Who decides

People often call an AI system sovereign because it was trained inside a country, paid for with public money, or built for a national language. Audrey Tang does not accept that shortcut. In her statement to the Senate of Canada she writes that "a model is not sovereign simply because it is national or speaks many languages. Sovereignty depends on the process."

Process sounds abstract, so make it concrete. Every AI or data project is a chain of decisions. Someone chooses the question. Someone decides what to collect and from whom. Someone labels the data, trains the model, tests it, decides where it runs, and decides what happens when it gets something wrong. A system belongs to a community when that community holds real authority somewhere along the chain, early enough to change what happens.

GainForest learned this during the XPRIZE Rainforest competition, where we co-led the ETH BiodivX team with researchers from ETH Zurich. Before the finals in July 2024, the team ran eight co-learning workshops with communities in and around Manaus. Four Indigenous and local communities formed an Indigenous Data Council. The council became an official member of the team with a voice in its decisions, and it approved the prices paid for the data communities collected during the finals. A benefit-sharing agreement made all of this formal.

Marina Mura, from the Mura people, chaired the council. She first came to one of our workshops in January 2024, curious about technology and, by her own account, afraid of it. Ten months later she gave the team's speech in Rio de Janeiro when ETH BiodivX won an XPRIZE bonus prize for co-developing inclusive technology. The team donated the full $250,000 to an endowment for Indigenous scientists in the Amazon.

In her field notes Marina explains why the arrangement worked. The team kept coming back with changes the village had asked for, and over time the village came to see that "it's not ETH BiodivX entering the village, but members of our community, as equals."

Compare that with a pattern you may have seen. A team designs a system, then holds a consultation to show it to the community. By then every important decision is made. Tang is blunt about this. Deliberation should help people "draw boundaries early", and it "should not be treated as symbolic consultation after a design has already been decided."

One more trap. Privacy protection matters, and it does not settle the question. A community can keep its data "technically protected and still lose authority", Tang writes, if its knowledge is absorbed into systems that it "cannot see, shape, correct or reject."

So the test for any project is short. Can the people it represents inspect it, correct it, reject it, move away from it, or retire it? If the answer is no at every step of the chain, the project is not theirs, whatever flag it flies.