Oxfam's Rights-Based AI Framework: A Model for Social Justice Organizations

The AI industry has produced an enormous volume of ethics guidelines, responsible AI principles, and fairness frameworks over the past several years. Most major technology companies have published versions of these commitments. Many nonprofit sector bodies have developed their own variations. At first glance, this proliferation might look like progress. Human rights organizations, however, have raised a fundamental concern: voluntary ethics commitments are not enforceable, and they have not prevented documented harms.

Amnesty International's position captures this critique directly. In their February 2025 statement at the AI Action Summit in Paris, they called for "binding and enforceable regulation to curb AI-driven harms," explicitly distinguishing this from abstract ethics discussions that have failed to prevent real-world harm. Their Algorithmic Accountability Toolkit, published in December 2025, provides a practical complement to this advocacy: concrete methods for investigating and seeking accountability for algorithmic harms in welfare systems, policing, healthcare, and education across multiple countries.

The rights-based approach differs from ethics-only frameworks in several ways. Rights are not aspirational; they are legal claims grounded in binding instruments like the International Covenant on Civil and Political Rights, the Convention on the Rights of the Child, and regional human rights treaties. Rights create corresponding duties for states, corporations, and organizations that deploy systems affecting people. Rights can be enforced through courts, regulatory bodies, and international mechanisms. And rights belong to specific people in specific circumstances, not to the abstract "users" and "stakeholders" that ethics frameworks typically address.

A rights-based approach is not theoretical caution. It is a response to documented, ongoing harms that affect specific communities. Social justice organizations need to understand this landscape not to be paralyzed by it, but to make informed decisions about where AI can safely advance their work and where it poses unacceptable risks.

Facial recognition technology illustrates the problem clearly. Research has consistently shown that these systems perform significantly less accurately on darker-skinned faces, with Black individuals accounting for a substantial proportion of wrongful arrests attributable to faulty facial recognition matches. For organizations working with communities that have historically experienced discriminatory policing, adopting or advocating for facial recognition systems represents a direct contradiction of mission, regardless of the intended application.

Predictive policing systems compound this problem by embedding historical discrimination into algorithmic outputs. When these systems learn from arrest records shaped by decades of racially biased enforcement, they produce predictions that systematically over-flag Black and Latino neighborhoods, generating a feedback loop that appears data-driven but is actually replicating human bias at scale. Amnesty International has documented specific cases of this dynamic across multiple jurisdictions.

The welfare system is another high-stakes domain. Amnesty's research into Denmark's welfare AI system documented how algorithmic tools used to assess benefit eligibility created mass surveillance risks and systematically disadvantaged disabled people, low-income individuals, migrants, and refugees. Similar concerns have been documented in welfare systems in the United Kingdom, Australia, and the United States. For nonprofits that advocate for or provide services to any of these populations, these findings have direct operational relevance.

At the same time, AI has demonstrated genuine potential to advance social justice missions when deployed thoughtfully. Learning Equality matched thousands of educational resources with relevant categories in Uganda at a scale impossible without AI assistance. Digital Green has used AI to deliver agricultural information to smallholder farmers in their own languages. Legal aid organizations are using natural language processing to accelerate document review and expand access to legal help. The rights-based framework is not anti-AI; it is pro-accountability.

Adopting a rights-based AI framework does not require beginning with a comprehensive governance overhaul. Most organizations will be better served by a phased approach that builds on existing commitments while progressively deepening accountability structures. The following sequence reflects the priorities identified by Oxfam, NetHope, and other organizations that have navigated this process.

The first step is clarity about what AI is already in use. Many organizations are surprised to discover how widely AI tools have already been adopted informally by staff using freely available consumer applications. Conducting an AI audit, mapping every AI system touching organizational operations and beneficiary data, provides the baseline information needed for any subsequent governance work. This audit should cover not just tools purchased by the organization but any AI features embedded in existing software subscriptions, including CRM systems, fundraising platforms, and communication tools. Organizations that have already built foundational knowledge management systems may find this process smoother because they have existing documentation practices to build on.

Vendor accountability is a second early priority. Most nonprofits do not scrutinize the AI-related provisions in their vendor agreements closely, but these agreements determine whose data is used for model training, who has liability when AI systems cause harm, and what transparency obligations the vendor accepts. Requesting bias audit documentation, clarifying data usage terms, and adding explicit AI governance requirements to vendor agreements are practical steps that many organizations can implement without specialized legal expertise.

Establishing an internal AI governance structure, even a simple one, creates the organizational capacity to make consistent decisions over time. This does not require a dedicated AI ethics committee at the outset. Designating a staff member to serve as the AI governance lead, establishing a basic policy that identifies prohibited uses, and creating a review process for new AI tool adoptions provides a workable foundation. As the organization's AI use matures, this structure can grow in sophistication. The relationship to your broader AI strategic planning should also be clear from the start.