Between the corporations optimizing us out of the org chart and the governments optimizing us out of the decision, there is a third path: bound the machine, and let institutions be human.

The Machine Must Stop — Image by adflegal.org

The Machine Must Stop — Image by adflegal.org

The Boundary Moves

In Neither Gods Nor Martyrs, I argued that the great debates over medical technology had been captured by two camps having the wrong conversation. On one side stood the transhumanists, who see the human body as legacy hardware awaiting an upgrade. On the other stood the bioconservatives, newly armed with Pope Leo XIV’s May 2026 encyclical Magnifica Humanitas, who warn that a civilization drunk on optimization will quietly decide that some lives are less worth keeping. Against both, I proposed a third position — call it therapeutic humanism, or the healing mean — built on a single load-bearing distinction: the difference between restoration and optimization. Restoration is bounded. It aims at a baseline, at returning a person to the full health proper to a human being, and it carries its own stopping point in its definition. You cannot over-cure malaria. Optimization is unbounded. It aims at a maximum, and a maximum has no natural ceiling; a project with no stopping point will, given enough power, keep going until it has remade everything, including the people it was supposed to serve.

That essay was about bodies. This one is not, because the most consequential application of the healing mean in the twenty-first century is not medical at all.

The optimization logic that the encyclical diagnosed does not live primarily in gene-editing labs or longevity clinics. It lives in the org chart, in the data center, and in the national-security state. It lives in the corporation that dreams of the near-zero-human company, restructuring itself at machine speed to maximize output while treating its own employees as friction. It lives in the individual professional who, offered a machine that thinks faster than he does, quietly stops thinking at all. And it lives in a government that has begun treating advanced software as a weapon — a thing to be gated, vetted, export-controlled, and, when the politics demand it, switched off globally overnight. Each of these is the same move the encyclical warned against, performed on a different patient. The person, the firm, and the state are each being reframed as systems to be maximized rather than institutions to be kept healthy, and in each case the unmaximized element — which is to say, the human element — starts to look like a liability on the balance sheet.

The public conversation about this, like the medical one before it, has collapsed into a false binary. One camp insists that we must accelerate without reservation, that every constraint on algorithmic capital or strategic AI is a tax paid to timidity, and that the humans displaced along the way are the necessary cost of progress. The other camp counsels rejection: refuse the tools, halt the data centers, freeze the models. Gods or martyrs, again. And again, both camps are answering a question that most people are not asking. Almost nobody wants an economy run by autonomous corporate software or a security state with a kill switch over the world’s intelligence infrastructure. Almost nobody wants to smash the machines, either. What people want — when you ask them directly — is that the technology make their work more human rather than less, their towns more prosperous rather than more extracted, and their governments more accountable rather than more opaque. That is the healing mean, applied beyond the body. It has simply lacked a name and a defense in this territory, because here too the megaphones have been held by the people arguing about gods.

This essay is an attempt to supply that defense. It proceeds through three concentric circles — the individual, the corporation, and the state — because the optimization logic operates at all three scales and the failures compound across them. A workforce whose judgment has atrophied cannot check a corporation that has automated its own governance, and a corporate sector organized around pure extraction cannot resist a state that has learned it can seize the switch. At each scale, I will apply the same diagnostic filter developed in the earlier essay, and at each scale I will try to show that the choice we actually face is not between acceleration and refusal, but between bounded restoration and unbounded optimization.

The Filter, Restated

Three criteria organize everything that follows, and each descends directly from the argument of Neither Gods Nor Martyrs and from the encyclical that anchored it.

The first is ontological dignity. Human worth is inherent and existential. It is, in Leo’s phrasing, “neither acquired nor earned,” and it “does not need to be justified” by output, efficiency, or utility. This is not an argument against performance review or compensation tied to output — those are legitimate management tools. The claim is narrower: an employee’s fitness for a role is a separate question from an employee’s worth as a person, and the two must not collapse into one number. A metric can say whether a job is being done well. It cannot say how much basic consideration a person is owed. The moment an institution lets the first quietly answer the second, low scorers stop looking merely reassignable and start looking dispensable — and a system that has erased the distinction will eventually act on it.

The second is finitude as foundation. Human limits — bounded attention, the need for rest, vulnerability, the fact that a person can hold only so much and decide only so fast — are not bugs to be automated away. They are the ground on which meaning, relationship, and institutional trust are built. The encyclical’s most quoted line, that “humanity flourishes not despite limitations, but often through them,” applies to organizations as much as to bodies. An institution that engineers away every human limit does not become a better institution. It becomes something else, and the something else has no particular reason to care about the creatures it was built to serve.

