2405.01470(2024). 7 Zhong*, Chen*, Sharma*, Brennan, and Gaikwad A Capability Approach Modules A-Module: The Non-optional Core A1. Functionings and capabilities as core concepts:CARE distinguishes ...

Pith/arXiv arXiv 2024

[1] [1]

Saleema Amershi, Dan Weld, Mihaela Vorvoreanu, Adam Fourney, Besmira Nushi, Penny Collisson, Jina Suh, Shamsi Iqbal, Paul N Bennett, Kori Inkpen, et al. 2019. Guidelines for human-AI interaction. InProceedings of the 2019 chi conference on human factors in computing systems. 1–13

2019

[2] [2]

Deborah J Anderson, Jonathan M Bearak, Frances W Grimstad, Thesla Palanee- Phillips, and Ariane van der Straten. 2025. Biomedical innovations in contra- ception: gaps, obstacles, and solutions for sexual and reproductive health.The Lancet406, 10515 (2025), 2119–2132

2025

[3] [3]

Anthropic. 2024. Claude 4.5. Large language model. Available at: https: //www.anthropic.com/claude

2024

[4] [4]

Erika Bonnevie, Tiffany D Lloyd, Sarah D Rosenberg, Kara Williams, Jaclyn Gold- barg, and Joe Smyser. 2021. Layla’s got you: developing a tailored contraception chatbot for Black and Hispanic young women.Health Education Journal80, 4 (2021), 413–424

2021

[5] [5]

Anjun Chen and Drake O Chen. 2023. Accuracy of chatbots in citing journal articles.JAMA Network Open6, 8 (2023), e2327647–e2327647

2023

[6] [6]

Catherine L Clelland, Stuart Moss, and James D Clelland. 2024. Warning: Artificial intelligence chatbots can generate inaccurate medical and scientific information and references.Exploration of Digital Health Technologies2, 1 (2024), 1–6

2024

[7] [7]

Kya family planning after marriage hoti hai?

Roshini Deva, Dhruv Ramani, Tanvi Divate, Suhani Jalota, and Azra Ismail. 2025. "Kya family planning after marriage hoti hai?": Integrating Cultural Sensitivity in an LLM Chatbot for Reproductive Health. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, NY, USA, Article...

arXiv 2025

[8] [8]

Google. 2026. Gemini 3 Flash. https://gemini.google.com/. Accessed: 2026-01-21

2026

[9] [9]

Google. 2026. Google AI Mode (January 2026 version). https://www.google.com/

2026

[10] [10]

Google. 2026. Google AI Overview (January 2026 version). https://www.google. com/

2026

[11] [11]

Girl, I’m so Serious

Connie Guan, Anya Bouzida, Ramzy M Oncy-Avila, Sanika Moharana, and Laurel D Riek. 2021. Taking an (embodied) cue from community health: Designing 6 “Girl, I’m so Serious”: CARE, a Capability Framework for Reproductive Equity in Human-AI Interaction dementia caregiver support technology to advance health equity. InProceedings of the 2021 CHI Conference on...

2021

[12] [12]

Calum Handforth and Kecia Bertermann. 2018. How Girl Effect built a chatbot. Girl Effect(2018)

2018

[13] [13]

It’s kind of like code-switching

Christina N Harrington, Radhika Garg, Amanda Woodward, and Dimitri Williams. 2022. “It’s kind of like code-switching”: Black older adults’ expe- riences with a voice assistant for health information seeking. InProceedings of the 2022 CHI Conference on Human Factors in Computing Systems. 1–15

2022

[14] [14]

Emma Harvey, Emily Sheng, Su Lin Blodgett, Alexandra Chouldechova, Jean Garcia-Gathright, Alexandra Olteanu, and Hanna Wallach. 2025. Understand- ing and Meeting Practitioner Needs When Measuring Representational Harms Caused by LLM-Based Systems.arXiv preprint arXiv:2506.04482(2025)

Pith/arXiv arXiv 2025

[15] [15]

Nchebe-Jah Iloanusi and Soon Ae Chun. 2024. AI impact on health equity for marginalized, racial, and ethnic minorities. InProceedings of the 25th annual international conference on digital government research. 841–848

2024

[16] [16]

