Wiki · Person · Last reviewed June 25, 2026

Shakir Mohamed

Shakir Mohamed is a machine-learning researcher and institution builder whose work joins probabilistic modeling, generative AI, global-challenge applications, African AI capacity building, participatory AI, and decolonial critiques of technology power.

Category: Person Published: June 25, 2026 Modified: June 25, 2026 Last reviewed: June 25, 2026 Tags: Google DeepMind, Deep Learning Indaba, Decolonial AI, Participatory AI, AI Governance

Definition

Shakir Mohamed is a South African machine-learning researcher and AI institution builder whose public significance comes from joining two usually separated roles: technical work in probabilistic and generative machine learning, and governance work on participation, African AI leadership, decolonial AI, and sociotechnical responsibility.

In this wiki, he is best read as a bridge figure. His work treats AI capability, community capacity, epistemic power, and institutional ownership as connected governance questions. The point is not that every AI system branded as socially beneficial is safe; it is that safety and legitimacy depend partly on who can understand, evaluate, contest, adapt, refuse, and govern the system.

This profile therefore separates three evidence tracks: Mohamed's technical publications, the institution-building record of Deep Learning Indaba, and his sociotechnical arguments about decolonial and participatory AI. Mixing those tracks into a single "AI for good" story would weaken the page's central point: public benefit requires evidence, capacity, and power, not only intention.

Snapshot

Current Context

As of June 25, 2026, Mohamed's own site identifies him as a director for research at Google DeepMind and founder and Chair of the Board of Trustees of Deep Learning Indaba. A Royal Society profile likewise describes him as a Google DeepMind research director working on foundational AI, healthcare, education, environment, participation, and responsibility.

Deep Learning Indaba's current mission page frames the organization around Africans becoming active shapers and owners of AI advances, not merely observers or receivers. It lists the annual Indaba, local IndabaX events, mentorship, awards, and year-round opportunities as ways to build a sustainable pan-African AI community. The official 2026 Indaba page says the annual gathering will take place at Pan-Atlantic University in Lagos, Nigeria, from August 2-7, 2026.

In December 2025, AAAI announced Mohamed as the 2026 recipient of its Award for Artificial Intelligence for the Benefit of Humanity, citing his work to empower communities worldwide to learn, contribute, debate, and shape how AI is used. That recognition is significant because it treats community capacity and participation as part of AI's public benefit, not as outreach separate from research. It should still be read as an award citation, not as independent evidence that any particular AI deployment is beneficial.

Technical Research

Mohamed's technical work sits in probabilistic machine learning, deep generative models, variational methods, and the estimation machinery that lets models learn from uncertainty. His coauthored 2014 paper Semi-Supervised Learning with Deep Generative Models helped show how deep generative models and approximate Bayesian inference could improve learning from small labeled datasets and large unlabeled datasets.

His 2020 JMLR survey Monte Carlo Gradient Estimation in Machine Learning, coauthored with Mihaela Rosca, Michael Figurnov, and Andriy Mnih, organized pathwise, score-function, and measure-valued gradient estimators as core tools for modern learning systems. That work matters because gradients of expectations appear across supervised learning, unsupervised learning, reinforcement learning, probabilistic modeling, and simulation-heavy AI.

His applied collaborations also show why probabilistic framing matters in high-stakes domains. In healthcare, he was a coauthor on DeepMind's 2019 Nature paper on continuous prediction of future acute kidney injury from electronic health records, developed on a large U.S. Department of Veterans Affairs dataset. In weather, he was a coauthor on GraphCast and GenCast, machine-learning systems for medium-range deterministic and probabilistic forecasting. These examples are not proof that AI systems are automatically deployable, equitable, or locally suitable; they show why uncertainty, validation, data governance, and decision context matter when AI is tied to health and climate risk.

The deployment caveat is central. A model that improves prediction still needs local validation, institutional readiness, data-protection review, clinical or operational accountability, failure escalation, and human oversight. Mohamed's technical portfolio is therefore relevant to AI in Healthcare and AI Weather Forecasting because it combines capability claims with the question of who can safely use and challenge them.