The third is restoration of the self. Technology passes the filter when it returns human beings to full agency — when it clears away the drudgery, the noise, and the curable dysfunction so that judgment, creativity, and responsibility can operate. It fails the filter when it replaces agency, when the human is reduced to approving what the machine has already decided, or removed from the loop entirely. Restoration preserves the agent. Substitution dissolves them. These are not the same act, and the whole argument turns on refusing to let them blur.

With the filter in hand, begin at the smallest scale: the single working mind.

Part One: The Individual — The Value of the Unassisted Mind

The Inversion of Professional Value

For sixty years, the knowledge economy ran on a simple bargain that Peter Drucker first described in 1959 when he coined the term “knowledge worker”: prove your worth by doing. Write the cleanest code, build the most accurate model, draft the most airtight contract, and the hierarchy would reward you. Execution was the proof of expertise, and the entire scaffolding of professional life — the universities, the case competitions, the junior-associate years, the career ladder itself — was built to produce and reward excellent executors.

That bargain is not so much being dissolved as rewritten, and the rewriting is happening faster than most job descriptions can keep up with. The exposure is concentrated in the white-collar heart of the knowledge economy — Goldman Sachs estimated in 2023 that roughly two-thirds of jobs in the United States and Europe carry some degree of automation exposure, concentrated in office administration, law, and other credentialed professions rather than in factories — but exposure to automation is not the same claim as disappearance of the job. What the more recent data actually shows is a shift in what a given role requires: the McKinsey Global Institute’s 2025 estimate of technically automatable work hours nearly doubled its own 2023 projection in the space of two years, and the World Economic Forum expects close to forty percent of the skills employers need to change by 2030. The honest reading of numbers moving that fast is not that the jobs are vanishing on a predictable schedule. It is that nobody — not the economists, not the employers, not the workers — can currently say with confidence what a given profession will require of the people in it five years out, and that uncertainty is itself the condition knowledge workers now have to operate inside.

In The Architect’s Inversion, I described what this does to the hierarchy of professional value. Execution — the procedural fluency that took a decade to build and that once sat at the top of the pyramid — becomes a commodity, produced on demand by what we might call System Zero: an external cognitive circuit that processes data, recognizes patterns, generates options, and scaffolds decisions before conscious human deliberation is even engaged. What becomes scarce, and therefore valuable, is everything upstream of execution: framing the right problem, specifying what a good outcome looks like, evaluating whether the machine’s flawless output actually serves the intended purpose, and owning the consequences when it does not. The craftsperson’s economy gives way to the architect’s economy, and the primary failure mode of professional life shifts from execution error — the bug, the typo, the miscalculation — to goal misalignment: the beautifully executed solution to the wrong problem, failing invisibly and at scale.

Read through the healing mean, the inversion itself is neither good nor bad. It is a change in where human value concentrates. The moral question is what we do to the human mind in response — and here the evidence should alarm anyone.

The Risk to the Instrument

The evidence here needs to be handled carefully, because it has been overstated in both directions online. A widely circulated MIT Media Lab study found that participants drafting essays with ChatGPT showed lower EEG brain engagement than those working unaided, with weaker activation in the prefrontal cortex specifically. It is a small, preliminary, task-specific result — it shows what happened during one kind of writing exercise, not that AI use causes lasting cognitive damage, and it should not be read as more than that. But it lines up with a broader and more carefully hedged pattern in the literature: several studies of knowledge workers have found that higher confidence in an AI’s answers correlates with less effort spent evaluating them, and that heavy, passive reliance can narrow the range of solutions a group produces for the same problem. Researchers call the underlying tendency automation bias — leaning on automated systems at the expense of independent checking — compounded by the ordinary human preference for the path of least cognitive effort.

None of this means using AI erodes the mind; it means how it is used matters. The risk is concentrated in passive, substitute use, and it lands hardest on people early in a career, who risk reaching for architectural judgment without ever having built the grinding, hands-on experience that judgment is usually made from.

Applying the Filter: Assisted, Not Substituted

This is where the restoration-versus-optimization distinction stops being a philosophical nicety and becomes a design specification.

