Decolonial AI: Reshaping AI Power

• Decolonial AI is an interdisciplinary field that interrogates and dismantles colonial patterns in AI systems by prioritizing marginalized knowledges and community sovereignty.
• It employs methodologies such as critical technical practice, participatory design, and community-governed data to contest extractive data regimes and epistemic violence.
• The approach drives reforms in AI governance and technical design, fostering ethical, context-aware, and justice-oriented innovations in technology.

Decolonial AI is an interdisciplinary field and critical practice directed at dismantling the colonial legacies and ongoing structures within artificial intelligence systems, research, and governance. It interrogates how AI reproduces patterns of epistemic, political, economic, and ecological domination rooted in colonial histories, and it advances methodologies and infrastructures to center marginalized knowledges, restore community sovereignty, and promote pluralistic, justice-oriented futures. Encompassing theory, methodological frameworks, technical pipelines, governance innovations, and participatory design processes, Decolonial AI is both diagnostic and reparative, oriented toward transforming the design, operation, and evaluation of AI to align with anti-colonial, abolitionist, and community-based principles.

1. Conceptual Foundations and Dimensions

Decolonial AI builds on foundational concepts from decolonial theory, postcolonial studies, and critical data science, translating them into the technical, institutional, and epistemic fabric of artificial intelligence. Key distinctions and terms:

• Coloniality of Power: Refers to the enduring matrix of power that survives beyond formal colonialism, structuring authority, economy, knowledge, and gender/sexuality in global hierarchies (Mohamed et al., 2020).
• Data Colonialism and Algorithmic Colonialism: Denote the extraction and commodification of human social life as data, and the imposition of AI systems that reproduce and automate forms of exploitation, dispossession, and control (A et al., 24 Nov 2025, Mollema, 2024).
• Epistemic Violence and Universalism: AI systems often encode and privilege Western forms of knowledge, marginalizing or erasing Indigenous, Afro-descendant, and Global South epistemologies (Mohamed et al., 2020, Varshney, 2023).
• Decoloniality: Extends beyond political independence to include the deconstruction of knowledge hierarchies, restoration of cultural and data sovereignty, and reimagining of technology to serve pluralistic and community-defined ends (Mollema, 2024).

Decolonial AI systematically interrogates three intersecting dimensions: political (governance, agency, regulation), ecological (resource extraction, environmental cost), and epistemic (whose knowledge, values, and subjectivities AI encodes and enforces) (Mollema, 2024).

2. Critical Diagnoses of Colonial Patterns in AI

A growing body of empirical and theoretical work diagnoses how AI perpetuates colonial logics at multiple system levels:

• Extractive Data and Labor Regimes: Data collection, annotation (“ghost work”), and attention economies exploit low-wage labor and digital production in, or from, the Global South while decision-making power and profit accrue to the Global North (A et al., 24 Nov 2025, Mohamed et al., 2020, Vargas-Solar, 2022).
• Platform Capitalism and Surveillance: Mainstream AI research and application (e.g., predictive policing, algorithmic sentencing, resource allocation) entrench hierarchies of dispossession, surveil marginalized populations, and reinforce the infrastructure of carceral and capitalist control (Earl, 2021, Wang et al., 8 Oct 2025).
• Universalizing Model Assumptions: Western-trained models, benchmarks, and loss functions impose normative standards, such as moral absolutism in LLM alignment, that erase context-specific ethical, cultural, and linguistic priorities (Varshney, 2023, Eze et al., 20 Oct 2025).
• Neglect of Systemic Harm: Superficial “diversity” efforts or token representation fail to address system-level oppressions or challenge the historical complicity of computational sciences with eugenics, race science, and scientific racism (Birhane et al., 2020).

These critiques demand an epistemic and practical rupture with historical and ongoing colonial structures across the AI pipeline.

3. Frameworks and Methodologies for Decolonial AI

Decolonial AI asserts a spectrum of methodologies, ranging from critical technical practice and participatory action research to postcolonial and abolitionist frameworks:

• Critical Technical Practice (CTP): Mandates reflexive co-design, transparency (datasheets, model cards), participatory harm audits, and explicit engagement with local value assumptions at every stage of AI development (Mohamed et al., 2020).
• CARE Principles: Formalizes community-governed data practices—Collective Benefit, Authority to Control, Responsibility, and Ethics—operationalized as explicit constraints in optimization and data management pipelines (Roberts et al., 2023). For example:

$$\text{Minimize}_{\theta}\;\mathcal{L}_{task}(\theta;D)$$

subject to:

$$\text{Collective Benefit:}\quad U_{ind}(\theta) - U_{baseline} \geq 0$$

and other CARE constraints.