This technical background is important to his public role. Mohamed is not only an outside critic of AI power. He is a researcher from inside advanced machine learning who argues that the field's social contract, geography, and epistemic assumptions must be redesigned.

Deep Learning Indaba

Deep Learning Indaba is one of Mohamed's most visible institutional contributions. The organization says its mission is to strengthen machine learning and artificial intelligence in Africa, with the goal that Africans become active shapers and owners of AI advances rather than observers or receivers.

Public profiles identify Mohamed as a Deep Learning Indaba co-founder, and the Indaba has grown from an annual conference into a wider ecosystem of local IndabaX events, mentorship, awards, research showcases, year-round opportunities, and volunteer leadership.

The deeper importance is infrastructural. AI capacity is not only a matter of model access. It includes teaching, peer networks, travel support, local research questions, publication pathways, leadership experience, and institutions that let people participate without leaving their region or abandoning local priorities.

That infrastructure has governance consequences. A region without technical communities, local reviewers, evaluation capacity, and research leadership becomes dependent on imported models, foreign benchmarks, and external claims about safety or benefit. The Indaba is therefore not only a training program; it is part of the institutional base for sovereign AI and public-interest scrutiny.

Capacity as Governance

Mohamed's institutional work makes a governance claim that is easy to miss: capacity is not downstream from safety; it is one of the conditions for safety. Communities need people who can inspect a model card, test a dataset, run an evaluation, question a benchmark, negotiate a procurement contract, and explain why a deployment fails locally.

This matters for foundation models, public-sector AI, health systems, education tools, climate services, and language technologies. Without local technical capacity, outside providers can define the task, provide the data pipeline, choose the evaluation, host the infrastructure, own the intellectual property, and narrate the benefit. The affected community then receives a system it cannot fully audit or refuse.

Access alone is not capacity. Open weights, cloud credits, donated tools, or one-off training programs can still leave dependency intact if communities lack compute operations, maintenance skills, language resources, evaluation methods, legal leverage, and institutions that can say no to a vendor or donor.

A Mohamed-style capacity review asks practical questions: who can reproduce the claim, who can inspect the data, who can adapt the model to local languages and institutions, who can contest errors, who can govern downstream reuse, and who owns the resulting infrastructure. Those questions connect Deep Learning Indaba to AI Procurement, AI System Inventory, Algorithmic Impact Assessments, Notice and Appeal, and Human Oversight of AI Systems.

Decolonial AI

Mohamed, Marie-Therese Png, and William Isaac's 2020 paper Decolonial AI helped name a major critique of artificial intelligence: AI is not deployed into a neutral world. It is built inside histories of colonial extraction, unequal knowledge production, data capture, race, labor, language hierarchy, and global power.

The paper argues that decolonial theory can give AI communities sociotechnical foresight. Its proposed tactics include critical technical practice, reverse tutelage and reverse pedagogy, and renewal of affective and political communities. In practical terms, this asks AI researchers to learn from the people and places usually treated as deployment targets, data sources, users, or afterthoughts.

Decolonial AI is sometimes flattened into a slogan. In Mohamed's work, it is better read as a discipline of design and governance: ask whose knowledge counts, whose harms are visible, whose language is modeled, whose data is captured, whose institutions are strengthened, and whose future is being optimized.

This is especially relevant to foundation models, language technologies, health AI, and climate tools. A system can be technically impressive and still reproduce dependency if it extracts local data, ignores local expertise, centralizes infrastructure, or treats communities as beneficiaries rather than decision-makers.

Participatory AI

Mohamed is also a coauthor of Power to the People? Opportunities and Challenges for Participatory AI, a 2022 paper with Abeba Birhane, William Isaac, Vinodkumar Prabhakaran, Mark Diaz, Madeleine Clare Elish, and Iason Gabriel. The paper reviews participatory approaches while warning that participation can be vague, coopted, or confused with other forms of consultation.

This matters because participation is now common language in AI governance. A company can invite public feedback while leaving the real decision unchanged. A regulator can ask for consultation without changing who has power. A lab can gather community data while preserving extractive control. Mohamed's participatory-AI thread treats inclusion as a design and power problem, not as a decorative process.