Minal Jain and Pradeep Yammiyavar. 2015. Game based learning tool seeking peer support for empowering adolescent girls in rural Assam. InProceedings of the 14th International Conference on Interaction Design and Children(Boston, Massachusetts)(IDC ’15). Association for Computing Machinery, New York, NY, USA, 275–278. doi:10.1145/2771839.2771895

doi:10.1145/2771839.2771895 2015

[17] [17]

Eunkyung Jo, Yuin Jeong, SoHyun Park, Daniel A Epstein, and Young-Ho Kim

[18] [18]

InPro- ceedings of the 2024 CHI Conference on Human Factors in Computing Systems

Understanding the impact of long-term memory on self-disclosure with large language model-driven chatbots for public health intervention. InPro- ceedings of the 2024 CHI Conference on Human Factors in Computing Systems. 1–21

2024

[19] [19]

Aman Khullar, Nikhil Nalin, Abhishek Prasad, Ann John Mampilli, and Neha Kumar. 2025. Nurturing Capabilities: Unpacking the Gap in Human-Centered Evaluations of AI-Based Systems. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–18

2025

[20] [20]

Duri Long and Brian Magerko. 2020. What is AI Literacy? Competencies and Design Considerations. InProceedings of the 2020 CHI Conference on Human Factors in Computing Systems(Honolulu, HI, USA)(CHI ’20). Association for Computing Machinery, New York, NY, USA, 1–16. doi:10.1145/3313831.3376727

doi:10.1145/3313831.3376727 2020

[21] [21]

Gabriel Edgar Villagomez Martinez, Clara del Carmen Flores-Acosta, Juan Anto- nio Soria López, Abel Guzman Lopez, Nubia Alejandra Wong Arce, Fernanda Isabel Velazquez Vega, et al. 2025. 105. Sexual and Reproductive Health Education for Adolescents: Opportunities and Challenges in the Social Media Era.Journal of Pediatric and Adolescent Gynecology38, 2 (2025), 281

2025

[22] [22]

Giselle Lorrane Nobre Melo, Nicoly Da Silva Menezes, Ingrid Moreira Miranda Da Silva, Luciano Arruda Teran, and Marcelle Pereira Mota. 2024. Inspecting the Accessibility of Instant Payment Systems from the Perspective of Low Literacy People. InProceedings of the XXII Brazilian Symposium on Human Factors in Com- puting Systems(Maceió, Brazil)(IHC ’23). Ass...

work page doi:10.1145/3638067.3638127 2024

[23] [23]

Microsoft. 2025. Copilot (Smart Mode). https://copilot.microsoft.com

2025

[24] [24]

Rhiana Mills, Emily Rose Mangone, Neal Lesh, Gayatri Jayal, Diwakar Mohan, and Paula Baraitser. 2024. Chatbots that deliver contraceptive support: systematic review.Journal of medical Internet research26 (2024), e46758

2024

[25] [25]

But I Won’t Say That It Was Bad Seeing a Real Vagina

Sara Moin, Manshul Belani, Pragya Singh, Nishtha Phutela, and Pushpendra Singh. 2025. "But I Won’t Say That It Was Bad Seeing a Real Vagina": Understand- ing Perspectives toward Learning Sensitive-Critical Health Topic. InProceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ’25). Association for Computing Machinery, New York, ...

doi:10.1145/3706598.3713807 2025

[26] [26]

Jakob Nielsen. 10. Usability heuristics for user interface design

[27] [27]

2000.Women and human development: The capabilities approach

Martha C Nussbaum. 2000.Women and human development: The capabilities approach. Vol. 3. Cambridge university press

2000

[28] [28]

OpenAI. 2025. Preparedness Framework. Version 2. Last updated: April 15, 2025

2025

[29] [29]

OpenAI. 2026. ChatGPT (GPT-5.2). https://chat.openai.com/. Accessed: January 21, 2026

2026

[30] [30]

2025.Health Equity

World Health Organization. 2025.Health Equity. Retrieved October 25, 2025 from https://www.who.int/health-topics/health-equity#tab=tab_1

2025

[31] [31]

Andrea G Parker, Laura M Vardoulakis, Jatin Alla, and Christina N Harrington

[32] [32]