The optimizing move treats the individual mind as a bottleneck. Its metric is throughput: how much human deliberation can be removed from the process, how quickly the machine’s output can be approved and shipped. The research literature calls this substitute use — the human accepts the AI’s output as the conclusion, offloading evaluation entirely. Substitute use is seductive because it is genuinely faster, and its costs are invisible in every quarterly metric. The thinking that was not done this quarter does not appear on any dashboard. It appears years later, as an organization full of people who can operate the machine but can no longer tell when the machine is wrong — flawless execution in service of goals nobody retained the capacity to interrogate.

The restorative move treats the individual mind as the asset the whole system exists to serve. Its metric is the quality and durability of human judgment. The literature calls this assisted use: the professional articulates their own analysis before consulting the machine, treats discrepancies between their reasoning and the model’s as productive friction, and constructs a final judgment they can genuinely defend. The machine is a scaffold for thought, not a replacement for it. And the design details matter measurably: one late-2025 study found that simple interface nudges prompting users to pause and reflect before accepting AI output significantly reduced overreliance on incorrect recommendations. Friction, it turns out, is not the enemy of a healthy cognitive practice. Friction is the resistance training. A workflow that never makes the human think is doing to the mind what a chauffeur does to the legs.

Notice that this is exactly the structure of the medical argument. The bounded, therapeutic use of the technology restores the person to full functioning: it clears away the rote work — the formatting, the boilerplate, the retrieval — that consumed attention without building anything, and returns that attention to the genuinely human work of framing, judging, and deciding. You cannot over-cure drudgery, any more than you can over-cure malaria; once the noise is gone and the mind is free, the therapeutic work is done. The unbounded, optimizing use keeps going past that point. It does not stop at removing the drudgery; it proceeds to remove the deliberation, then the evaluation, then the human — because its goal was never health. Its goal was maximum, and a maximum has no stopping point.

The criterion of restoration of the self gives every professional, and every designer of professional tools, a usable test: after a year of working this way, is the human more capable of independent judgment, or less? If the answer is less, the workflow has crossed from therapy into the thing the encyclical warned about — and no productivity statistic can buy that back.

Part Two: The Corporation — The Autopoietic Bureaucracy and Its Footprint

From Labor to Capital, From Company to Organism

Scale the optimization logic up from the single mind to the firm and it acquires a new and stranger character. A corporation is, formally, a system for coordinating human effort toward a purpose. What happens when the coordination no longer requires the humans?

The economic engine driving this question is straightforward. Human labor is the classic non-accumulable input: you cannot stockpile attention, you cannot compound a workforce the way you compound capital, and every additional employee arrives with salary, sleep, and a sense of purpose attached. Compute is accumulable. Multi-agent AI systems convert what was once a labor cost into a capital asset — and capital, unlike labor, can be scaled by purchase order. For a firm facing diminishing returns, the temptation is not merely to automate tasks but to automate the organization itself: the ideal, openly discussed in parts of the industry, of the near-zero-human company, in which AI orchestrators write the rules, define the roles, route the messages, and continuously restructure the whole apparatus at machine speed. Call it the autopoietic bureaucracy — a bureaucracy that produces and reproduces itself, with the humans reduced to a shrinking compliance layer somewhere near the bottom.

Run the filter over this ideal and it fails on every criterion at once. It fails ontological dignity, because it completes the move the encyclical named as the age’s most insidious temptation: it judges a person’s worth by their output, and then discovers that machines produce more output than people do. On the ledger of the autopoietic firm, an employee is not a person whose flourishing the institution exists to enable; they are a liability line competing against a compute line, and losing. It fails finitude as foundation, because everything that makes a company a human institution — the trust built in slow conversations, the judgment accumulated through years of shared failure and success, the loyalty that survives a bad quarter and rallies on a good one — rests precisely on the foundation of human limitation, on the fact that people must rely on one another because none of them can do everything. An organization restructured hourly by an orchestrator has no such foundation. And it fails restoration of the self for the reason Part One established: a firm that removes human judgment from its own governance is not augmenting its people but dissolving them, and it inherits the architect economy’s signature catastrophe — the flawless, machine-speed execution of a strategy no human retained the ability to question.

The bounded alternative is not a smaller ambition; it is a different one. The restorative firm automates exactly as aggressively — but it automates toward a baseline, not a maximum. The baseline is an organization in which every human role is genuinely human: where the machine drafts and the person decides, where the machine monitors and the person leads, where the workflows are deliberately designed for assisted rather than substitute use, and where the measure of an AI deployment is not how many salaries it eliminated but how much judgment, creativity, and responsibility it returned to the people who remain. That is a company using technology to become more fully what a company is — a community of persons coordinating toward a purpose — rather than to become something else that no longer needs the persons.