• Participatory Design and Vernacularization: Employs participatory action research, co-speculative design, and the vernacularization of taxonomies of harm. Frameworks such as those for AI safety require the translation of universalized categories into locally meaningful, situated, and community-governed forms, often via iterative ethnographic engagement and context-specific crosswalks (Kennedy et al., 2024, Kenny et al., 2024).
• Abolitionist and African Data Ethics: Pursues the destruction and transformation of oppressive infrastructures, privileging Black or Indigenous autonomy, self-determination, communalist practice, and the operationalization of relational philosophies such as Ubuntu (Earl, 2021, Barrett et al., 22 Feb 2025, A et al., 24 Nov 2025).
• Alignment Beyond Moral Absolutism: Advances model and alignment openness, participatory value co-creation, and the inclusion of excluded knowledges (e.g., viśeṣa-dharma) via modular, adapter-based architectures (Varshney, 2023).

A sample table summarizes key frameworks operationalized in recent literature:

Framework	Key Principles	Example Reference
CARE	Collective Benefit, Authority, Responsibility, Ethics	(Roberts et al., 2023)
African Data Ethics	Power asymmetry, Self-determination, Communalism	(Barrett et al., 22 Feb 2025)
Vernacularization	Participatory, context-attentive harm taxonomies	(Kennedy et al., 2024)
Open Alignment	Openness, viśeṣa-dharma (context morality)	(Varshney, 2023)

4. Technical Interventions and System Architectures

Decolonial AI practice instantiates alternative technical and organizational architectures that resist colonial logics:

• Community-Governed Datasets: Data curation, ingestion, and label ontologies are co-designed by local or affected communities; local governance bodies are empowered to review, veto, and update data usage (Vargas-Solar, 2022, A et al., 24 Nov 2025, Alimujiang, 21 Oct 2025).
• Model Openness and Context-Aware Modularization: Open-source models (weights, data, adapters) enable fine-tuning for local epistemologies and linguistic diversity. Adapter orchestration (e.g., LoRA + bandit controllers) can instantiate context-specific moralities (viśeṣa-dharma) in LLMs at inference time (Varshney, 2023).
• Infrastructure and Sustainability-Empowered Pipelines: Streaming, edge computing, federated learning, and low-cost hardware enable operational autonomy in resource-constrained or infrastructurally colonized regions, reducing dependency on external cloud providers and their governance regimes (Reddyhoff, 2022, Segun et al., 12 Aug 2025).
• Participatory Governance and Evaluation: Multi-stakeholder councils, audit labs, and community ratification of evaluation metrics become normative, with preference judgments and weighted value adherence scores replacing capability-only "leaderboards" (Wang et al., 8 Oct 2025).
• Decentralized and Commons-Based Data Infrastructures: Community-run archives, co-operatives, and distributed ledgers encode permissioning and benefit-sharing rules, enabling data and model sovereignty (Vargas-Solar, 2022).
• Slow-Media and Temporal Decolonization: Systems design intentionally preserves cultural temporality, prioritizing organic human cycles over computational efficiency or "real-time" imperatives (Alimujiang, 21 Oct 2025).

5. Governance, Ethics, and Accountability

Decolonial AI recasts data and AI governance as a site of contestation and regime change, embedding accountability and ownership in community, state, and transnational structures:

• Data and Research Sovereignty: Research agendas, funding priorities, credit, authorship, and institutional infrastructure must be locally anchored and designed to be self-sustaining beyond initial grants (Reddyhoff, 2022, Barrett et al., 22 Feb 2025).
• Multi-Level Institutionalization: From community data councils and open audit labs, through national statistical offices, to continental alliances (African Union AI Institutes), governance must be structured to allow bottom-up as well as harmonized top-down oversight (Segun et al., 12 Aug 2025, Eze et al., 20 Oct 2025).
• Metrics and Benchmarks: Standard technical metrics are augmented or replaced by domain-specific indicators of trust, participation, collective flourishing, and harm reduction, including context-layered F₁ scores for multilingual benchmark suites (Segun et al., 12 Aug 2025, Kennedy et al., 2024).
• Legislation and Formal Enforcement: Legal frameworks, such as cultural heritage protection laws or the formalization of CARE/FAIR principles for Indigenous and marginalized populations, are integral to ensuring the authority, consent, and benefit rights of communities (Roberts et al., 2023).
• Ethics Beyond Compliance: Ethics becomes not merely a compliance checkbox but a practice of continual participatory judgement, mutual accountability, and redress for harms, grounded in locally operationalized principles rather than abstract universals (Barrett et al., 22 Feb 2025, Wang et al., 8 Oct 2025).

