Exploring the Data Behind AI and What It Means for Peacebuilding: A Jargon-Free GuideBlog Details

BY Nathan Coyle / ON 15 JULY, 2026

Exploring the Data Behind AI and What It Means for Peacebuilding: A Jargon-Free Guide

One of the most-shared posts from PeaceTech Alliance is PeaceTech 101: A Simple Guide for Non-Tech People. In it, we mentioned, almost in passing, that "over 90% of datasets used to train AI systems originate from Europe and North America, while more than 80% of violent conflicts occur outside the Global North." We cited that from existing research, MIT and UCDP, ACLED and PRIO, and moved on.

This time, we didn't want to just repeat someone else's number. We wanted to build it ourselves, check it against fresh, current data, and see exactly how bad the gap really is. Every chart below is live and interactive, built straight from the underlying dataset, not a screenshot. Scroll through, hover the charts, and see for yourself.

Where Does This Data Actually Come From, and Why Should You Trust It?

Quick definition before we start: you'll see the terms "Global North" and "Global South" a lot below. Roughly speaking, Global North means wealthier, higher-income countries, and Global South means everywhere else. It's not a perfect label, and while a lot of the community is unhappy with that distinction, it is a working definition people use, hence why it has been used here. Before we get into the charts, it's worth being upfront about where these numbers come from, because that's exactly the kind of thing PeaceTech should never gloss over.

• Who has the AI computers: tracked by Epoch AI, an independent research organisation that publishes its global compute data openly under a Creative Commons licence.
• Where conflict is happening: tracked by ACLED (the Armed Conflict Location & Event Data Project), one of the most widely used, independently verified conflict databases in the world, relied on by the UN, journalists, and researchers everywhere. We only used their published, aggregated country-level event counts, in line with their terms of use.
• Who's building AI's training data: tracked by the Data Provenance Initiative, an academic audit project that traces where AI training datasets actually come from and who compiled them.

None of these numbers are ours. They're independent, publicly available, and already trusted by researchers and journalists individually. What hadn't been done before, as far as we could find, was lining all three up side by side, country by country, to see exactly how far apart they really are. That's the part we built.

So What Did We Actually Compare?

Three things, country by country: where the physical computers that power AI actually sit, where the data used to train AI comes from, and where armed conflict is happening right now. Three separate maps of the world, laid on top of each other.

• AI compute means the physical hardware, specialised chips, that actually run AI systems. Someone owns them, and they have to be somewhere.
• AI training data means the huge collections of text, speech, and video that AI systems learn from. Someone has to gather and organise that, usually a university, company, or lab.
• Conflict data here specifically means events that target civilians directly, tracked by one of the most established independent conflict databases in the world.

Here's the headline picture once we lined all three up:

All three numbers point the same direction. Almost all AI computing power sits in wealthy countries. Almost all civilian-targeting conflict is happening in poorer countries. And almost all AI training data was put together by organisations in wealthy countries too.

Just How Lopsided Is "AI Compute", Really?

Let's zoom into the compute side on its own.

North America alone holds more AI computing power than the rest of the world combined. Africa, on this measure, holds essentially none. If AI infrastructure were a road network, entire continents would be missing from the map.

Who's Actually Building the Training Data?

AI systems learn from examples. Somebody has to collect and organise those examples, and that somebody is based somewhere.

Same pattern again. The organisations building the datasets that shape how AI "understands" the world are overwhelmingly based in North America and Europe. That matters, because whoever builds the dataset also decides what's included, what's left out, and whose voice counts as a reliable source.

And Where Is the Conflict Actually Happening?

This is where the picture flips completely. The regions with almost no AI infrastructure, Latin America, Africa, and the Middle East, are exactly the regions carrying the largest share of civilian-targeting violence. North America, which holds nearly all the world's AI compute, barely registers here.

What Happens When You Put Both Maps Together?

Here's the moment those two pictures collide. Every dot below is a country. The further right, the more AI compute it has. The higher up, the more civilian-targeting conflict it has recorded. The size of the dot shows how many AI training datasets trace back to that country. Hover any dot to see exactly which country it is.

Look at the top-left corner. That's where the countries facing the most civilian-targeting conflict sit, and almost none of them have any meaningful AI compute at all. Of the fifteen countries with the highest number of civilian-targeting conflict events in this period, twelve have no recorded AI infrastructure whatsoever. Together, those fifteen countries account for just 0.06% of the world's AI compute capacity.

So What Does This Actually Mean for Peacebuilding?

This isn't really a story about anyone doing something wrong on purpose. It's a story about geography and history, money, chips, and research institutions have historically clustered in wealthy countries, the same way a lot of infrastructure does.

But it matters here specifically, because AI is increasingly being proposed as a tool for conflict early warning, humanitarian response, and peacebuilding analysis. If the infrastructure and training data behind those tools are built almost entirely away from the places facing conflict, then the people who understand that conflict best, on the ground, have had the least influence over how the technology actually works.

• Lived context matters most. No dataset built somewhere else can fully replace it.
• Infrastructure shapes influence. Whoever owns the compute and the data has an outsized say in how these tools get built.
• The gap is measurable, not just felt. That's the whole point of this post, now there are numbers to point to.

This is the exact tension we raised in PeaceTech 101: PeaceTech isn't defined by the technology itself, but by who governs it, who's included in building it, and whose knowledge counts. Now that argument has a dataset behind it.

Want to Dig Into the Full Country List?

Search or sort any of the 241 countries and territories in the dataset directly, right here.

The full dataset, methodology, and source citations are published openly through Harvard Dataverse. We're also putting together a full research project page with the complete academic methodology behind this, separate from this post, so keep an eye out for that.

This is a first step, not a final word. If you're working at the intersection of AI and peacebuilding, we'd like to hear from you.