The Self-Closing Loop

AI, Power, and the End of the Public Good

By Mike Stancil

Artificial intelligence is being actively abandoned as a public good. The vision of empowerment is being quietly killed by corporate restructuring filings and confidential IPO prospectuses, while the public conversation is stuck arguing about whether it’s a force for good or not.

While this tension feels like another flavor of “market freedom” versus “government overreach” or even the boogieman, “socialism,” this shift matters in ways that exceed economic disagreement. It may be among the more consequential decisions our civilization makes without fully realizing it’s making one.

We’ve been here before. It’s not a new pattern. The history of transformative technology is not a history of inevitable privatization or inevitable public control. It’s a history of well-worn sequences. Electricity was built by private capital, consolidated into near-monopolies, and then finally regulated. Not because regulation was the original plan, but because the abuse of private dominance eventually made it unavoidable. The railroad industry followed the same arc. So did telecommunications. 

What we tend to call the "utility model" was almost never the design. It was the correction. And that correction was only possible because the technology scaled slowly enough that governance could eventually catch up to it.

Something is different now. Not in the pattern, but in its speed. The window between the emergence of a transformative technology and its deep entrenchment in economic life is closing faster than the institutions designed to act within it can move. The window in which a different governance arrangement is still possible may be closed before we get a chance to stop it.

This raises several questions that build toward an uncomfortable conclusion:

Does a for-profit service that depends structurally on public utilities get to draw on those shared resources without limit, at whatever scale its capital can sustain, regardless of what that costs the public that built and maintains them? 

Does a for-profit service that becomes necessary for economic participation get to remain governed primarily by shareholder interest while the tool itself operates at a scale capable of reshaping the economy it now inhabits? 

And is there a point at which private market penetration becomes so complete that the utility model of access isn’t merely unachieved, but structurally foreclosed. Closed not by a democratic decision, but by the weight of arrangements too entrenched to revisit?

These questions are being answered as I type this, in ways that may or may not be reversible, by a small number of people who are not held accountable to the public whose infrastructure they are drawing on.

The Infrastructure Problem

The first question many of us feel acutely if we’re anywhere near a potential data center site. For me, this became reality with the sudden closure and sale of Pittsburgh International Raceway (Pitt Race,) a world class autocross and competitive kart racing complex near Pittsburgh. Just as my son was getting into racing, the facility was abruptly sold and is now in the planning stages of becoming a data center.

The 400-acre site sold for $50 million (approximately $130,000 per acre) to an entity linked to Provident Data Centers, a Dallas-based real estate firm. The logic of the acquisition was straightforward. Pitt Race had what data center developers need most: land, an onsite electrical substation, and a water reservoir.

What the community around Pitt Race gets in return is less clear. Wampum is a small, economically depressed borough in Lawrence County, where median household incomes hover just above the poverty line. The promise of data centers in rural communities has always been jobs, but the evidence for that promise is thin. Research has consistently shown that economic benefits of a typical large data center decline substantially after the construction phase, leaving communities with a surge of temporary work followed by a facility that operates largely without them. A Rice University study found that hyperscale data center developers choose to build in rural communities not to create jobs or grow local economies, but because of low costs (cheaper land, cheaper power, fewer zoning complications). The community is the input, not the beneficiary.

The workers who run these facilities, by and large, will not come from Wampum. The skills required are specialized, and the people who have them tend to live in places with more infrastructure, more amenities, and more opportunity than a rural borough in western Pennsylvania can offer. What arrives instead is a structure that draws heavily on shared public resources while returning relatively little to the people who live around it.

That grid burden is not an empty talking point. In the PJM electricity market (which covers western Pennsylvania and stretches from Illinois to North Carolina), data centers accounted for an estimated $9.3 billion price increase in the 2025-26 capacity market alone, with average residential bills expected to rise by $16 a month in Ohio and $18 a month in western Maryland. The people least equipped to absorb those increases are exactly the people living in the communities data centers have chosen for their low costs. Low-income residents already spend up to 20% of their income on energy, compared to 3% for higher-income households, and even before the data center boom, one in four households reported difficulty paying their energy bills or maintaining safe indoor temperatures.

