AI Governance Has a Thanksgiving Problem
You know the story: two peoples, one table, mutual exchange. We repeat the myth of the Pilgrims and Wampanoag eating together in gratitude as a founding tale of collaboration.
But look closely and it fails. What followed that meal, if it happened as told, was land theft, disease, war, and erasure. We keep the myth because it flatters us. It claims the table was shared when the reality was taking.
AI governance has its own version, and we are living inside it.
The Table We Keep Setting
I haven’t been in the rooms where AI policy gets written. But I’ve spent years in monitoring and evaluation, and I know what extraction dressed as collaboration looks like.
It looks like “participatory” design processes where the questions were already set before anyone arrived. Community members give their time, their stories, their data. They never see the report. Donors require extensive data collection that serves their reporting cycles, not the program. Ownership defaults to the funder or the INGO, never to the community. The people who provided the raw material don’t get to decide what gets built from it.
We call this “data for learning.” We call it “accountability.” We put it in logic models and theory of change diagrams, and we thank communities for their contributions in the acknowledgments. I’ve written those reports. I’ve designed those surveys.
AI governance is running the same playbook with more compute and higher stakes.
Training data gets scraped the way communities get surveyed. It’s there. It’s cheaper than asking. The people who provide it don’t have the power to refuse. Consent is reduced to a checkbox, the same as the signature line on a data release form no one reads. Value flows one direction, and the people whose knowledge trained the model never see what gets built from it.
The M&E version: “We collected data in partnership with communities.”
The AI governance version: “We included diverse stakeholders in our process.”
Same grammar. Same function. Both frame a taking as a sharing.
What Got Taken to Build the Table
Start with the material. Rare earths from Congo, Chile, and China. Content moderation in Nairobi. Data labeling in Caracas at rates that would not pass in San Francisco. Energy drawn from places that will feel climate damage first and benefit least.
Then the knowledge. Data scraped without consent: books, images, conversations, medical records, sacred texts. Languages treated as raw material rather than living traditions held by living people. If a model answers a question about Navajo cosmology, where did that knowledge come from, who was asked, and what returns?
The outcome is model monoculture. A few systems built by a few companies, trained on a flattened mass of human expression, marketed as default infrastructure for reasoning and sold back to us at a premium. This is not a metaphor for colonialism. It is the same logic with new tools: extraction, consolidation, dependency. A monocrop planted over an ecosystem.
The table was built from what was taken, and the feast is made of it.
What Would Be Different
I’m drawing on principles that some Indigenous scholars have offered as challenges to Western governance models. This isn’t a survey of Indigenous thought. It’s a gesture toward what current AI governance lacks, and what taking these ideas seriously would require.
Reciprocity. Robin Wall Kimmerer writes about this in terms of gift economies: if something is given, something must return. Not as transaction, but as relationship. What returns from AI systems to the communities whose data trained them? To the languages that were scraped? To the artists whose work became a training set? Current governance has no mechanism for this. It doesn’t even have the question. Taking reciprocity seriously would mean actual return: revenue sharing, community benefit agreements, ownership stakes. The question isn’t whether AI companies can afford this. The question is why the current model assumes extraction is free.
Consent as ongoing relationship. The First Nations Information Governance Centre developed the OCAP principles: Ownership, Control, Access, Possession. Consent isn’t a checkbox. It’s relational and revocable, maintained over time, not captured once. A checkbox lets you take and leave. A relationship requires you to stay accountable. Taking this seriously would mean communities decide whether data is collected, how it’s used, who sees it, and where it lives. Consent can be revoked. Agreements get revisited. This is harder than a checkbox, and that difficulty is the point.
Stewardship over ownership. The idea that a corporation can own a model trained on the commons of human expression only makes sense inside a property regime that treats knowledge as commodity. Many Indigenous legal traditions don’t. Land, water, knowledge: held in trust, not possessed. Stewardship would mean limits on intellectual property claims over models trained on public data. It would mean some knowledge stays out of training sets entirely because it wasn’t offered. It would mean holding these systems in trust rather than as assets.
Decisions made with future generations in view. The Haudenosaunee principle: consider the impact seven generations out. AI development operates on quarterly cycles. That gap isn’t a detail. It’s a fundamental difference in what governance is for. Taking this seriously would mean slower timelines, fewer models, precautionary defaults, and the genuine option to say “not yet” or “not this.” Speed forecloses deliberation and forces adoption. Slowing down is a policy choice. What are we in a hurry for anyway?
These are not alternatives to governance. They are tested systems that the dominant model ignored because ignoring them paid. This is not a roadmap. It is a different orientation to power. The question is not how to govern AI responsibly. The question is what we would have to give up to govern it justly.
The Myth Persists
The Thanksgiving story lasts because it lets us feel grateful without feeling indebted. It turns taking into sharing.
AI governance tells itself a similar story right now. Inclusive processes. Responsible development. A seat at the table for everyone.
But the table was built from what was taken, and the feast is made of it.
The difference is that AI governance is young. The patterns are forming, not fixed. There is still time to change the story, or better, to stop telling stories and start redistributing power.
Whether that happens is an open question. I don’t know the answer. But I think it’s worth sitting with the question through the holiday. Let it unsettle the meal a little.
Anthralytic is a strategy and evaluation studio using human expertise, data, and AI to help mission-driven teams clarify and amplify their impact.
Spot on. Your dissection of "extraction dressed as collaboration" perfectly connects with your prior work on accountability in AI. How can we truly shift teh default ownership model?