The AI divide is already here, and it's widening faster than the digital divide ever did.

Here is a strange fact about the United States in 2026. This country builds most of the world's frontier AI models. It hosts roughly three quarters of the global data center capacity under construction.¹ Its companies set the pace for the entire industry. And yet, when Microsoft's AI Economy Institute ranked countries by what share of the working-age population actually uses AI tools, the United States came in 21st.²

Twenty-first. Behind the UAE, behind Singapore, behind a long list of smaller, highly digitized nations. The country that invented the technology is not the country that uses it most broadly.

That gap between building AI and using AI is the first clue that something is off. The second clue sits inside the national numbers. Just over 30 percent of working-age Americans now use AI tools, up three points from the end of 2025.³ Sounds like healthy growth, until you break it down by county. In metropolitan counties, usage averages 32.9 percent. In rural counties, it's 16.2 percent.³ Half. Not a few points behind. Half.

This is the AI divide: the widening gap between the people, places, and organizations that are putting AI to work and those that are not. It follows the fault lines of the old digital divide, which is depressing but not surprising. What should worry us more is that this divide is compounding faster than any technology gap America has seen, and the problems it creates are already visible if you know where to look.

A divide with three layers

When people hear "AI divide," most picture an access problem: some people have the tools, others don't. That framing is comfortable because access problems have familiar fixes. Build the broadband. Subsidize the devices. Done.

But access is the layer where the AI divide is smallest. ChatGPT, Claude, Gemini, and Copilot are free or nearly free, and they run on any smartphone. A general-purpose technology has never been distributed this widely, this fast, at this low a price. Previous general-purpose technologies took decades to reach most of the population. Electricity needed about fifty years.⁴ Generative AI reached hundreds of millions of people in under two.

So if the tools are essentially free and available to anyone with a phone, why do adoption rates diverge so sharply? Because the divide isn't primarily about access. It has two deeper layers.

The second layer is usage and skill. Owning access to a tool and knowing how to get value from it are different things. Researchers analyzing billions of search queries found that early interest in ChatGPT clustered heavily on the West Coast and stayed persistently low across Appalachia and the Gulf states. When they controlled for everything they could measure, education emerged as the single strongest predictor of who engaged with generative AI.⁵ Not income. Not age. Education. The people best positioned to extract value from a thinking tool are the people who already had the most invested in thinking work.

The third layer is organizational. Individuals adopt tools; organizations adopt systems. A Census Bureau working paper on AI adoption across American firms found what its authors called an AI divide within the business population itself: a small set of firms with particular characteristics, typically larger, younger, better capitalized, and already tech-forward, leading the diffusion curve, with a long tail of laggards behind them.⁶ Whether your employer builds AI into daily workflows matters more for your career than whether you personally experiment with a chatbot on weekends. And that decision is completely out of most workers' hands.

Three layers, then: access, skill, and organizational integration. America has mostly solved the first. The second and third are where the divide lives, and they're the hard ones.

The map tells the story

Pull up a county-level map of AI usage and you're looking at a familiar picture. The coasts glow. University towns glow. The corridors around Austin, Atlanta, Nashville, and Columbus glow. Large stretches of rural America, Appalachia, and the Deep South stay dim.

Microsoft's data scientists found something uncomfortable when they dug into the rural-urban gap. The obvious explanations, older populations, lower incomes, less education, explain part of it. But even after controlling for all of those factors, a large gap remains. Juan Lavista Ferres, who leads the research, called the finding striking and noted that this pattern matches the structural rural-urban technology divides he grew up seeing in Latin America, divides that people there treated as permanent facts of life.⁷

Let that sink in. The pattern of technology inequality that Americans have historically associated with developing economies is now measurably present inside the United States, and it persists after you strip out every demographic explanation.

Why does a usage gap on a map matter? Because AI adoption isn't just a consumption statistic, like which counties watch more streaming video. Early evidence suggests it functions as a leading indicator of productivity and wage growth. The counties adopting AI fastest are positioning their workers and businesses to produce more per hour, command higher wages, and attract investment. The counties adopting slowest are not standing still. Relative to everyone else, they're falling backward.

Rural America has lived through this movie before. Electrification, interstate highways, broadband: each time, the places that got infrastructure late spent decades catching up, and some never did. The difference this time is speed. Broadband diffusion gave laggard regions a decade or more to close gaps. AI capability is improving on a cycle measured in months. A county that's two years behind on AI adoption in 2026 isn't two years behind a fixed target. It's two years behind a target that's accelerating away.

Education compounds the problem instead of fixing it

The finding that education predicts AI adoption better than any other variable deserves more attention than it gets, because it inverts one of the great hopes for this technology.

