Oil or soil

Most workshops about AI open with a slide that explains what AI is. We open with a laptop and a webcam.

In June 2026 we ran this exercise at Parque das Tribos, an urban Indigenous neighbourhood in Manaus, with about thirty people from Sateré-Mawé, Ticuna, Baré and non-Indigenous families. Many of them already used AI tools for school, so we skipped the speech about AI arriving and went straight to how it works. Each group trained an image classifier on their own selfies with Teachable Machine, a free tool that runs in the browser. People got the main idea within minutes. The model finds patterns in the examples you feed it, and that is all it does.

The interesting moment comes when the examples are uneven. Give the model a hundred pictures of one person and ten of another, and it will be confident about the first and confused by the second. Nothing in the code is against anyone. The model never got to know the second person. Add more pictures and the problem goes away.

The same thing happens at the scale of the planet. GBIF, the Global Biodiversity Information Facility, is the largest open database of where species have been seen. Scientists, governments and AI models all draw on it. When we checked in October 2026 it held about 3.96 billion records. About three in four came from North America and Europe, and the United States alone accounted for a third. Brazil, which holds most of the Amazon, accounted for just over 1 percent. The Democratic Republic of the Congo, home to most of the Congo Basin rainforest, accounted for 0.04 percent.

A model trained on that map knows a lot about birds in Ohio and very little about the forest around Manaus. It is the webcam exercise again, with the whole world in front of the camera.

There are two ways people talk about fixing this. The first treats data like oil. You find it, extract it, ship it somewhere to be refined, and sell what comes out. In that picture the answer to the gap is to send more scientists and sensors to the tropics and bring the data home.

The second treats data like soil. Audrey Tang, Taiwan's former digital minister, described it in a statement to the Senate of Canada: soil asks "who tends the data, understands its seasons and meanings, can correct when it becomes harmful and decides what should be planted." In that picture the answer is to help the people who already live in those places collect data on their own terms, keep it close, and decide what leaves.

The two pictures lead to very different projects. In the first, communities are a source. In the second, they are in charge of the ground. This playbook works from the second picture. The lessons that follow show what that takes: rules agreed before anything is collected, the right to say no, tools that run locally, a council that decides, and fair pay for the work.