RAIN — Week of September 07, 2026

CRAI News

How CRAI Built an Open-Source AI Grant Writing Tool for Nonprofits

Ask Amy

What determines the responses that any AI platform provides? And why are the answers so different?    Rebeca from Alamosa, Colorado

Rebeca also told me which tools she uses: "I use several: Ask Leo from the Brave browser, Gab AI, Copilot and Gemini."

There are many layers here. I am going to give you the 10,000 foot view and spare you the technical nuances.

The models. Underneath every AI platform is a model. The biggest ones come from OpenAI (ChatGPT), Anthropic (Claude), and Google (Gemini). Each of those companies soaked up all the information possible from the internet. Every model was trained on all this data to answer and to make choices when we prompt it. Models do not think. They were trained to come up with the most probable answer. All the models differ, because they were built by different companies and different people. In some subjects they differ quite a bit.

The sources. If the training data was deemed insufficient to answer then the AI tool will search the web. Ask about a current event, ask for citations, ask it to double-check, and it goes looking. The go-to sources will differ between models. ChatGPT recently minimized how it cites Reddit. I took note because I often use Reddit for local answers like "who is the best propane company?"

You. If you have instructions in your AI tool, or memory turned on so it has learned your preferences, that will color the answers you receive. I have some very specific instructions in my Claude account. I do not have those same instructions in Gemini so I do not routinely use Gemini for deep thinking questions.

Rebeca, only Gemini puts you straight into one company's model. Copilot runs more than one model, but whether you get to choose depends on which version you have. Ask Leo and Gab AI are harnesses running several models underneath, and if you do not pick one the harness picks for you. Go find the picker.

Then ask two of them. Compare the answers. Ask again and get a consensus. Rinse and repeat. But always be a little skeptical. This is new technology and we are all figuring it out together.

Until next week, Amy

Every column starts with a reader question. Yours can be next: ruralai.org/rain

Stories of the Week

Artificial intelligence in rural health care: A policy roadmap for ensuring ethical and equitable usage and closing gaps in access

The National Rural Health Association has published a policy roadmap on artificial intelligence in rural health care, focused on ethical and equitable use and on closing access gaps. It is NRHA's formal advocacy position as federal rules for health AI take shape.

Rural hospital boards and clinic directors now have a citable framework to bring into vendor meetings and comment letters. Worth reading before the next ambient-scribe or diagnostic-AI purchase.

Trump officials say AI will help save rural health care. Some leaders in the field don't believe it

CMS administrator Mehmet Oz and other administration officials are pitching AI, including "AI-based avatars," as the fix for rural systems absorbing nearly $1 trillion in Medicaid cuts alongside the $50 billion rural health transformation fund. Providers across Maine and Louisiana told STAT the savings have not shown up.

Penobscot Community Health Care's ambient scribes cut documentation time but not cost, and weak Maine broadband limits anything internet-dependent. Rural leaders expect the real gains to concentrate in large systems that can afford AI-ready data infrastructure. Also covered by STAT Morning Rounds.

From barbed wire to data: How AI is changing ranching in the American West

Axios reports on AI reshaping ranching across the American West, where herd data and automated monitoring are displacing practices built around barbed wire and daily eyes on the livestock. The tools reach grazing decisions, animal health, and labor allocation.

Ranchers face real upfront costs and cellular dead zones across the same ground the sensors are meant to cover. Extension offices and ag lenders will field those questions first.

The complex corporate web behind a $3.2 billion AI data center

An update on the rural data center accountability story RAIN covered recently: a June fire at TeraWulf's $3.2 billion Lake Mariner campus in Somerset, New York found no working alarm, no suppression system, and three dead hydrants. Anthropic compute runs through the site, which Fluidstack operates and Google backstops through warrants for a 14 percent stake.

A 2019 promise of 165 permanent jobs is now 35 to 40 at full buildout, and New York's Responsible Data Center Development Act remains unsigned. Rural towns negotiating host agreements should put hydrants, job counts, and verification into the permit, not the press release.

Expanding AI access and cyber defense for federal, state, local, and tribal governments

OpenAI and GSA signed a 27-month agreement, running October 1, 2026 through December 31, 2028, that drops the $15 per user monthly license fee to $0 for federal, state, local, and tribal governments and cuts usage costs by half. Every verified government entity is approved for Daybreak Blue cyber-defense access at 50 percent off.

Tribal governments and small county offices are eligible for the first time, which matters most where one staffer covers all of IT and security. An overview webinar runs September 14 at 11:15 a.m. ET.

Most Data Center Bans Are in Metro Areas, Despite More Facilities Heading for Rural Areas

An update on the data center backlash RAIN tracked earlier: of nearly 700 local bans and moratoriums, only 25 percent are in rural counties, even though two-thirds of new AI campuses are planned there. Trackers now count 580 opposition groups with more than 640,000 members.

Wiscasset, Maine votes in November on a 180-day moratorium, and Texas has at least 248 proposed facilities, most of them rural. Counties whose land use code does not yet define "data center" are where projects land first.

Human Input: Colorado's data center fights are about more than simply anti-AI sentiment

An update on Colorado's siting fights: with two statewide bills dead since May, the rules are being written county by county. Weld County approved Global AI's permit 5-0 even though the filing listed as few as 18 permanent jobs against public promises of more than 200.

Colorado Springs residents got C-weighted low-frequency noise monitoring and ten years of utility reporting attached as permit conditions, which bind after the goodwill fades. Loveland votes on a moratorium September 15, and Colorado Springs hears appeals September 17.