The optimistic story about AI was always egalitarian. These tools write, analyze, code, and summarize. In principle, they hand every person capabilities that used to require an expensive degree. A talented worker without credentials could use AI to close the gap with credentialed peers. Early studies of AI in customer service and writing tasks reinforced this hope, showing the largest gains going to the least experienced workers.^8,9

That leveling effect is real within organizations that deploy AI well. But at the population level, the opposite dynamic is winning. The people using these tools most intensively are the already-educated, already-connected, already-advantaged. Researchers studying household internet behavior found the adoption gap between high-income and low-income households, and between younger and older households, isn't shrinking as the technology matures. It's growing.¹⁰

This is what scholars of the original digital divide called the second-level divide,¹¹ and we appear to be replaying it at higher speed. In the 2000s, the first-level question was who had internet access. By the 2010s, access had spread, but researchers found a deeper split in how people used it: some households used the internet for education, income, and skill-building, while others used it almost entirely for entertainment. Access equalized; benefit didn't.

The same split is forming around AI, except the stakes are higher. The internet mostly changed how people found information. AI changes how people produce work. A college-educated professional who spends 2026 learning to delegate research, drafting, and analysis to AI is building a compounding productivity advantage. A worker who encounters AI only as the system that screens their job application, sets their gig-work rates, or denies their insurance claim experiences the same technology as pure downside. Both are living in the AI economy. Only one of them is benefiting from it.

Small organizations are on the wrong side of the line

The organizational layer of the divide may end up being the most consequential, because it determines which layer of the economy captures AI's productivity gains.

Large enterprises have AI strategies, AI budgets, AI governance committees, and increasingly, AI teams. They can afford enterprise licenses, custom integrations, and consultants. They can absorb a failed pilot and try again. The Census Bureau research found exactly what you'd expect: firm size, age, and prior technology sophistication strongly predict AI adoption. The firms already leading pulled further ahead.

Now consider the other end of the spectrum. The United States has roughly 33 million small businesses¹² and about 1.8 million nonprofits.¹³ Most operate with no IT department, let alone an AI strategy. The executive director of a 20-person nonprofit is not evaluating retrieval pipelines. She's trying to make payroll and file the 990. For organizations like hers, AI adoption comes down to bandwidth, and the bandwidth doesn't exist.

The problems this creates run in two directions.

First, competitive: small firms compete with large ones for customers, contracts, and talent. If AI delivers even a modest productivity edge and that edge accrues mainly to organizations above a certain size, market concentration accelerates. The local accounting firm, the regional insurer, the independent retailer all lose ground to national players whose cost structures now bake in AI-driven efficiencies theirs don't.

Second, and less discussed: the organizations furthest behind on AI adoption are disproportionately the ones serving the populations furthest behind on AI adoption. Community nonprofits, rural health clinics, legal aid offices, small school districts, food banks. These organizations serve the low-income, rural, older, and less-educated Americans on the wrong side of every adoption statistic in this piece. When a well-funded hospital system uses AI to cut administrative costs and a community clinic can't, the clinic's patients pay twice: once through the general divide, and again through their institutions' place in it.

The AI divide, in other words, runs between institutions as much as between individuals, and it stacks disadvantage on top of disadvantage.

The problems are not hypothetical

The damage is easier to see in specifics, because "some people use AI more than others" undersells it badly.

Start with wages. If AI adoption tracks with productivity growth, and productivity growth tracks with wage growth, then the adoption map is a preview of the wage map. Regional income divergence in the United States, which narrowed for most of the twentieth century, has been widening since the 1980s.¹⁴ AI is set to pour accelerant on that trend. The metro areas capturing AI-driven productivity gains will pull further away from the rural and post-industrial regions that aren't, and the political consequences of regional economic divergence are already a defining feature of American life. We do not need more of them.

Then labor markets. AI exposure isn't evenly distributed across occupations, and neither is the ability to adapt. A worker whose employer trains them to work alongside AI experiences the technology as augmentation. A worker whose employer simply automates their tasks, or who works in a small firm with no capacity to retrain anyone, experiences it as displacement. The difference between those two outcomes often isn't the worker's talent or effort. It's which side of the organizational divide their employer sits on. That's a recipe for a labor market where similar workers doing similar jobs end up on wildly different trajectories based on institutional luck.

Then the information environment. AI literacy doubles as a defensive skill. People who understand how these systems work, roughly what they can and can't do, and how they fail are better equipped to spot AI-generated scams, synthetic media, and manipulated content. Communication researchers studying AI competence have found that users with low AI literacy are the most vulnerable to algorithmic manipulation, misinformation, and data-driven exploitation.¹⁵ The same populations lagging in adoption are the populations most exposed to AI's misuse. They're getting the risks without the benefits, which is the worst possible trade.

