AI Needs a Grange

One block from the White House stands a 12-story building, the National Grange of the Patrons of Husbandry, erected in 1960—long the lone private holdout on a block of executive branch offices. The Grange isn’t the political force it once was. Still, its presence so close to the president reminds us that one of America’s most consequential community movements understood that local grievances sometimes require national attention, even when they resist entirely national solutions.

The Grangers are often misremembered as a populist farmers’ movement that protested the power of the railroads and big business generally. This isn’t quite true: Nearly everyone in the Grange wanted railroads, supported freer trade, and pursued international markets. In many ways, its worldview was more cosmopolitan than the Gilded Age’s big businesses, which largely favored protectionist tariffs.

What got Grangers engaged wasn’t the technology or big business itself, but decisions with vast, sometimes existential, consequences for their communities being made without them. Railroads were American big business and the most pervasive technology of the Second Industrial Revolution; their choices about everything from routing to rates to timetables could determine a community’s fate, without anyone asking the community. Granges’ agitation ultimately helped set in motion the Interstate Commerce Commission, uniform national standards for railroads, and, over time, rate regulation—which addressed real problems.

The Grange movement should inform the current debate over AI, especially its belief that decisions belong near the people who live with them, and that they cannot decide well without honest information. The two of us—one an executive at a nonpartisan national foundation, the other president of a right-of-center think tank—arrive at that same conviction from different directions and don’t agree on every remedy.

To a family weighing a new data center, artificial intelligence isn’t just an abstraction. It is whether the water table drops, whether the power bill climbs (or falls), whether new jobs are the jobs that were promised. Surveys find ordinary people are more skeptical of AI’s benefits than the experts and executives building it. And it’s obvious why. Such decisions affect the most elemental things a community has—its land, its water, its work, its sense of itself. That may explain the bipartisan, highly emotional reaction to data centers. Even when the industry answers one of these concerns—new data centers almost all have “closed loop” water systems that won’t touch aquifers—new ones arise.

Isaiah Berlin argued that many of humanity’s most important values are plural. Liberty and security, innovation and stability, efficiency and self-determination can all be worthy ends but can’t be reconciled. No formula will maximize them all at once. Communities will reasonably disagree about the tradeoffs they are willing to make between economic growth, privacy, environmental protection, employment, and cultural preservation. If Berlin was right, then governance cannot concoct a single perfect answer for everyone.  Pluralism doesn’t always beat standardization, of course. A country needs one system of weights and measures. We needed one railroad gauge. But a standard of that kind is an interface, not the goal. We both believe the balance is to standardize the connections where society must, and leave the choices about what to build on them to those who live with the results.    

Homogeneous AI models carry a real risk of flattening impressions across the population: Small nudges toward common forms of expression or common ways of approaching a question can become socially significant when they occur across hundreds of millions of interactions each day. Reviewing evidence across linguistics, psychology, and cognitive science, a recent paper in Trends in Cognitive Sciences argues that large language models do indeed decrease human diversity along three dimensions at once—language, perspective, and reasoning—by mirroring the dominant patterns in their training data, and that the effect compounds as more people lean on the same handful of models.

In dealing with AI, as with other new technologies, therefore, a society should preserve a diversity in decisions and approaches, ensure communities have clear information, and centralize authority only where an identified externality crosses community lines. This is roughly what the Grange ended up with: state action first, which didn’t work and proved impossible anyway, then federal rules for what plainly outran state lines.

Almost everything then turns on scale, and the current debate gets scale wrong, focusing nearly all its attention on the states and the federal government. Local governments should never try to regulate an AI model, which lives on a global network and, to the extent it needs specific governing, will have to be governed federally or internationally. The need for federal privacy laws—ideally, uniform ones that preempt state standards—is already clear; other classes of regulation may become so in time.

Our argument concerns two places where AI touches ground: the physical plant that sits on a community’s land and draws on its water and power, and the decisions local institutions make about how to use these tools.

Much of what AI will change most—how teachers teach, how lawyers advocate, whether a data center gets built—happens in communities. A locality that refuses a data center should be free to do so; the costs it weighs—water, noise, land, aesthetics, the tax base—fall mostly on its own residents, and the facility will find a home in one of the many places competing to host it. It’s connected to the global internet, so its location doesn’t matter. A county that blocks an interstate transmission line is different: It pushes costs onto ratepayers hundreds of miles away, which is a spillover that justifies federal preemption. (Pending federal legislation, intended largely for the energy industry but clearly relevant to data centers, offers such preemption.)