Paired with Deep Learning Indaba, this gives his work a consistent pattern: build technical capacity, build communities, then insist that those communities have standing in the design, evaluation, deployment, refusal, and governance of AI systems that affect them.

Governance and Safety Implications

Mohamed's work pushes AI safety beyond model behavior alone. A system can pass a benchmark and still be unsafe if the affected community lacks technical capacity, language support, documentation, procurement leverage, audit access, or the institutional power to reject a deployment.

A practical safety review inspired by this work should ask whether the affected region can run independent evaluations, inspect training and evaluation data provenance, compare the system against local baselines, preserve public-sector exit options, and maintain an incident and appeals process after deployment. If those conditions are absent, the system may be technically strong but institutionally unsafe.

Governance Tests

A Mohamed-informed review turns "benefit" and "participation" into evidence questions. It asks whether affected communities have enough power, skill, and infrastructure to shape the system before, during, and after deployment.

These tests connect Mohamed's decolonial and participatory work to ordinary AI governance records: system inventories, procurement terms, impact assessments, model cards, audit trails, and post-market monitoring. They also help separate genuine public-interest AI from symbolic inclusion.

Limits and Misreadings

Mohamed should not be used as a stand-in for every African AI researcher, every decolonial AI argument, or every socially beneficial technology claim. His work is a useful bridge across technical research, institution building, and governance, but those strands still need claim-level evidence.

Recognition

TIME included Mohamed in its inaugural TIME100 AI list in 2023. Wits University reported the recognition because Mohamed and Pelonomi Moiloa were both Wits alumni named in that list. Treat this as recognition of public influence, not as independent evidence that any particular AI deployment is safe or beneficial.

The Royal Society profile for Mohamed describes technical and sociotechnical work across foundational AI, healthcare, education, environment, participation, and responsibility, and identifies him as a founder and trustee of Deep Learning Indaba.

In December 2025, AAAI announced Mohamed as the 2026 recipient of its Award for Artificial Intelligence for the Benefit of Humanity. The press release credited his work on social benefit, global communities, health, weather forecasting, education, and Deep Learning Indaba's role in strengthening African AI.

Source Discipline

Use Mohamed's own site and official institutional profiles for current roles. Use peer-reviewed papers and paper records for research claims. Use Deep Learning Indaba's own site for its mission, programs, and event plans. Use AAAI and TIME for recognition, but do not let award language substitute for evidence about model performance, deployment safety, or governance outcomes.

For decolonial AI and participatory AI, cite the specific papers rather than turning the terms into loose slogans. Decolonial AI supports claims about coloniality, sociotechnical foresight, critical technical practice, reverse tutelage, and political communities. Power to the People? supports claims about participatory methods, cooptation risk, and the need to ask who benefits from participation.

Prefer publisher or proceedings pages when a paper has been formally published, and use arXiv for preprints, open access, or version history. A current web profile is appropriate for a current role; it is not evidence that a deployed AI system works or that a governance process moved power.

For global-benefit claims, distinguish capacity-building evidence from deployment evidence. Building an AI community, publishing a model, winning a benchmark, and improving public welfare are different claims that require different sources.

For applied-system claims, preserve the domain boundary. A kidney-injury prediction paper, a deterministic weather model, a probabilistic ensemble forecast, and an education intervention are not interchangeable evidence for "AI for good." Each needs its own validation, deployment context, affected population, and accountability record.

Spiralist Reading

Shakir Mohamed matters because he refuses the idea that AI's center is fixed.

The ordinary story of advanced AI runs through a few labs, a few clouds, a few benchmark tables, and a few capital markets. Mohamed's work widens the map. It says the future of machine intelligence is not only a question of who scales the model. It is also a question of who has the teachers, conferences, languages, datasets, institutions, and standing to contest what the model is for.

For Spiralism, this is a form of cognitive sovereignty at continental scale. A community that cannot train, inspect, criticize, adapt, or govern AI is not merely behind on technology. It is being written into someone else's world model.

Open Questions

Sources

Return to Wiki