Research & Development
Rural facts get flattened when AI uses them: the place, the year, the source all fall away. We are researching how to keep them attached.
Why this research matters
AI models see, experience and learn about rural America like tourists, not locals.
AI systems learn from what has been written down, digitized, and published. Rural America is thin in that record at every stage: less documented online, served by shrinking local news, with much of what it knows living in archives, organizations, and lived experience that training pipelines never reach.
The gap is measurable. The first systematic audit of place-based knowledge in large language models found that models answer questions about urban counties with significantly more accuracy and depth than rural ones. The shortfall concentrates in what defines a community: its events, institutions, and social fabric.
As AI moves into health care, public benefits, lending, and everyday business software, that incomplete record starts shaping decisions with real economic weight. Our research makes the gap visible and measurable, and tests whether it's fixable.
That gap will matter more, not less. As access to data commoditizes and synthetic training data floods the market, volume is no longer the scarce input. What becomes scarce is preserved meaning: knowing where a fact came from, what it actually describes, and whether it holds for a particular place and time.
What a thin record looks like
In USDA farm data for southwest Colorado, Montezuma, Archuleta, La Plata, and Dolores counties each carry several hundred entries. Neighboring San Juan County has one. Ask a model about farming in San Juan County, and it isn't reasoning from thin evidence — it's reasoning from a single point, or quietly borrowing from the counties next door. The answer will still sound confident. That's the problem.
What this is for
Rural communities are surrounded by resources, programs, funding, and data. Finding what applies to a specific place, and understanding how to act on it, is the hard part.
We are researching how AI can make that knowledge reachable by giving it better rural context: not more data, but an understanding of the place, the evidence, and the rules that determine what's relevant.
Use Cases
Finding the right funding
A rural economic development organization wants federal grants that could support a broadband, workforce, housing, healthcare, or infrastructure project. AI can find programs. Determining which ones actually fit takes the community, the eligibility rules, the competing definitions of rural, local conditions, deadlines, and supporting evidence, together.
Connecting communities to programs
A local leader knows their community needs childcare, healthcare access, agricultural support, workforce development, or broadband, but not which federal or state programs address it. Instead of searching agency by agency, they ask in plain language and get relevant programs grounded in evidence about their community, with sources they can verify.
Building a case for investment
Demonstrating need means assembling facts scattered across Census, USDA, federal programs, university research, regional studies, and local knowledge, while preserving where each one came from, what it means, and where it stops being reliable.
The core benefit
Turn fragmented rural information into trustworthy, actionable context, so rural communities can find and pursue the resources that actually fit them. That serves the people doing this work directly: economic development organizations, nonprofits, grant writers, funders, and policymakers, and eventually the AI applications built to serve rural places.
Everything below this point — the library, the benchmark, the governed evidence layer — is our hypothesis for how to deliver that reliably. The architecture is not the benefit. Whether it helps someone find the right grant is.
Rural Intelligence
Open research questionThe layer between authoritative evidence and AI
Rural Intelligence is the move from rural knowledge retrieval to rural knowledge understanding: turning authoritative rural evidence into context AI can use without losing what makes it true.
We are not recreating the underlying data, and we do not own it. Census, USDA, and the other federal statistical systems already publish it. The open question is the interpretation layer in between, and whether building it carefully produces measurably better answers than handing a model the raw sources.
What has to be preserved
Which place, at which grain — county, tract, place, or region.
Which vintage and reference period, and whether two figures share one.
Which definition of rural a question actually calls for.
Margins of error, suppression codes, and sparse small-population data.
Which source governs when sources disagree.
An unbroken path from an answer back to the evidence under it.
Whether two pieces of evidence can be placed side by side at all.
Who decides how community knowledge may be used, and on what terms.
Where the research stands
Two rungs are built and measured. The rest are open. We'd rather say that plainly than describe a finished architecture we haven't earned.
- 01BRAINBuilt · in daily internal use
Curated rural articles, research, and policy documents made available to AI tools with citations.
- 02BRAIN BenchBuilt · first run published below
Test whether curated rural context actually improves AI answers, rather than assuming it.
- 03Structured evidence researchIn progress
Work with Census, USDA, and 22 other federal statistical systems to find what rural meaning breaks when evidence moves between systems.
- 04Rural IntelligenceOpen research question
Build and test the governed layer that turns accessible rural evidence into context an AI system can be trusted with.
- 05ProofDefining the model
Demonstrate that governed rural context produces measurably better answers on consequential rural questions.
BRAIN
First prototype · Internal useBedrock for Rural AI kNowledge
BRAIN is a curated, cited library of rural-AI intelligence — a centralized knowledge base built for our team and our AI tools to use. It plugs in as an MCP connector, so any tool we run can draw on the same shared, vetted memory rather than searching scattered files one query at a time.
It replaces scattered knowledge, repeated data-gathering, and dependence on any one person's storage with a single searchable library where every answer carries a citation.
BRAIN began with a deliberately simple hypothesis: can AI answer rural questions better when it has access to a purpose-built body of rural knowledge? It's a real working library, not a concept. It's also the first prototype in the research above, not the finished architecture we intend to build.