In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems

Participatory AI Considerations for Advancing Racial Health Equity. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems. 1–24

2025

[33] [33]

Planned Parenthood Bot. [n. d.]. Planned Parenthood Case Study — work.co. https://work.co/clients/planned-parenthood/

[34] [34]

Rishadur Rahman, Nafis Irtiza Tripto, Mohammed Eunus Ali, Sajid Hasan Apon, and Rifat Shahriyar

Rifat Rahman, Md. Rishadur Rahman, Nafis Irtiza Tripto, Mohammed Eunus Ali, Sajid Hasan Apon, and Rifat Shahriyar. 2021. AdolescentBot: Understanding Opportunities for Chatbots in Combating Adolescent Sexual and Reproductive Health Problems in Bangladesh. InProceedings of the 2021 CHI Conference on Human Factors in Computing Systems(Yokohama, Japan)(CHI ’...

arXiv 2021

[35] [35]

2017.Wellbeing, freedom and social justice

Ingrid Robeyns. 2017.Wellbeing, freedom and social justice. Open Book, Cam- bridge, England

2017

[36] [36]

2017.Wellbeing, freedom and social justice: The capability ap- proach re-examined

Ingrid Robeyns. 2017.Wellbeing, freedom and social justice: The capability ap- proach re-examined. Open book publishers

2017

[37] [37]

Andrew D Selbst, Danah Boyd, Sorelle A Friedler, Suresh Venkatasubramanian, and Janet Vertesi. 2019. Fairness and abstraction in sociotechnical systems. In Proceedings of the conference on fairness, accountability, and transparency. 59–68

2019

[38] [38]

Amartya Sen. 1993. Capability and well-being73.The quality of life30 (1993), 270–293

1993

[39] [39]

1995.Inequality reexamined

Amartya Sen. 1995.Inequality reexamined. Harvard university press

1995

[40] [40]

Amartya Sen. 1999. Commodities and capabilities.OUP Catalogue(1999)

1999

[41] [41]

1999.Development as freedom

Amartya Sen. 1999.Development as freedom. Anchor

1999

[42] [42]

Amartya Sen, S Anand, and F Peter. 2004. Why health equity? (2004)

2004

[43] [43]

Melanie Shackleford, Anna Horvath, Mayra Repetto, Andrea Thi, Rory Twells, Maggie Sanders, Stephanie Fernandez, Dale Netski, Kavita Batra, Nadia Gomez, et al. 2024. An analysis of oral contraceptive related videos on TikTok.AJOG Global Reports4, 3 (2024), 100364

2024

[44] [44]

Anika Sharma, Malavika Mampally, Chidaksh Ravuru, Kandyce Brennan, and Neil Gaikwad. 2025. Can LLMs Understand What We Cannot Say? Measuring Multilevel Alignment Through Abortion Stigma Across Cognitive, Interpersonal, and Structural Levels.arXiv preprint arXiv:2512.13142(2025)

arXiv 2025

[45] [45]

Prasann Singhal, Tanya Goyal, Jiacheng Xu, and Greg Durrett. 2023. A long way to go: Investigating length correlations in rlhf.arXiv preprint arXiv:2310.03716 (2023)

Pith/arXiv arXiv 2023

[46] [46]

Piya Sorcar, Benjamin Strauber, Prashant Loyalka, Neha Kumar, and Shelley Goldman. 2017. Sidestepping the Elephant in the Classroom: Using Culturally Localized Technology To Teach Around Taboos. InProceedings of the 2017 CHI Conference on Human Factors in Computing Systems(Denver, Colorado, USA) (CHI ’17). Association for Computing Machinery, New York, NY...

doi:10.1145/3025453.3025958 2017

[47] [47]

Who Knows? Maybe it Really Works

Huiyun Tang, Gabriele Lenzini, Samuel Greiff, Björn Rohles, and Anastasia Sergeeva. 2024. “Who Knows? Maybe it Really Works”: Analysing Users’ Per- ceptions of Health Misinformation on Social Media. InProceedings of the 2024 ACM Designing Interactive Systems Conference. 1499–1517

2024

[48] [48]