But the optimization logic inside a firm does not stay inside the firm. The same impulse that reduces employees to a liability line reduces a host community to a line item too — and it is in that encounter with the physical world that the logic of algorithmic capitalism becomes most visible to the people living next to it.

The Black Box in the Cornfield

The AI economy is being poured in concrete at historic speed, and the way it lands on the ground reveals which logic is running the enterprise. The extractive pattern is by now well documented, and I spent two long essays — on brownfield siting and on economic reciprocity — mapping it. A developer, shielded by non-disclosure agreements, quietly options a few hundred acres of farmland. A rezoning application appears with little warning. A windowless monolith rises, drawing gigawatt-scale power from a regional grid that was never designed for it and water from an aquifer the neighbors drink from, while returning to the host community a modest property-tax line, a chain-link fence, and — after the construction crews leave — remarkably few jobs. The most automated hyperscale campuses can operate with as few as twenty to thirty permanent staff per hundred megawatts of capacity; one 1.1-million-square-foot facility outside Reno disclosed just seventy-three permanent positions. A twentieth-century mill of comparable footprint would have anchored a thousand families for two generations.

Communities have noticed, and the arithmetic of their response is the most underappreciated business story in America. The American Farmland Trust reports the loss of roughly two thousand acres of farmland every day to non-agricultural development, and data centers are an accelerating driver of it. Organized opposition — 142 activist groups across twenty-seven states as of early 2025, rising to at least 188 groups in forty states within months — has blocked or delayed tens of billions of dollars in projects, with one industry tracker counting ninety-eight billion dollars stalled in a single quarter of 2025. Project cancellations rose from two in all of 2023 to twenty-five in 2025. Loudoun County, Virginia — the most permissive data-center market on earth — ended by-right development in March 2025. A court voided the rezoning behind the Prince William Digital Gateway, once planned as the world’s largest campus, and one of its two major developers walked away. By April 2026, polling found that voter comfort with a new data center in their own community had collapsed from sixty-nine percent to thirty-five percent in three years, and seventy-three percent of Virginia voters blamed data centers for rising electricity bills. More than a hundred local moratoriums are now in force, twelve states have introduced related legislation, and a federal moratorium bill — however unlikely to pass — sits in the Senate with prominent sponsors. This is what the collapse of a social license looks like in real time.

The extractive developers tend to read this as a public-relations problem. It is not. It is the same diagnostic signal the filter predicts: an unbounded optimization — maximum compute, maximum speed to market, minimum obligation to the host — treating a human community the way the autopoietic firm treats its employees, as friction to be routed around. And communities, unlike employees, can vote.

The Floor: Brownfields and Reciprocity

The bounded alternative already exists, is already operating, and is — this is the point the industry keeps missing — more competitive, not less. It has two components, which I have elsewhere called the third and fourth pillars of responsible digital infrastructure: build on the scars, and pay the place back.

Build on the scars. The United States holds an inventory of more than 450,000 brownfield sites — the former steel mills, aluminum smelters, textile plants, and coal stations that deindustrialization left behind, concentrated precisely in the communities that lost the most when the plants closed. These sites carry, beneath their contamination, exactly what a hyperscale campus needs: heavy industrial zoning, water rights, rail and fiber corridors, and above all pre-existing high-capacity grid interconnection, the asset that greenfield developers now wait five to eight years in utility queues to obtain. Google proved the template in 2015 at the retired Widows Creek coal plant in Alabama, inheriting transmission and water infrastructure and leaving Jackson County with more jobs, at higher wages, than the coal plant had provided at its close. Homer City, Pennsylvania — for decades the state’s largest coal plant, decommissioned in 2023 — is being reborn as a 4.5-gigawatt energy and compute campus on its 3,200-acre site, leveraging dual grid interconnections that a greenfield project could not replicate in a decade, in a county whose economy the plant’s closure had hollowed out. Meta’s Forest City campus in North Carolina, built on a remediated textile and boat-manufacturing brownfield, is now the single largest tax revenue source in Rutherford County, supplying roughly fourteen percent of county revenue on land that had sat contaminated, vacant, and untaxable. Every one of these projects converted an industrial wound into an economic anchor, and every one of them moved faster and met less resistance than its greenfield competitors — because it asked the community to heal something rather than to sacrifice something.

