On November 11, 2026, researchers Madiha Zahrah Choksi, Reeda Shimaz Huda, Rita Faia Marques, Anja Thieme, and Cecily Morrison released an arXiv preprint titled Terms of Care: Designing for Participation in Data Governance by Disability Communities. The study introduces a modular data licensing framework that enables marginalized communities to govern dataset access, balancing fair artificial intelligence representation with robust safeguards against algorithmic exploitation, decontextualization, and unconsented commercial reuse.
What happened
Modern generative artificial intelligence systems and foundation models frequently fail to represent people with disabilities respectfully. Systematic audits of commercial image generation services show pronounced defect rates and stereotypical distortions when depicting individuals with physical disabilities, including dwarfism, blindness, and low vision. Addressing these representational gaps requires curating dedicated community-specific training datasets. However, depositing demographic imagery into open repositories under permissive licenses like CC0 or MIT presents severe hazards: marginalized groups risk non-consensual exploitation, automated biometric scraping, deepfake creation, and commercial commodification.
To resolve this fundamental governance dilemma, researchers from Columbia University, the Georgia Institute of Technology, and Microsoft Research partnered with three disability advocacy organizations through an international NGO. Published on the Cornell University-hosted arXiv repository, their investigation examined how high-context communities can participate directly in data stewardship. The research team co-designed a participatory governance framework called Terms of Care, replacing one-size-fits-all open licenses with modular, context-sensitive licensing agreements.
Methodology and key findings
The authors structured their investigation in two empirical phases. In Phase 1, the researchers conducted semi-structured qualitative interviews with project leads from each participating organization. These leads served as community data stewards, coordinating internal decisions and helping members form collective perspectives on AI representation. Concurrently, each organization assembled and annotated a specialized dataset of 400 images reflecting lived disability experience. In Phase 2, the team delivered a data governance webinar and guided participants through a modular licensing activity adapted from Creative Commons clauses, allowing communities to configure permitted uses, sector-specific restrictions, and cultural safeguarding requirements.
The findings revealed that data governance in high-context communities depends on organizational memory, hybrid authority, and historical exposure to harm rather than abstract legal clauses:
- Child protection and local regulation: Participant organization P1, an advocacy group in Kenya representing people with visual impairments, prioritized strict safeguards for minors and compliance with the Kenyan Data Protection Act, responding to prior institutional incidents where outside contractors bypassed internal channels.
- Anti-surveillance and dignity: Organization P2 permitted educational and research reuse while enacting categorical bans on biometric surveillance, facial recognition, policing, military deployment, defamatory media, and re-identification.
- Sector-specific bans: Organization P3, a 501(c)(3) body representing people with dwarfism, barred commercial reuse and instituted explicit exclusions against pharmaceutical marketing and entertainment uses like circuses or wrestling. These restrictions honored organizational memory resisting corrective medicalization and stigmatizing public spectacles.
Despite the success of the modular licensing activity, organizational leaders voiced critical uncertainties regarding downstream enforcement. Participants questioned whether data platforms like Hugging Face would verify compliance or whether the burden of identifying violations and pursuing legal remedies would fall entirely on under-resourced community organizations.
Why it matters for research practice
For institutional review boards (IRBs), data managers, and research ethics committees, the Terms of Care framework exposes significant shortcomings in conventional data management protocols. Traditional research ethics has relied almost exclusively on transactional, one-time individual consent forms, followed by binary distribution choices that treat data as either fully restricted or completely public. In high-context communities, data stewardship is inherently relational, shaped by multi-generational trust and shared norms.
Applying methods grounded in community-based participatory research (CBPR) to the curation of machine learning training sets directly strengthens data integrity. As regulatory frameworks such as Article 10 data and data governance standards under the EU AI Act require developers to prevent discriminatory bias in high-risk models, researchers must document provenance and community consent rigorously. Coupling modular licenses with standardized metadata schemas such as datasheets for datasets provides clear transparency into what populations permitted, where data may flow, and which applications remain strictly prohibited.
What changes for researchers
Research institutions, repository curators, and AI engineering teams should adopt several concrete operational practices from this study:
- Move beyond static open licenses: Replace generic permissive licenses on sensitive demographic datasets with modular covenants that distinguish non-profit research from commercial exploitation and high-risk applications.
- Fund community data stewards: Recognize community liaisons as formal co-investigators and data stewards, allocating grant funding and institutional resources to support collective deliberation on data governance.
- Build relational repository infrastructure: Public and institutional repositories must develop machine-readable permission flags, user verification mechanisms, and reporting channels so enforcement does not remain the sole responsibility of vulnerable communities.