That the affluent will enjoy AI while exporting its burdens has a similar answer. The fix is not to take decisions away from poorer places, but to hold data center builders to their promises to pay their own way and more. In any case, the Constitution already forbids a locality to discriminate against interstate commerce; nothing we propose changes that.

The most useful thing the national and state governments can usually do is to guarantee that those closest to AI’s effects have the information to make good decisions.

The most serious people working on AI governance have already given careful thought to community voice. The Collective Intelligence Project argues that democracy is a set of accountable institutions that “process and act on decentralized information.” We agree. Where we part ways is on what comes next. CIP’s instinct is to gather public preferences through carefully designed assemblies and route them into the places where models are built and released: training, fine-tuning, and even the decision to ship a new model at all. It is a thoughtful design. But we see it differently.

If knowledge really is decentralized, the answer is to leave the decision where people live, with those who have a real stake in it. A thousand school boards each deciding what to do with a chatbot are processing decentralized information. A representative sample convened to advise a developer or a government, however thoughtfully assembled, gathers it up and hands it over. Peculiar local circumstances can’t be taken into account.

Unlike developers, local authorities are answerable to the parents whose children use these tools. If a product is shown to harm children in ways no board would knowingly accept, existing child-protection and consumer law should apply everywhere, and if not, new laws can be written. Short of that, it’s better to have thousands of school boards make different mistakes than one national authority make the same mistake for every child in the country.

The instinct to keep such decisions plural is global. Audrey Tang, Taiwan’s first digital minister, argues that governments should reach for a national rule only where the risk is genuinely systemic, leaving the rest to deliberation closer to the ground. Tang and the economist Glen Weyl put the structural problem plainly: A uniform rule ends up either too rigid or too complex to function. They use the speed limit as an example, noting that it is often set too high or too low for the road. Indeed, the United States had a federal speed limit of 55 miles per hour for over 20 years, and almost nobody liked it.

What about states? A typical state is too big to weigh local peculiarities; California and Texas hold as much difference inside their borders as the country did a century ago, so a single statewide rule reproduces, on a smaller map, the misfit of a single federal speed limit. Yet the state is also too small to catch the problems that genuinely cross community lines. If a risk is so large that no community can manage it, it is almost by definition national or international. Fifty separate regimes will neither contain it nor agree with one another.

States tried to set railroad rates in the 1870s under Granger pressure. But as more of what a farmer shipped crossed state lines, state rate-setting mattered less, and once the Supreme Court held in the 1886 Wabash case that a network spanning many states could be governed by none alone, the question moved to Washington; the Interstate Commerce Commission followed within a year. AI follows the same pattern. The models, like the rates, are a national matter; the data center, like the depot, sits in a particular town.

This is why many recent state laws pose problems. Two kinds of laws failed this test, in opposite directions. The first, California’s SB 1047, vetoed in 2024, reaches up. Its narrower successor, SB 53, along with New York’s RAISE Act (which, as amended in March 2026, is quite similar to SB 53), focuses on frontier AI models themselves—the handful of largest systems regulated for genuinely catastrophic risks like a model’s contribution to a chemical or biological weapon. But if this threat is genuine (and it’s unclear whether it is for current models), that risk is national, even global, in scale. A single state cannot contain it. The second wave reaches sideways, dropping a wholly new regime onto activity that was never regulated at the state level. Colorado’s never-enforced 2024 AI Act, the first of its kind, would have demanded elaborate compliance—duties of care, risk assessments, management programs—on anyone deploying AI in employment, housing, lending, and the like. Within two years, it was challenged in court, its enforcement was stayed, and the statute was repealed and replaced with a narrower framework built around disclosure. Many of the 2024 bill’s stated goals were unobjectionable and even laudable—“the AI did it” should never become an excuse for violating civil rights laws—but implementation was nearly impossible. It would have built a state bureaucracy before anyone could weigh the externalities or even knew what they were. And it points to a broader fact: Nearly all regulation of the models themselves is essentially impossible to implement at the state level, since the models live on an open, public internet.

That doesn’t mean a state should sit on its hands. The two of us don’t agree on every adjacent question—state-level child-protection standards, say, or employment law—but even there we can agree on a test for state measures: Does it enlarge local decision and debate, or collapse it? That points toward a few guidelines: a preference for disclosures over outright prohibition, for liability standards over rigid rules, and for leaving communities room to set their own course.

The same test runs in the other direction. Preemption is a form of centralization, and it should be used only when a specific problem crosses state lines, and there is evidence that states and communities can’t handle it. We agree that some things shouldn’t be touched at all: local zoning decisions on data centers, state tax policy toward AI, generally applicable civil rights and consumer protection law, and the rules a state sets for its own use of AI. Most serious federal legislative proposals leave these alone.