Pay the place back. The second component converts the data center from a consumer of a region into a participant in it, and its instrument is the legally binding Community Benefits Agreement — not the press release, not the pledge, but the contract. The city of Lancaster, Pennsylvania, negotiated the current American benchmark: in exchange for hosting up to ten billion dollars of development, the city secured a hard, covenanted cap of twenty thousand gallons of municipal water per day per campus, twenty million dollars in phased payments into named community funds with local oversight, a one-hundred-percent clean energy commitment with a contractual floor of sixty percent and a five-million-dollar contingent fund behind it, and the retained right to seek injunctions — up to complete shutdown of operations — if the non-financial commitments are breached. Beyond contracts, the physics of the facility itself can be turned outward. In the Helsinki region, Microsoft’s new data centers will feed their waste heat into Fortum’s district heating network, eventually covering roughly forty percent of the annual heating demand of three municipalities serving a quarter-million people and displacing some four hundred thousand tonnes of carbon dioxide per year; in Toronto, Telehouse now routes its server heat into Enwave’s district system, where it helps warm the city’s drinking water. And the redundant fiber that every campus must pull into a region can be given a local off-ramp: Meta’s Forest City fiber build was extended, in partnership with a state nonprofit, into roughly one hundred miles of new backbone serving nearly thirty schools, colleges, and community institutions across five rural counties, plus free public Wi-Fi for thousands of families who previously had none.

Hold the two corporate modes side by side and the encyclical’s categories map onto them precisely. The extractive model is the ceiling project: greenfield land cleared for speed, resources drawn down and nothing returned, obligations minimized behind NDAs — the corporate Babel, climbing on the strength of its own capital and trampling the neighbors on the way up. The reciprocal model is the floor project: industrial scars remediated, waste heat returned as warmth, private fiber opened as public lifeline, obligations written into enforceable covenants — the rebuilt city, where the rejected sites become the cornerstone. One of these models is currently losing sixty-four billion dollars to community opposition and manufacturing the regulatory backlash that will constrain the entire industry. The other is being cited by federal agencies as the template and welcomed by the very counties that are suing everyone else. The healing mean, it turns out, is not just the ethical strategy. It is the durable one.

Part Three: The State — The Weapon That Thinks

The Reclassification

The third circle is the widest, and the events inside it are so recent that most of the public has not yet absorbed what they mean. Over roughly eighteen months, the United States government has quietly reclassified frontier artificial intelligence. It is no longer treated purely as a commercial software product, governed by product liability and market competition. It is treated as a dual-use strategic asset — something closer to fissile material than to an app — governed by the instincts the state reserves for weapons: non-proliferation, export control, customer vetting, pre-release review, and secrecy.

The documentary record is public, if rarely read. In November 2025 the White House launched the Genesis Mission, a Department of Energy program mobilizing the seventeen national laboratories and their roughly forty thousand scientists to pool the government’s classified scientific datasets and train foundation models behind a vetted-access wall, framed by the administration itself as comparable in urgency to the Manhattan Project. In April 2026 the Office of Science and Technology Policy formally designated industrial-scale foreign “distillation” of American models — the harvesting of a frontier model’s capabilities through millions of automated queries — as a national-security threat, and a House bill, the Deterring American AI Model Theft Act, moved to answer it with sanctions and entity listings. In June 2026 an executive order established a framework under which frontier developers give the government pre-release access to their most capable models for up to thirty days, with the National Security Agency deciding, through a classified benchmarking process, which models qualify. The framework is formally voluntary. Within a month, the two leading American labs had both shipped their flagship releases through it — one after having a model forcibly taken offline, the other by opening its launch to roughly twenty government-disclosed organizations while the general public waited.

The fear organizing all of this is real, whatever one makes of its framing. Anthropic has alleged, in submissions to the Senate, that Chinese labs ran coordinated extraction campaigns against its Claude models totaling tens of millions of exchanges through tens of thousands of fraudulent accounts — allegations the named companies deny and that have not been independently verified, but that Washington has treated as confirmation of the underlying strategic logic: if a model’s outputs can be mined for its cognitive structure, then releasing the model broadly is itself a proliferation event. Once that sentence is accepted, everything else follows. The gate at the top of the release escalator. The classified benchmarks. The vetted-partner tiers. The kill switch.