6. Case Studies and Practical Implementations

A spectrum of case studies demonstrates how decolonial AI principles are enacted in practice:

• Performative AI in Cultural Heritage: Feedback-driven mixed-reality art experiences with live Dutar performance embed oral histories and images into model architecture, foregrounding moments where human performers override or re-anchor AI suggestions to maintain cultural survivance (Alimujiang, 21 Oct 2025).
• African AI Safety Agenda: A five-point plan for AI governance in Africa centers on rights-based frameworks, inclusive benchmarks for 25+ languages, indigenous AI Safety Institutes, and continental forums that resist monolingual, infrastructural, and computational colonialism (Segun et al., 12 Aug 2025).
• Co-Speculative Design for Marginalized Youth: Workshops with BIPOC youth leverage speculative design to critique surveillance, algorithmic exclusion, and propose alternative, communally governed infrastructures and metrics (e.g., equity index) (Kenny et al., 2024).
• Community-Led Model Evaluation and Deployment: Participatory co-design with abolitionist and restorative justice practitioners grounds model evaluation in community-defined values, with open governance over data and algorithmic features (Wang et al., 8 Oct 2025).
• Deployment of AI in Low-Resource Public Health: Participatory calibration and edge-enabled deployment of environmental monitoring sensors in Uganda ensures research and technical sovereignty, equitable authorship, and budgeted mentoring for local contributors (Reddyhoff, 2022).

7. Challenges, Controversies, and Ongoing Directions

Implementing decolonial AI raises unresolved challenges:

• Measuring and Redressing Systemic Harm: The field lacks standardized, quantitative frameworks for structural oppression. Metric development is suggested as a research frontier (Birhane et al., 2020).
• Resource Intensity of Participatory and Vernacularization Practices: Gathering, maintaining, and updating context-specific epistemologies, taxonomies, and metrics is resource-intensive and requires expertise not always accessible to under-resourced communities (Kennedy et al., 2024).
• Institutional Resistance and Epistemic Inertia: Entrenched academic, corporate, and regulatory structures may resist the de-doctrinalization of "universal" AI principles or standards (Mollema, 2024, A et al., 24 Nov 2025).
• Translating Principles into Scalable Architectures: While adapter-based alignment, federated learning, and data-governance technical pipelines are invoked as promising, the generalization of these approaches across diverse community contexts remains an open technical and governance problem (Varshney, 2023).
• Global Pluralization, Not Relativism: The field must guard against a superficial relativism that detaches local epistemologies from mutual accountability, embracing instead a vigilant pluralism (Mohamed et al., 2020).

Continued progress in decolonial AI demands rigorous interdisciplinary collaboration, infrastructural re-building, and both technical and philosophical innovation to realize technology as a commons in service to global, community-defined justice and flourishing.

---

References

• "Decolonial AI: Decolonial Theory as Sociotechnical Foresight in Artificial Intelligence" (Mohamed et al., 2020)
• "Decolonial AI as Disenclosure" (Mollema, 2024)
• "African Data Ethics: A Discursive Framework for Black Decolonial Data Science" (Barrett et al., 22 Feb 2025)
• "Dependency, Data and Decolonisation: A Framework for Decolonial Thinking in Collaborative AI Research" (Reddyhoff, 2022)
• "Data Flows and Colonial Regimes in Africa" (A et al., 24 Nov 2025)
• "Decolonial AI Alignment: Openness, Viśeṣa-Dharma, and Including Excluded Knowledges" (Varshney, 2023)
• "AI for Abolition? A Participatory Design Approach" (Wang et al., 8 Oct 2025)
• "Towards decolonising computational sciences" (Birhane et al., 2020)
• "Evaluating LLMs for Career Guidance: Comparative Analysis of Computing Competency Recommendations Across Ten African Countries" (Eze et al., 20 Oct 2025)
• "Vernacularizing Taxonomies of Harm is Essential for Operationalizing Holistic AI Safety" (Kennedy et al., 2024)
• "When Strings Tug at Algorithm: Human-AI Sovereignty and Entanglement in Nomadic Improvisational Music Performance as a Decolonial Exploration" (Alimujiang, 21 Oct 2025)
• "Calling for a feminist revolt to decolonise data and algorithms in the age of Datification" (Vargas-Solar, 2022)
• "In Consideration of Indigenous Data Sovereignty: Data Mining as a Colonial Practice" (Roberts et al., 2023)
• "Toward an African Agenda for AI Safety" (Segun et al., 12 Aug 2025)
• "Reimagining AI: Exploring Speculative Design Workshops for Supporting BIPOC Youth Critical AI Literacies" (Kenny et al., 2024)
• "Towards an Abolitionist AI: the role of Historically Black Colleges and Universities" (Earl, 2021)