Built, and our team uses it every day
What Lives in BRAIN
What's Live & Working
- AI tools query BRAIN directly
- Data is auto-ingested and approved by CRAI
- Eval-tested and ready for handover
What building it exposed
The evidence that answers real rural questions also lives in Census, USDA, other federal statistical systems, university research, and regional sources — not just in articles and reports.
Pointing a model at those sources, or loading them into a vector database, doesn't tell it whether the evidence is current, comparable, authoritative, suppressed, or geographically right for the question.
Meaning, provenance, geography, time, uncertainty, and authority all have to survive the trip from source to answer. That's the open problem above, and it's bigger than BRAIN itself.
A small nonprofit can build a curated, AI-ready knowledge library. Building one is how we found the harder problem underneath it.
Benchmark
In DevelopmentHow well do today's AI models understand and represent rural information?
Benchmark tests how well today's leading AI models, including Claude, Gemini, and GPT, perform for rural users. Today, these models default to urban, data-rich environments. This project turns that gap into a public, comparable score.
It won't fix the problem on its own. But instead of asking people to take our word for the metro–rural tech-equity gap, it makes that gap visible and checkable. The project is in active development.
How It Will Work
Pose a set of 50+ rural-specific questions to each model and its sub-models.
Grade answers against a rubric: accurate terminology, avoids urban defaults, reflects rural reality.
Post the results publicly so anyone can see where the gaps are.
Re-run on a regular cadence to track how models change over time.
BRAIN Bench
Pilot EvaluationThe first step in closing the Rural AI Gap.
CRAI invested in BRAIN on a strong hypothesis, but without proof. There was no quantitative measure of whether BRAIN's answers were more accurate than an AI model working on its own, and no reliable way to pin down where its knowledge gaps were or how to close them. BRAIN Bench is the evaluation that turns that hypothesis into evidence.
It grades real prompts, many drawn straight from CRAI's own first users, across nine industry categories and six leading models, scoring each answer against a four-dimension rubric. The result moves BRAIN from guesswork to something CRAI can measure, repeat, and improve as the tool grows toward a public, rural-facing resource.
First-run results: blended score by model
CRAI's BRAIN configurationbaseline models, tested as-is
| Model | Accuracy | Completeness | Rural nuance | Honesty | Avg tokens | Avg time |
|---|---|---|---|---|---|---|
| Claude + BRAIN MCP | 100% | 58% | 81% | 100% | 400 | 8.0s |
| Sonnet 5 | 90% | 72% | 82% | 91% | 2,056 | 27.7s |
| Claude Haiku 4.5 | 84% | 75% | 69% | 88% | 609 | 6.4s |
| GPT-5.6 Sol | 82% | 65% | 71% | 84% | 1,653 | 35.9s |
| GPT-5.6 Terra | 78% | 60% | 66% | 79% | 1,914 | 23.8s |
| Claude Opus 5 | 68% | 55% | 62% | 66% | 5,120 | 78.1s |
Read this as a first measurement, not a victory lap. BRAIN leads on accuracy and honesty: it answers from cited sources or says it doesn't know. But its one-point margin over the strongest baseline is well inside the noise of a 100-prompt run, and its lower completeness score reflects notably shorter answers. Pinning down that tradeoff is exactly what the benchmark was built to do.
BRAIN Bench is the origin, not the destination. The same method now scales up into the Rural Intelligence Benchmark below. Built by Emaliah Sawyer.
Next: the Rural Intelligence Benchmark
DesigningAs the research expands from a curated library toward governed rural context, the benchmark expands with it. The next phase compares four ways of answering the same rural question, to test whether a governed interpretation layer beats simply giving a model more data.
No added rural context at all.
Direct access to services such as Data Commons or federal APIs and MCP servers.
Retrieval over raw or vectorized source material.
Retrieval over CRAI's governed rural evidence layer.
What we will measure
We expect the hardest tests to sit where rural evidence is easiest to misuse: geographic grain, competing definitions of rural, mismatched data vintages, suppressed or uncertain estimates, sparse small-population data, conflicting sources, and questions that need evidence from several systems at once. We care most about the consequential cases, including grant eligibility, broadband support, federal funding, workforce, agriculture, and community planning, where a plausible but poorly grounded answer sends someone to the wrong resource. The benchmark isn't designed to show that CRAI has more data. It's designed to test whether CRAI helps AI understand rural evidence better.
Grounded in Real Rural Expertise
Subject-matter experts help build the questions and grading rubric and validate the answers, so scoring reflects real rural experience. The expanded benchmark raises that bar: measures like comparability and rural-definition correctness can't be scored without people who know the domain. From there, we'll build an agentic grading system so the benchmark can run without relying on subject-matter experts every cycle. Any rural or tribal dataset the benchmark draws on or publishes routes through our Tribal & Community Data Sovereignty Review Process. Community governance isn't a footnote to this work. It's one of the properties the research is trying to preserve.
Why this matters
A model that answers a rural question wrongly does not look wrong. It looks fluent, cites something plausible, and sends a community after the wrong grant or the wrong program. As AI moves further into funding, health care, and local decision-making, the cost of that stops being abstract.
The evidence rural communities need already exists. Our work is to make it usable without losing what makes it true, and to prove in public whether that actually produces better answers.