Source-linked AI summary
Non-portability of Algorithmic Fairness in India
Nithya Sambasivan, Erin Arnesen, Ben Hutchinson, Vinodkumar Prabhakaran
TL;DR
The paper asks how well Western-rooted algorithmic fairness assumptions apply in India, where institutions, infrastructures, and values differ. Drawing on expert interviews and analysis of deployments and policies, it identifies challenges spanning data and models, ML makers, and AI aspirations. It argues that fairness work must engage local structures through recontextualising, empowering, and enabling.
Problem
Western-rooted fairness research draws on particular groups, datasets, measures, and legal tenets, while its assumptions about institutions and infrastructure may not generalize to India.
Source: §1 · p. 1Method
The paper synthesizes 36 interviews with Indian researchers and activists and analysis of algorithmic deployments and policies in India.
Source: §2 · p. 2Results
The study identifies flawed data and model assumptions, ML makers’ distance, and AI aspiration as challenge clusters amid large model–community distance in India.
Source: §3 · p. 2Takeaways & Limitations
The authors call for fairness research that engages local structures through recontextualising, empowering communities, and enabling ecosystems.
Source: §1 · p. 2, §4 · p. 3, §4.2 · p. 4
Abstract
from arXiv · showhide
Conventional algorithmic fairness is Western in its sub-groups, values, and optimizations. In this paper, we ask how portable the assumptions of this largely Western take on algorithmic fairness are to a different geo-cultural context such as India. Based on 36 expert interviews with Indian scholars, and an analysis of emerging algorithmic deployments in India, we identify three clusters of challenges that engulf the large distance between machine learning models and oppressed communities in India. We argue that a mere translation of technical fairness work to Indian subgroups may serve only as a window dressing, and instead, call for a collective re-imagining of Fair-ML, by re-contextualising data and models, empowering oppressed communities, and more importantly, enabling ecosystems.
1 Introduction
Conventional Fair-ML is rooted in Western concerns, datasets, measures, and legal tenets, yet its assumptions are increasingly treated as universal. The paper asks whether those assumptions travel to India, where relevant institutions and infrastructures may differ.
1 Introduction
- Western fairness research centers Western injustices, datasets, measurement scales, and legal tenets, while its assumptions about institutions and infrastructure may not generalize to non-Western countries.Examples include race and gender, ImageNet, the Fitzpatrick scale, and equal opportunity.Source: §1 · p. 1
- In the US, facial-recognition fairness failures led to bans and moratoria amid supportive research, public data and laws, responsive industry, government involvement, and media scrutiny.The paper notes that such factors may be absent or irrelevant elsewhere, raising questions about fairness mechanisms and local social, economic, and infrastructural conditions.Source: §1 · p. 1
- The position paper combines interviews with 36 Indian scholars and activists with analysis of emerging algorithmic deployments, addressing a gap in fairness discussions for India's large, diverse population.The introduction identifies public-sector applications including predictive policing, facial recognition, and agriculture.Source: §1 · p. 1
- The authors argue that optimizing model fairness alone may be tokenistic where distance between models and oppressed communities is large, and call for recontextualising, empowering, and enabling.They frame the response as collective responsibility for interdisciplinary Fair-ML research.Source: §1 · p. 2
2 Methodology
The study synthesizes expert interviews with analysis of algorithmic deployments and policies in India. Its respondents span disciplines and social-justice areas, and the deployment and policy review begins in 2009.
2 Methodology
- The study synthesizes 36 interviews with researchers and activists and observations of algorithmic deployments and policies in India.The interviewees worked across computer science, activism, law and public policy, STS, development economics, sociology, and journalism.Source: §2 · p. 2
- All respondents had 10–30 years of social-justice experience in India, including work on caste, gender, labour, disability, surveillance, health, and constitutional rights.The disciplines represented were Computer Science (11), Activism (9), Law and Public Policy (6), Science and Technology Studies (5), Development Economics (2), Sociology (2), and Journalism (1).Source: §2 · p. 2
- The deployment and policy analysis began in 2009 and drew on news publications, policy documents, grassroots forums, and prior research on automation in India.The authors direct readers to Appendix A for the collection, categorization, and synthesis process.Source: §2 · p. 2
3 Summary of Findings
The study identifies three clusters of challenges that widen the distance between models deployed in India and the communities they affect. Together, they show why localizing model outputs alone may backfire.