The industry is aware of this tension. In early 2026, executives from major AI companies traveled to the White House to sign a nonbinding "ratepayer protection pledge," agreeing to "build, bring, or buy" the power needed for their facilities. A Harvard utility law expert called the pledge meaningless, noting that utilities and utility regulators (not the president or tech companies) hold the pen on who actually pays for grid expansions.

A nonbinding pledge. From the companies whose infrastructure is drawing on public utilities built and maintained by the people least positioned to absorb the cost.

The Governance Problem

The second question gives me the most concern. We have a new technology advancing and revolutionizing nearly every aspect of our lives. Specifically, it’s incredibly impactful on our economy and participation in it. While those mechanisms are typically regulated by government intervention, the current system is not showing much oversight for the sake of progress. 

There is a version of this question that gets dismissed quickly, framed as an argument about wealth inequality or the familiar critique of unchecked corporate power. That framing undersells it considerably and gets shelved as a partisan distraction. What is being asked here is something more specific and more structurally serious: whether a tool that has become (or is rapidly becoming) a prerequisite for economic participation can remain governed by the same logic we apply to a company that makes better trucks.

The logic here is not neutral. Shareholder governance optimizes for return. That’s not a criticism, it's how the system works. The problem is not that AI companies are run by people trying to make money. The problem is that a tool optimized for return, when that tool is also the infrastructure through which people access economic opportunity, will naturally be built and priced and developed in ways that serve the customers most capable of generating that return.

The people best served by AI will be the people who can pay most for it, work most fluidly with it, and live in the environments where its benefits accumulate. The people in Wampum are not, and will not be, the optimization target.

This produces stratification, not as a side effect but as a structural output. The communities least equipped to absorb the costs of AI infrastructure (higher energy bills, displacement of local institutions, elimination of the kinds of jobs that don't require a computer science degree) are the same communities being chosen as sites for that infrastructure because of their low costs and limited political resistance.

The extraction and the exclusion are not separate phenomena. They are efficiencies.

Here is where the human question becomes analytically important, separate from pure pathos. The people making foundational decisions about how this technology is built, priced, deployed, and governed are people for whom the consequences of those decisions are largely detached.

This is a question about feedback. Democratic systems and functional markets both depend, at some level, on decision-makers experiencing the results of their decisions. A regulator who uses the public transit system has a different relationship to transit policy than one who doesn't. A legislator whose children will compete for the jobs being automated has a different relationship to labor policy than one whose children will not.

The feedback loop that normally disciplines decision-making (however imperfectly) is severed when the people making the decisions are insulated from their consequences by a degree of wealth that isn't just large but categorically different from the wealth of the people affected.

The executives running the most powerful AI companies are not merely rich. They inhabit an economic reality so removed from average American life that the tools they are building, and the tradeoffs they are making in building them, cannot be evaluated by them against any lived experience of what those tradeoffs cost. We’ve seen what that detachment looks like in practice. 

The day after Donald Trump's second inauguration, OpenAI CEO Sam Altman stood in the Oval Office to announce a $500 billion infrastructure deal, telling the president that "for AGI to get built here, we wouldn't be able to do this without you." The deal was presented to the public as a jobs program. The jobs it will primarily create are construction roles situated near data center locations, temporary by nature, in communities chosen for their cheap land and available power. The people making that announcement, and the people cheering it, will not be the people absorbing the consequences when the construction crews leave. That distance was the necessary condition under which the decision was possible.

That isn't a moral failing, necessarily. It's a structural one, and it would be true regardless of the specific individuals involved.

What makes the current moment different from previous concentrations of industrial power is the mechanism available for ensuring that concentration persists. Citizens United didn't create the relationship between capital and legislation. It industrialized it.

