Summary:

We present Frontier 2025, a curated review of breakthroughs and deployments across Africa’s AI landscape. The project documents where progress is real, measurable, and oriented toward public benefit: data sovereignty, physical-world applications, and emerging governance capacity. Our aim is to move beyond forecasts and take a clear account of what has actually been built.

Why this project

Much writing on AI focuses on future scenarios: dramatic gains, catastrophic risks, or sweeping transformations. Less attention is paid to careful stock-taking. What has already changed? What capacity has been built? Where have institutions, datasets, and systems quietly shifted the frontier?

What Frontier 2025 is

Frontier Progress 2025 is an attempt to answer those questions for Africa’s AI ecosystem. It curates milestones from research, datasets, deployments, and policy across the continent and beyond, with a focus on public-interest impact. This is in the spirit of year-end reviews in science and engineering: careful records meant to guide the next stage of work. The project consists of a public gallery of entries—datasets, research papers, deployments, partnerships, and policies—each evaluated along three dimensions:

• Surprisal S(p): How unexpected the breakthrough is, and how likely it is to generalize.
• Impact: The qualitative magnitude of the change (e.g., systemic, humanitarian, institutional).
• Theory: A short narrative explaining why this represents a genuine shift rather than incremental progress

Three patterns from the 2025 frontier

Across the entries, three themes recur.

1. Reclaiming digital agency

A central story of 2025 is the move from data scarcity to data construction. While the vast majority of African languages remain unsupported by major language models, communities have begun to respond with large-scale, locally governed datasets.

The release of multi-thousand-hour speech corpora, new multilingual text collections, and stewardship frameworks marks a shift from critique to infrastructure. The question is no longer only why African languages are excluded, but how to build and govern the resources that change this.

This matters not just for inclusion, but for long-run alignment: models trained on locally grounded data better reflect social norms, domains, and needs that matter for public use.

2. AI in the physical world

Another pattern is the movement of AI beyond text and chat into physical and social systems. Weather forecasting models now outperform traditional methods in data-sparse regions and are being deployed directly to users. In healthcare, AI-assisted diagnostics and clinical tools are moving from pilots toward integration. In agriculture and infrastructure, similar transitions are beginning. These are not abstract benchmarks. They are systems touching climate resilience, disease detection, and service delivery. At the same time, they expose bottlenecks: compute access, connectivity, and maintenance capacity. Initiatives aimed at local AI infrastructure reflect growing awareness that models alone are not enough.

3. Governance and human capital

Policy has begun to shift from statements of principle to institutional form. Continental strategies, national partnerships with frontier labs, and regulatory coordination point to a more mature governance layer. In parallel, research into data labour, annotation supply chains, and “alignment debt” highlights the human costs embedded in AI systems. With hundreds of millions of young people entering African labour markets in coming decades, the choice is stark: build pathways for participation as designers and stewards of systems, or reproduce extractive roles at scale. 

“Things going surprisingly well”

A core section of the project is a list of 50 developments that, taken together, suggest real momentum: deployed systems, foundational research, new datasets, and working partnerships.

These are not claims that “everything is fine.” Many constraints remain: funding gaps, compute access, fragile institutions, and uneven capacity. But the list serves as a counterweight to narratives that see only stagnation or dependency.

For the EA community, this kind of evidence matters. It helps ground discussions about where marginal resources can do the most good, and where early investments are already compounding.

Relevance to effective altruism

From an EA perspective, Frontier 2025 connects to several priorities:

• Global public goods.
   Open datasets, forecasting systems, and safety frameworks are classic public goods with high spillovers.
• AI safety and alignment.
   Local data, governance capacity, and institutional maturity shape how safe systems can be built and deployed in practice, not just in theory
• Cause prioritisation under uncertainty.
   Mapping surprisal against impact offers one way to reason about where attention might be most valuable when outcomes are still unclear.

The project does not claim to settle these questions. It aims to provide better empirical footing for them. Many important efforts are missing. Some entries will age poorly. That is expected. The value lies in making the map explicit, so it can be criticised and improved.

A call for contributions

Frontier 2025 is a living document. If you are working on something that advances data, safety, governance, or societal impact—especially in low-resource contexts—we want to hear about it. Progress is easier to shape when it is first made visible.

If you’re interested in contributing or suggesting entries, you can reach us at jonas@equiano.institute.