Hindou Oumarou Ibrahim stood before delegates in Geneva and raised something that should be obvious. And yet it still isn't binding standard in any international framework. If an artificial intelligence system is going to use data from indigenous territory, the community living there has to give its free, prior, and informed consent before the project starts, not after the model has already been trained.
That's the definition that matters here. Free, prior, and informed consent is a mechanism because it establishes who decides and at what point in the process. Not because it sounds good in a press release. Without that mechanism, what exists is symbolic consultation, a photo with indigenous leaders on stage while the technical decisions have already been made elsewhere.
Ibrahim, coordinator of the Association of Peuhl Women and Indigenous Peoples of Chad, has spent years at UN forums insisting on something specific. The environmental and territorial data feeding AI models, from satellite monitoring to predictive deforestation systems, comes from lands that indigenous peoples have managed, known, and inhabited long before any AI lab existed. That this knowledge gets used without consultation, and that the benefits of the resulting systems don't flow back to those communities, is not a design accident. It's the design.
The event was the Global Dialogue on AI Governance, held in Geneva. One of those spaces where the UN brings together governments, tech companies, and, increasingly at least on paper, indigenous representatives under the umbrella of initiatives like AI for Good. The format will feel familiar to anyone who's followed climate summits: panels, good-will declarations, closing photos. What's still missing is an enforcement mechanism. No international treaty today obligates an AI lab to obtain indigenous consent before collecting or using territorial data. Ibrahim's demand doesn't ask for something new in the vocabulary of human rights. Free, prior, and informed consent has existed as a standard since the International Labour Organization's Convention 169 of 1989. What she's asking is that this standard be applied to a technology nobody imagined in 1989.
Here it's worth asking something directly. Why doesn't discursive inclusion in forums like Geneva translate into real decision-making power? Being on the panel doesn't grant control over data flow. You can speak in the plenary session and simultaneously lack any legal tool to block a satellite monitoring company from processing images of your territory, cross-referencing them with environmental data, and training a predictive model that later gets sold as a conservation tool. Participating in the debate and owning the data are different things. Conflating them lets institutions like the UN present themselves as inclusive without ceding anything structural.
This pattern didn't start with AI. In 1992 the Convention on Biological Diversity was signed, and in 2010 the Nagoya Protocol, both as a response to decades of biopiracy. Pharmaceutical companies patenting compounds derived from medicinal plants used by indigenous peoples for generations, without payment, without recognition, without consultation. The turmeric case is textbook: a 1995 US patent claimed as an invention the plant's wound-healing use, knowledge documented in Ayurvedic texts centuries old, until India managed to have it revoked by presenting those same texts as evidence of "prior art." Neem, a tree used in India as a natural insecticide and antiseptic, was the subject of similar patents in Europe during the nineties, also overturned after long and costly litigation. Nagoya tried to close that legal gap by requiring access and benefit-sharing agreements. It took nearly twenty years to be partially implemented, and compliance remains weak in most signatory countries. That's the uncomfortable data point. When a specific legal framework existed for this problem, it worked only halfway and with two decades of delay. There's no reason to think AI governance, which doesn't even have that framework yet, will move any faster.
The parallel with today's algorithmic extraction isn't metaphorical. It's structural. Before, the vehicle of extraction was the pharmaceutical patent on a molecule derived from a plant known to indigenous communities. Now it's data flow. Satellite sensors capture images of forest cover, soil moisture, wildlife migration patterns over territories that indigenous communities manage and monitor with knowledge accumulated over generations. That knowledge, which signals anticipate drought, how a specific ecosystem behaves in cycles no satellite has yet observed, doesn't appear in the dataset but silently validates the models trained on the imagery. The community is left out of the flow at the moment of collection and out of the distribution at the moment of monetization. It's the same architecture of extraction without return, only now it runs in the cloud instead of being registered at a patent office.
And who benefits, technically and economically, from that consent not existing? The answer has layers. AI labs training environmental models get access to territorial data without consultation costs or licensing negotiations, which accelerates development and cuts spending. Satellite monitoring companies and providers of predictive deforestation or biodiversity models sell these systems to governments and NGOs as conservation tools, without the "shared benefit" framework ever including the communities whose territory and knowledge sustain the validity of those models. States and corporations that unilaterally define what constitutes "benefit" for an indigenous community retain the power to decide without negotiating. Sometimes that benefit is a seat on a panel. Sometimes it's a mention in an impact report. And the UN, meanwhile, gains something different: symbolic legitimacy. It can show that it included indigenous voices in Geneva without committing to any binding mechanism that would cost something to the governments funding it.
Technical flows get optimized for what's measurable and scalable. Consent doesn't qualify. A satellite sensor doesn't ask. An environmental data tracker doesn't pause to negotiate terms with the community living in the photographed area. The entire architecture, from collection to model training, is designed on the assumption that territorial data is an available resource, not a good over which someone holds prior rights. Changing that isn't a privacy-policy tweak. It's redesigning the flow from the first point of contact with the data.
What would it technically mean to build a system that actually requires consent? Concrete proposals already exist, though none solves the problem on its own. Data trusts, legal structures where a trust manages access to data on behalf of a community with the capacity to negotiate terms and block unauthorized uses, would let an indigenous community control who accesses data from its territory and under what conditions, rather than discovering after the fact that a model was already trained on its lands. Community data licenses, similar to those already being discussed for indigenous Amazonian languages, could require that any commercial use of territorial data go through an explicit agreement with benefit-sharing, replicating in the AI space what Nagoya tried to do with genetic resources. And distributed data governance, where the decision over access doesn't rest with a single external actor but with a committee that has real representation of the affected community, is technically viable today. It requires no new technology. It requires institutional willingness to cede control.
None of these architectures is an easy or definitive solution. Data trusts can be captured by intermediaries who end up acting on the community's behalf without being accountable to it. Community licenses depend on the legal and technical capacity to negotiate them existing, something many communities lack without outside support, and that outside support brings its own power asymmetries. Distributed governance sounds good on paper and is hard to sustain when data volume grows faster than any committee's ability to review it. There's no settled answer about which of these mechanisms would work best, and I suspect nobody has one yet. This is more complicated than it looks, as explored in The Generosity in the Doorway. Data without context. Communities without a voice.
How would we get consent to stop being just a phrase uttered in Geneva?