The AI companies most capable of shaping the regulatory environment governing AI are also the companies with the resources to fund the political outcomes that keep that environment ideal. This isn't a conspiracy, it's the ordinary operation of the political economy at a scale and speed that existing democratic institutions were not designed to handle. The feedback mechanism that was supposed to exist (voters, through their representatives, setting the conditions under which industries operate) is being outpaced by the capital available to shape what those representatives believe, prioritize, and ultimately decide.

The numbers are not subtle. Top tech and AI companies spent more than $100 million to influence government policy in 2025. OpenAI alone increased its lobbying spend from $260,000 in 2023 to $1.76 million in 2024, and continued accelerating from there. The return on that investment arrived quickly. Big spenders were rewarded with policy shifts, notably support for building data centers and a reversal of the White House's ban on selling advanced AI chips to China. The companies spending the most on shaping AI policy are the same companies whose valuations depend on AI policy remaining unrestrictive. That’s not a coincidence to be noted and moved past. It’s the mechanism by which the governance question gets answered before it can be seriously asked.

The Loop

Each of those stages is serious on its own. Together, they create something qualitatively different: a system with the means and the incentive to ensure its own continuation regardless of whether that continuation serves the public.

The recursive quality of this is what distinguishes it from ordinary monopoly power. Standard monopoly analysis asks whether a dominant firm is effectively excluding competitors. That’s a serious problem, but it is a bounded one. Regulators have broken up monopolies before, however imperfectly and uncomfortably late. What is being described here is something harder to correct, because it operates at the level of the conditions under which correction is evaluated and attempted.

A sufficiently entrenched AI system doesn’t need to prevent regulation through crude interference. It shapes the environment in which regulation is conceived. It optimizes labor markets in ways that concentrate the people capable of building alternatives within a handful of firms, making competition structurally difficult before any regulatory question arises. It funds the political outcomes that determine who sits on the committees writing AI policy and what frameworks they find credible. And it does something more unsettling than either of those: it increasingly mediates how the public understands AI itself.

This last point deserves to stand out on its own for its implications. The tool most capable of shaping information environments at scale is owned by the people who benefit from particular narratives about AI persisting. The question of whether AI should be regulated like a public utility, whether its development serves the public interest, whether its costs are being fairly distributed, these questions are increasingly being encountered by ordinary people through AI-assisted search, AI-generated summaries, AI-curated feeds. Evidence from the 2024-2025 electoral cycle shows how AI systems were already optimizing content for maximum emotional impact across multiple countries, not through conspiracy but through the ordinary operation of engagement-maximizing systems built to serve the priorities of the people who built them. 

The information environment in which the public forms opinions about AI governance is not neutral. It is infrastructure owned by one of the interested parties.

This is where the loop closes in a way that is genuinely difficult to reverse. A monopoly can be broken apart. A corrupted regulator can be replaced. An information environment shaped by the tool being regulated, in ways that make alternatives harder to imagine and harder to argue for, is a different kind of problem that compounds with each passing quarter of market entrenchment.

We have handled governance issues before, with electricity and with railroads and with telecommunications. They were handled imperfectly, after considerable damage, but we managed it. What made those answers possible was that the systems being regulated could not, themselves, shape the terms of the debate about their regulation. The printing press that published arguments for breaking up Standard Oil was not owned by Standard Oil.

The question of whether we can get to a similar answer now isn’t purely a question of political will. It’s a question of whether the window remains open and if we can still see it clearly enough to move toward it.

The Threshold Problem

This begs the final question: is there a point where we can't go back, but where the control loop hasn't fully closed?

Imagine a threshold of AI penetration into every part of the economy at which point it can never be extracted. Not regulated, not restructured, but effectively and meaningfully removed. The threshold isn't a dramatic moment. There is no announcement, no event that marks the crossing. It’s a condition that arrives gradually and is recognized, if at all, only in retrospect.

