Distrust: The Most Effective AI Operating Model
By Andrew Aitken, Founder and Executive Director, Center for Rural AI
Every AI answer is a claim to be tested, and what better way than using another model for verification?
Rural people learned a long time ago that trust is something you earn, not something you hand over. You check the hours on a used tractor before you buy it. You read the water-rights agreement twice before you sign. You call a neighbor who has done the job before you start it yourself. Verify first, rely second. It is a habit born of distance and thin margins, and it's exactly the discipline the AI economy needs right now.
At the Center for Rural AI, we have built that habit into how we work. Not distrust of any one tool, and not cynicism about what AI can do for rural communities. What we practice is structured distrust: we assume any single model can be confidently, fluently wrong, and we believe the fastest way to expose that error is making the models argue with one another.
A model that sounds certain has told you nothing about whether it is right. Useful information lives in disagreement.
How it works in practice
Every piece of work we produce starts in Claude — drafting, reasoning, structuring the problem, development. From there, we hand it to a rival:
When the work is code, ChatGPT is the adversary. We give it Claude's output with one job: break it. Find the edge case, the security hole, the assumption that holds in a demo but fails on a rural clinic's aging hardware or a farm with spotty connectivity.
When the work is research, Gemini is the adversary. We take a claim or a synthesis and ask a different model to challenge the sources, flag what is overstated, and identify what was left out.
The version that survives the cross-examination is the version we ship. Anything that cannot withstand a second model trying to tear it down is not ready to put in front of a stakeholder.
Why distrust is a design choice
Every model carries the blind spots of the data it was trained on. For rural work, that is not an abstraction. AI systems trained overwhelmingly on urban, English-dominant data default to answers that make sense in San Francisco and quietly fall apart in the Navajo Nation or rural Wyoming. A single model will deliver those wrong answers with the same polish it brings to the right ones. Confidence is not a signal you can rely on.
Adversarial checking catches what confidence hides. When two models reach the same conclusion by different routes, that agreement means something. When they disagree, the disagreement is a map; it points straight at the claim worth verifying by hand. Either way, we learn more than one model alone could tell us.
This is not a novel idea so much as an old one wearing new clothes. Peer review, red-teaming, a second set of eyes on a grant before it goes out the door: every serious field already knows that important work should be attacked before it is trusted. We are applying that standard to AI itself.
What this means for the people we serve
For the rural colleges, communities, and funders we work with, the cost of being wrong is real and local. A mistaken figure in a grant proposal, a security flaw in a community-built tool, a research claim that stretches past its evidence; each of these costs trust that rural institutions cannot easily earn back. Adversarial review is how we protect that trust before it is at risk.
It is also why we benchmark. Our work measuring how accurately large language models handle rural data answers a specific question: where can these tools be relied on for rural questions, and where can they not? Knowing the boundary is what lets us use AI aggressively inside it and cautiously outside it.
Used well, distrust is not paralysis, and it is not pessimism. It is how you earn the right to be confident. We would far rather find the flaw in our own work than have a stakeholder find it in the field — and the surest way we know to find it is to let a rival model go looking first.
Andrew Aitken is the Founder and Executive Director of the Center for Rural AI (ruralai.org), a fiscally sponsored project of SW Community Foundation based in Durango, Colorado. CRAI is partnered with the AI Institute at Fort Lewis College.