3 Summary of Findings
- The study finds that conventional fairness approaches may be inappropriate, insufficient, or inimical in India when they fail to engage with local structures.The authors describe model–community distance across technical, social, ethical, temporal, and physical dimensions; Figure 1 frames the clusters through this distance.Source: §3 · p. 2, Figure 1 · p. 2
- Data and model assumptions can fail through infrastructural distortions, incomplete data favoring middle-class men, differing discrimination proxies, and challenges posed by reservations and quotas.The paper notes that 50% of people lack internet access, especially women and lower-caste people.Source: §3 · p. 2
- Respondents describe Indian users as agency-less data subjects, with poor recourse and redress for people disempowered by caste, religion, or gender.Human infrastructures such as street-side bureaucrats and frontline workers are removed in ML systems, and respondents also identify limited workforce diversity.Source: §3 · p. 2
- AI is aspirationally viewed in India, with high-stakes reliance and trust alongside limited transparency and few empowered actors to question algorithmic interventions.Respondents point to a lack of journalists, activists, and researchers serving as an accountability buffer; AI's neutral and human-free associations lend credence to its authority.Source: §3 · p. 2
4 Towards Algorithmic Fairness in India
The paper proposes three pathways for addressing the distance between AI systems and communities in India: recontextualising data and models, empowering communities, and enabling accountability ecosystems. Together, they call for end-to-end, collaborative Fair-ML research.
4 Towards Algorithmic Fairness in India
- The agenda calls for designing around end-to-end chains of algorithmic power and cross-disciplinary, cross-institutional collaboration.It considers how AI systems are conceived, produced, used, appropriated, assessed, and contested in India.Source: §4 · p. 3
4.1 Recontextualising data and models
- Data and model recontextualisation requires accounting for distortions and data voids while involving communities in data stewardship, consent, and category design.The paper proposes relationship-based transitive consent and discusses collective consent, data trusts, and data co-ops as ways to support community agency and data reliability.Source: §4.1 · p. 3
- Fairness testing should address Indian axes of historical injustice and operational challenges, including caste representation, code-mixing, and quota allocation.The paper notes that readily available Indian name datasets for fairness evaluation are lacking, and that matching under quota constraints has received little ML research attention.Source: §4.1 · p. 3
4.2 Empowering communities
- Empowering communities entails participatory Fair-ML knowledge systems and designing for heterogeneous literacies, costs, connectivity, languages, and devices.The paper also calls for care in deployments, meaningful recourse, and engagement with user feedback in India.Source: §4.2 · pp. 3–4
4.3 Enabling fair ML ecosystems
- Enabling accountability requires engagement across civil society, media, industry, judiciary, and state, alongside transparency about datasets, processes, models, limitations, and failures.The paper also calls for APIs, documentation, and socio-economic datasets to support accountability.Source: §4.3 · p. 4
5 Conclusion
The paper argues that Western efforts to make AI fair may not transfer readily to India, based on an empirical bottom-up analysis. It proposes a holistic Fair-ML research agenda that reflects local considerations.
5 Conclusion
- Western efforts to make AI fair may not transfer readily to India, as shown through an empirical bottom-up analysis.The paper argues for moving beyond localising data and model fairness alone.Source: §5 · p. 4
- The paper calls for general approaches to fairness that reflect local considerations, including Indian concepts of fairness where relevant.It presents a holistic research agenda for operationalising Fair-ML in India.Source: §5 · p. 4
A Methodology details
The study combines expert interviews and analysis of algorithmic deployments and policies in India. It uses purposeful recruitment and inductive coding, with stated procedures for protecting participants and describing the researchers’ positions.
A Methodology details
- The study combined interviews with 36 experienced researchers, activists, and lawyers with analysis of Indian deployments and policies beginning in 2009.Interviewees represented seven disciplines and had 10–30 years of experience working with marginalised communities or on social justice.Source: A Methodology details · p. 8
- Researchers recruited experts purposively until saturation, conducted English-language interviews, and used inductive coding to develop and iteratively apply three power-structure categories.Two team members created a code book from the themes identified in transcripts.Source: A Methodology details · p. 8
- The team used informed consent, restricted data access, and de-identification procedures, and disclosed its members’ backgrounds and privileged class and/or caste positions.Participants could choose anonymity or acknowledgement; identifiable details were redacted when quoting them.Source: A Methodology details · p. 8
B Axes of Discrimination in India
The section introduces axes of potential ML unfairness in India and says the table draws on interview themes, secondary research, and authoritative statistics.
B Axes of Discrimination in India
- Table 1 organizes potential ML unfairness in India around sub-groups, proxies, and harms.The accompanying text says the axes were raised in interviews and enriched with secondary research and statistics from authoritative sources.Source: Table 1 · p. 9, B Axes of Discrimination in India · p. 9