It’s worth being precise about what kind of threshold we are describing, because there are actually two, and they close at different times.

The first is the economic extraction threshold, the point at which AI is so embedded in the functioning of critical systems that removing it would cause more disruption than governing it badly. This threshold has already been crossed in several sectors, quietly and without much public deliberation. Nearly two-thirds of U.S. hospitals using Epic's electronic health record system had adopted ambient AI documentation tools as of June 2025, with the technology listening to doctor-patient conversations in real time and generating clinical records automatically. 

Cleveland Clinic, Mass General Brigham, and Emory Healthcare are not running pilots anymore, they are counting adoption in thousands of providers and building institutional workflows around systems they did not build and do not control. In consumer finance, AI credit scoring has become the core infrastructure for modern lending. It’s no longer a supplementary tool layered on top of existing processes, but the decisioning layer itself, handling approvals, risk assessments, and loan determinations at a scale and speed that human review can’t match. 

In hiring, AI-generated rankings and scores are now influencing who gets interviewed across industries, with regulators in New York and California beginning to acknowledge that a human nominally reviewing an AI decision isn’t the same thing as a human making one.

The question of who governs these embedded systems, on what terms, and in whose interest is already live in each of these sectors. It’s being answered primarily by the companies that built the tools, because no adequate public framework exists to answer it otherwise. 

The economic extraction threshold doesn't announce itself. It simply means that the conversation shifts from "should we allow this" to "how do we manage what we already have." That shift has already happened in hospital documentation and consumer credit. It’s in progress in hiring, in legal research, in logistics, in education. The trajectory is not difficult to project.

The second threshold is harder to define and more consequential: the political correction threshold. It’s the point at which the concentration of capital, legislative influence, and control over information environments makes meaningful democratic correction structurally impossible. Not politically difficult, but structurally impossible. This threshold has not yet closed. But the previous section of this essay describes the mechanism by which it is closing, and the pace of that closure is not independent of the pace of economic entrenchment. 

Every quarter in which AI becomes more operationally necessary is a quarter in which the argument for leaving its governance undisturbed becomes easier to make and harder to refute.

The gap between these two thresholds is where this essay lives, and where the urgency of the question originates. If both thresholds have already closed, there is nothing left to argue for. If neither has closed, the argument is important but not yet pressing. The specific danger of the present moment is that the first threshold is closing sector by sector while the second remains technically open. Which creates the precise conditions under which inaction feels most defensible.

“There is still time. The system is still correctable. The window hasn't closed yet.”

That framing is the system's greatest advantage. The most dangerous stretch of the slope is the part that still looks like flat ground.

What makes this different from every prior version of the same structural problem is that those systems could not reshape the conditions under which they were evaluated. The railroad companies could lobby Congress and corrupt regulators, and many did. But they could not reach into the daily information environment of every American and shape what they understood about railroad regulation, what alternatives they could imagine, what felt possible and what felt radical. Artificial Intelligence can. It already does, in ways that are neither conspiratorial nor accidental but simply the ordinary output of systems optimized to hold attention and reflect back what users are already inclined to believe.

The extraction threshold and the correction threshold are therefore not independent. As AI becomes more economically necessary, the information environment in which people form opinions about its governance is increasingly mediated by the thing being governed. The window doesn't just close because of lobbying and capital accumulation. It closes because the capacity to imagine it closing or imagining the path to escape is itself subject to the same forces of entrenchment.

It’s an argument for honesty about what kind of moment this is. The question of whether AI should be governed in the public interest is still answerable. The people with the most to lose from a different answer have not yet finished purchasing the conditions under which the question gets asked. 

There is still time. The economic extraction threshold has crossed in some sectors but not all. The political correction threshold remains open.

The window is a condition, not a permanent feature of the landscape. The recursive loop of economic necessity, legislative capture, and information environment control described in the previous section, is not waiting for permission to close it. It’s already closing. The question is not whether we want a different arrangement. It is whether we are willing to say so while saying so still matters.