Good local decisions require transparency. When a data center developer seeks a zoning variance, a subsidy, or favorable tax treatment, communities are sometimes asked to decide without the necessary information. A county board of a few volunteers and a part-time clerk must contend with a corporate behemoth that has done this a hundred times in a hundred towns. How much water will the data center consume? Will the town need new generation capacity, and what will that do to rates? How many jobs will it create? Companies often have detailed answers to these questions. Communities often don’t. That means communities routinely make hard-to-reverse decisions without the right information. Sometimes the problem is worse than asymmetry. In the summer of 2026, local officials in Louisiana’s Rapides Parish negotiated a data center deal under nondisclosure agreements, and residents learned what was coming only after the fact. Pennsylvania Governor Josh Shapiro has since barred the agencies he oversees from signing such agreements on data center projects. Disclosure at the decision point is a community’s best protection against government officials end-running the democratic process.

The Grangers’ core grievance was similar: Railroads could know what routing, schedule, and rate decisions would do to specific communities, and those communities learned only after the fact. Towns competed fiercely for the railroad and sometimes subsidized it. Boosters usually won in the long run; a rail yard might bring jobs to one town and become a noisy eyesore in the next, but at least each town could argue about it. What towns could not see, or contest, were the rates that would decide whether their farmers and merchants could compete once the line arrived.

The federal answer dealt mostly with the rates—although the actual ability of the ICC to regulate varied quite a lot over time—but the system did result in relatively uniform national rate regulation by the 1940s. In an environment with slow communications and little data, this was a blunt response and, ultimately, it proved ineffective in maintaining a healthy railroad business: over a fifth of the nation’s freight mileage was owned by bankrupt companies in the 1970s. This is why disclosures at specific decision points, not a national planning bureaucracy or sweeping bans, may be the best option: they let communities do what they already can, and do it better, without imposing uniform rules.

Much of this is an admission of what is not yet known. Outside the catastrophic cases, we should be slow to move against a harm before we are sure one exists and have evidence that the remedy we propose will address it. Whether AI tutoring helps or hurts an eighth grader’s social development, whether a town’s automated permitting speeds decisions or alters the character of the place, cannot be settled in a lab, and the answer may differ between communities. That’s why it’s best to answer when and where the data shows up.

The best protection against avoidable losses is variety itself. Fifty states, more than 35,000 sub-county governments, and several thousand school districts, each setting its own course, are how a society finds out what works best: a slow, distributed experiment no central body could run.

The problem is that an experiment only teaches if its results travel. Letting each community choose differently does little good unless each can see what the others have learned. A thousand communities, each striking their own bargain with an LLM, where the costs of that bargain fall mainly on the people who strike it, are running an experiment whose results society needs.

In other words, messy is often good. Americans reached that conclusion once Prohibition failed, handing choices back to states, counties, and towns. When a genuine externality does come into view, the task is to identify it and where its costs fall; otherwise, let communities weigh the risks against the rewards, settle on what success would look like for them, and work toward it.

There is a catch, and it falls squarely on the higher levels of government. None of this community-level governance is possible unless communities have something to govern. The economics of building these models push hard toward concentration. So the most useful thing Washington and the states can do is not to write the rules communities should be writing for themselves, but to keep the ground beneath them open: low barriers to entry, enforcement of existing antitrust laws, provision for models whose weights, standards, and underlying code are open enough to build on, and the ability to move between providers. Legal tweaks in all of these areas may be needed, but some that may seem attractive could backfire: Above all, policies that block market entry, allow only certain models, or discourage alternative approaches should be met with the utmost suspicion. Providing this diversity of models and approaches is the best antidote to the problem that communities could all end up with models that “think” in the same way.

This is where the Grange has the most to teach: National power can protect local life, but it does so best by keeping markets open rather than by deciding for the places it protects. This can mean more federal control in certain cases, but it also means localities might well have more to say than states.

The hollowing-out of community institutions—the local paper, the civic association, the Rotary Clubs and Masonic Temples where a community made decisions—is among the quiet losses of the past half-century. AI arrives as one more threat to them, but also as the first chance in a long while for a community to shape its own information infrastructure.

To keep decisions where people live is to keep the friction that makes a society more than the average of its parts. The Grange understood that a community’s fate should not be settled by people who never live with the consequences. A century and a half later, with a technology that may prove as transformative as the railroads, the task is the same: Give communities the means to shape their own future and flourish on their own terms.

The post AI Needs a Grange appeared first on Washington Monthly.