Then public services and civic life. Government agencies, courts, and benefits systems are deploying AI to process claims, flag fraud, and allocate resources. The people subject to those systems, disproportionately lower-income Americans, have the least understanding of how the systems work and the least capacity to contest their errors. An algorithmic denial that a tech-savvy professional would recognize, document, and appeal simply stands for someone who doesn't know an algorithm was involved. The divide determines who benefits from AI, and it determines who can defend themselves against it.

And finally, trust. Attitudes toward AI in America are polarizing along the same geographic and educational lines as usage.^16,17 That makes sense: if your only contact with AI is hearing that it might take your job, skepticism is rational. But it creates a vicious cycle. Low adoption breeds distrust, distrust suppresses adoption, and the communities that most need the productivity gains talk themselves out of the tools that could deliver them. Meanwhile, national AI policy gets written by and for the high-adoption side of the divide, which deepens the alienation. A general-purpose technology that half the country embraces and half the country resents is not a stable foundation for anything.

Why this one moves faster

Every point above had an analogue in the digital divide of the 1990s and 2000s, and the country eventually muddled through that one. So why treat this as more than a rerun?

Three reasons.

The first is compounding. Internet access was mostly a binary: once you had it, you had roughly what everyone else had. AI skill compounds. Every month of hands-on use makes the next month more valuable, because users learn to delegate bigger tasks, chain tools together, and redesign how they work. The gap between an experienced AI user and a non-user isn't static like the gap between a connected and unconnected household. It grows with time, automatically.

The second is that the technology itself is improving underneath the adoption curve. The AI a laggard eventually adopts in 2028 will be far more capable than what leaders use today, which sounds like good news for late adopters. It isn't, because the leaders will have moved to whatever exists in 2028 and will know how to use it. When the target accelerates, the distance behind grows on its own.

The third is that the window for cheap intervention is short. Adoption patterns are hardening right now. Organizational habits, regional attitudes, and educational trajectories being set in 2025 and 2026 will shape who benefits from AI for the next decade. The digital divide took twenty years to fully harden into today's map. This one is hardening in three or four.

What closing it would actually take

The instinct will be to reach for the old playbook: infrastructure spending and device access. That playbook targets the layer of the divide that's already mostly solved. The real work sits in the other two layers, and it's less glamorous.

Skill has to be treated as the primary bottleneck. That means AI literacy woven into community colleges, workforce boards, libraries, and adult education, not as computer science but as practical fluency: what these tools do, how to get value from them, how to check their output, and how to spot their misuse. The research is unambiguous that education drives adoption,⁵ which makes education systems, formal and informal, the best point of intervention available.

Organizational adoption needs its own strategy, because small organizations won't close the gap alone. The MSPs, industry associations, community foundations, and regional development agencies that already serve small businesses and nonprofits are the natural delivery channel for AI capability, the same way agricultural extension services once carried farming innovation into every county in America. Some of that support will come from the market. Some of it, particularly for nonprofits and rural institutions, will need philanthropic and public backing, because the organizations that need help most can least afford to buy it.

And the communities behind on adoption need a reason to engage that isn't fear. Nobody adopts a technology they've only heard described as a job destroyer. The most persuasive case for AI in a rural county isn't a Silicon Valley keynote. It's the clinic that cut its billing backlog, the family farm that automated its compliance paperwork, the small manufacturer that won a contract it couldn't have serviced before. Local proof beats national hype, every time.

None of this is exotic. All of it is doable. The question is whether it happens while the divide is still a gap, or after it has hardened into a structure.

The United States has run this experiment before. It let a powerful new technology diffuse on market logic alone, watched the benefits pool where advantage already lived, and then spent two decades and tens of billions of dollars trying to retrofit equity onto a finished map. The AI divide offers a rare second chance to act early, while the map is still being drawn. The data says we have a few years. The pattern says we'll waste them unless enough people decide not to.

Talk with an advisor about closing the gap →

↑ Back to top

Who publishes this

This report is published by ETTE (Empowerment Through Technology & Education, Inc.), a managed IT services provider based in Washington, DC. ETTE spends most of its time with exactly the organizations this report identifies as furthest behind — small businesses and nonprofits with no IT department, let alone an AI strategy — helping them figure out where AI actually helps, building the guardrails to adopt it safely, and training staff to use it well.

We wrote this because we kept seeing the pattern described above play out client by client before we saw it confirmed in the data. If something here is wrong, out of date, or missing a source, tell us and we'll fix it.