So What Do We Do?

No one has a complete answer. Anyone who tells you otherwise is either selling something or hasn't read the previous sections carefully enough.

But "no complete answer" isn't the same as "no direction." If the argument made here is correct, the most urgent task isn't finding the perfect policy solution. It's refusing to let the question go unanswered by default, because that's exactly what happens when the window closes. It happens not through a decision, but through the accumulation of moments in which nobody decided anything.

The most glaring gap is one that's received almost no serious attention: we don't have agreed-upon metrics for either threshold. We can observe that ambient AI documentation is running in nearly two-thirds of American hospitals. We can observe that AI credit scoring has become the core infrastructure of modern lending. What we can't do is point to any framework, developed by a public institution or with any democratic origin, that tells us what level of market entrenchment constitutes an economic extraction threshold, who's responsible for monitoring how close we're getting, or what triggers a mandatory public review before the line is crossed. That framework doesn't exist. Its absence isn't an oversight. It's a condition that benefits the people whose interest lies in thresholds remaining invisible and crossed without fanfare.

So what would it look like to make the threshold visible? Financial regulators track systemic risk in banking with designated intervention points that trigger independent review before a crisis becomes unmanageable. There's no equivalent for AI entrenchment in healthcare, in credit, in hiring, in the legal system. Building one, under some form of democratic authority, is probably the precondition for everything else on this list. Without it, every other intervention is reactive — responding to a threshold that's already been crossed rather than acting while there's still room to act.

Governments at every level are among the largest buyers of AI services, and most of them are purchasing without conditions. Every hospital system signing a contract with an ambient AI vendor, every municipal government deploying a hiring algorithm, every public university licensing a generative AI platform is making a procurement decision that could carry requirements for transparency, auditability, interoperability, and meaningful human override. The EU has already developed model contractual clauses for public AI procurement that embed accountability directly into the purchasing relationship. American public institutions could do the same tomorrow without waiting for federal legislation. The question isn't whether the lever exists. It's whether anyone is reaching for it.

In 2025 alone, 1,208 AI-related bills were introduced across all 50 states, with 145 enacted into law. Virginia and Georgia have moved toward data center moratoriums. Sanders and Ocasio-Cortez introduced federal legislation in March 2026 conditioning new data center construction on the enactment of AI regulation. The Trump administration responded with an executive order directing the DOJ to challenge state AI laws it considers burdensome. That response is worth sitting with for a moment. The administration isn't ignoring state-level action. It's actively working to suppress it, which tells you the people with the most to lose from a different answer understand exactly how much fragmented state resistance could matter if it ever became coordinated. The question isn't whether resistance exists. It's whether it can find coherence before it's made irrelevant.

The EU AI Act's transparency provisions require disclosure, auditability, and human oversight for AI systems making consequential decisions about people's access to credit, employment, healthcare, and housing. They establish the principle that a system affecting people's lives can be required to explain itself. That's not a radical idea. It's the minimum condition of accountability in a democratic society. The question is whether enough people understand that its absence here isn't a technical limitation or a gap waiting to be filled. It's a choice that's being actively maintained.

Underneath all of this is a civic question that may matter more than any of the policy ones. The feedback loops this essay has described aren't permanent features of the landscape. They're conditions that persist in the absence of organized, informed, persistent pressure from people who understand what's at stake. The threshold isn't crossed by a single decision. It's approached incrementally, in quarterly earnings calls and procurement meetings and zoning hearings and lobbying disclosures, by people who are mostly not thinking about thresholds at all.

Which means the people who are thinking about thresholds have an opportunity, for as long as that opportunity remains open. The argument is still being made, in state legislatures and academic papers and essays like this one, in rooms that the people with the most to lose from a different answer have not yet fully entered.

The window is still open.

What we do with that is the only question that matters right now.