AI and the Ice Cream Parlor Problem

There's an ice cream shop near where I live in Durango that's been around quite a long time. On a summer Saturday, the line runs out the door and halfway down the block.

When you get to the front, you pay on a register that looks like it was salvaged from a 1940s soda fountain. Cash only. No receipt unless you ask. The owner has been running it this way for decades and sees no reason to change.

I spent most of my career in Silicon Valley. I helped create open-source ecosystems, launched a couple of startups, and watched the data economy get constructed from the inside. I understand the theory that AI systems reflect their training data. Moving to rural Colorado brought that concept home.

That ice cream transaction doesn't exist in any structured, machine-readable way that modern AI systems are trained on. No POS record, no API call, no inventory system, no digital customer history. The sale happens, money changes hands, but it leaves almost no usable signal for the systems increasingly tasked with understanding the economy.

And this is not an outlier. It's the feed store running on handshake credit. The hunting guide who books by phone and collects cash at the trailhead. The farm stand that moves six figures in produce each season and issues zero digital receipts. The contractor whose customer list is a spiral notebook in his truck. Such transactions are how a significant portion of the rural economy operates. The activity is very real, yet the data is sparse, delayed, fragmented, or non-attributable.

Some of that is due to infrastructure. The FCC estimates roughly 28% of rural Americans still lack access to modern broadband speeds. But connectivity isn't the whole story. Even where the infrastructure exists, adoption is uneven; cost, trust, habit, and perceived value all shape whether a business fully digitizes its operations.

The result is not that rural economies are invisible. It's that they are poorly instrumented. AI systems don't learn from reality. They learn from what's measured. And large parts of rural America aren't measured in ways those systems can easily ingest.

We already see the effects. Model performance degrades as you move away from dense, data-rich environments. Geographic bias shows up in everything from business recommendations to economic assumptions. At the same time, the AI economy is concentrating in the opposite direction — Brookings Metro found the San Francisco Bay Area alone accounts for roughly 13% of U.S. AI-related job postings, and about 30 metro areas generate close to two-thirds of the national total.

That creates a compounding effect. Rural economies generate less structured data. Models trained on existing data perform worse in rural contexts. Businesses see less value in adopting those tools, so adoption stays low, so less data gets generated — and the next generation of systems inherits the same bias, often amplified. And the consequences go beyond inconvenience: credit models trained on incomplete signals misprice risk, insurance models miss context, supply chains optimize around what they can see, and policy decisions rely on datasets that systematically underrepresent entire categories of economic activity.

That ice cream shop is doing just fine without being part of the data economy. The question is what happens as more decisions — financial, operational, and political — are made by systems that have no meaningful record of its existence. And more importantly: does rural America find a way to become instrumented on its own terms, or does it remain an afterthought in systems that increasingly define the mainstream economy?

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.