Bonnie Tran and Lee Na Choi. 2018. Menstrual Maze: A Toy Exploring Public Engagement in Menstrual Health Education. InExtended Abstracts of the 2018 CHI Conference on Human Factors in Computing Systems(Montreal QC, Canada) (CHI EA ’18). Association for Computing Machinery, New York, NY, USA, 1–6. doi:10.1145/3170427.3180649

doi:10.1145/3170427.3180649 2018

[49] [49]

Anupriya Tuli, Shaan Chopra, Neha Kumar, and Pushpendra Singh. 2018. Learn- ing from and with Menstrupedia: Towards Menstrual Health Education in India. Proc. ACM Hum.-Comput. Interact.2, CSCW, Article 174 (Nov. 2018), 20 pages. doi:10.1145/3274443

work page doi:10.1145/3274443 2018

[50] [50]

It’s a girl thing

Anupriya Tuli, Shruti Dalvi, Neha Kumar, and Pushpendra Singh. 2019. “It’s a girl thing”: Examining Challenges and Opportunities around Menstrual Health Education in India.ACM Trans. Comput.-Hum. Interact.26, 5, Article 29 (July 2019), 24 pages. doi:10.1145/3325282

work page doi:10.1145/3325282 2019

[51] [51]

Anupriya Tuli, Azra Ismail, Karthik S Bhat, Pushpendra Singh, and Neha Kumar

[52] [52]

Information-Backward but Sex-Forward

“Information-Backward but Sex-Forward”: Navigating Masculinity towards Intimate Wellbeing and Heterosexual Relationships. InProceedings of the 2023 CHI Conference on Human Factors in Computing Systems(Hamburg, Germany) (CHI ’23). Association for Computing Machinery, New York, NY, USA, Article 39, 16 pages. doi:10.1145/3544548.3581478

doi:10.1145/3544548.3581478 2023

[53] [53]

UNESCO. 2023. Comprehensive sexuality education: For healthy, informed and empowered learners

2023

[54] [54]

UNESCO. 2024. Comprehensive Sexuality Education. https://www.unesco.org/ en/health-education/cse. Accessed January 2026

2024

[55] [55]

Hanna Wallach, Meera Desai, A Feder Cooper, Angelina Wang, Chad Atalla, Solon Barocas, Su Lin Blodgett, Alexandra Chouldechova, Emily Corvi, P Alex Dow, et al. 2025. Position: Evaluating generative ai systems is a social science measurement challenge.arXiv preprint arXiv:2502.00561(2025)

Pith/arXiv arXiv 2025

[56] [56]

William H Walters and Esther Isabelle Wilder. 2023. Fabrication and errors in the bibliographic citations generated by ChatGPT.Scientific Reports13, 1 (2023), 14045

2023

[57] [57]

Laura Weidinger, Jonathan Uesato, Maribeth Rauh, Conor Griffin, Po-Sen Huang, John Mellor, Amelia Glaese, Myra Cheng, Borja Balle, Atoosa Kasirzadeh, et al

[58] [58]

InProceedings of the 2022 ACM conference on fairness, accountability, and transparency

Taxonomy of risks posed by language models. InProceedings of the 2022 ACM conference on fairness, accountability, and transparency. 214–229

2022

[59] [59]

Wai Lok Woo, Bin Gao, Raid Rafi Omar Al-Nima, and Wing-Kuen Ling. 2020. Development of conversational artificial intelligence for pandemic healthcare query support.International Journal of Automation, Artificial Intelligence and Machine Learning1, 1 (2020), 54–79

2020

[60] [60]

Jenny Wu, Esmé Trahair, Megan Happ, and Jonas Swartz. 2023. TikTok,# IUD, and user experience with intrauterine devices reported on social media.Obstetrics & Gynecology141, 1 (2023), 215–217

2023

[61] [61]

Girl, I’m so Serious

Wenting Zhao, Xiang Ren, Jack Hessel, Claire Cardie, Yejin Choi, and Yuntian Deng. 2024. Wildchat: 1m chatgpt interaction logs in the wild.arXiv preprint arXiv:2405.01470(2024). 7 Zhong*, Chen*, Sharma*, Brennan, and Gaikwad A Capability Approach Modules A-Module: The Non-optional Core A1. Functionings and capabilities as core concepts:CARE distinguishes ...

Pith/arXiv arXiv 2024