Three Days in June

The kill switch is not a metaphor. On June 9, 2026, Anthropic released Claude Fable 5, the public version of its most capable model tier, wrapped in safety classifiers designed to refuse the riskiest requests. On June 12, a directive from the Commerce Department’s Bureau of Industry and Security — citing a jailbreak demonstration — barred foreign nationals from accessing it, including Anthropic’s own foreign-born engineers. Because a cloud model cannot sort its users by passport in real time, the company’s only path to compliance was to shut the model down for everyone on Earth, the same day. A model that had existed in public for roughly seventy-two hours vanished globally. Allied institutions lost access along with everyone else, including the United Kingdom’s AI Security Institute — the body whose job was to safety-test the very systems that had just been switched off. An American startup dependent on the model sued the government, describing the harm as existential. Anthropic’s own assessment of the triggering evidence was openly skeptical: the jailbreak was narrow and non-universal, amounted to asking the model to read a codebase and identify flaws, surfaced vulnerabilities that were minor and already known, and demonstrated capabilities the company’s subsequent testing found in rival and even older models. Access was restored on July 1, after a retrained classifier and a separate government review — nineteen days of global darkness traced back to a technique that ordinary public models could largely match.

The episode cannot be understood apart from what preceded it. Earlier in 2026, Anthropic and the Pentagon — recently rebranded the Department of War — had fallen out over a contract. The government wanted the model available for any lawful purpose. The company refused to remove two guardrails: no fully autonomous weapons, meaning no AI making final kill decisions without a human, and no mass domestic surveillance of Americans. The response was to designate the company a supply-chain risk — a label historically reserved for foreign adversaries — until a federal judge blocked the designation in March, observing that branding an American company a potential saboteur for disagreeing with its government had an “Orwellian” quality. The June export order landed on that already-scorched ground, three days after the model shipped and weeks before the company’s planned public offering, and a Pentagon official greeted the shutdown by posting that some things matter more than “revenue cycles, clickbait, and pre-IPO valuation.”

Set aside, for a moment, every question about motive. Attend only to what was demonstrated, because demonstrations do not expire. The world now knows, as a matter of public record, that the people who control the model weights can be compelled to flip a switch — globally, overnight, on the strength of a domestic political dispute. Every government, every enterprise, and every allied research institute watched it happen, and they are building accordingly. Within weeks, the European Commission had published its Cloud and AI Development Act, a sovereignty framework that would condition public-sector contracts on graduated guarantees against exactly this kind of foreign control, with its strictest tiers barring third-country corporate control outright. The containment strategy, aimed at preserving an American monopoly, is manufacturing the multipolar, fragmented world it was meant to prevent.

Applying the Filter: Babel with a Clearance

Run the diagnostic over the security state’s posture and the pattern from Parts One and Two reappears, at the largest possible scale.

The state, like the firm, is optimizing — but its maximum is not capital. It is absolute security, absolute advantage, absolute control over a technology it has decided is existential. And like every unbounded optimization, it has no natural stopping point. If a narrow jailbreak justifies a global shutdown, then by the same standard nearly every frontier model on the market must come down; the logic does not restrain itself, it must be restrained. If pre-release review is prudent for thirty days, it is more prudent for ninety. If one adversary’s distillation justifies gating a model from foreign nationals, the next disclosure justifies gating it from everyone. The process now forming — opaque, ad hoc, capability-by-capability approval with no published criteria and no statutory basis — is precisely what an optimization looks like before it has been given a boundary: a perpetual, discretionary preclearance regime, a standing kill switch held over the world’s software, answerable in practice to no one.

The filter’s criteria name what this costs. Ontological dignity, at the civic scale, is the principle that the citizen is not raw material for the state’s strategic calculations — that a researcher in Stockholm, a startup’s engineers in Canada, and a safety institute in London do not exist at the pleasure of another country’s threat assessment. The June shutdown treated all of them as acceptable rounding errors in a security equation, which is exactly how an optimization treats everything that is not its maximum. Finitude as foundation, at this scale, is the constitutional insight itself: that state power must be bounded — by published rules, by courts, by the separation of the switch from the hand that benefits from flipping it — because a power without limits does not become a better protector. It becomes the thing you need protection from. And restoration of the self, at this scale, is the preservation of human agency at the points where it matters most, which is precisely where Anthropic drew its two red lines. A machine that makes the final kill decision has not augmented military judgment; it has replaced the one form of judgment that a society can least afford to automate. A surveillance system aimed at a government’s own citizens has not restored anyone’s security; it has dissolved the boundary that made the citizens a public rather than a monitored population.

None of this argues for a defenseless state, any more than Part One argued for abandoning the tools or Part Two for halting the data centers. The healing mean is not pacifism about technology; it is the insistence on a stopping point. And the stopping point here is legible in the government’s own better instincts. The same June executive order that created the review gate also directed something genuinely restorative: the hardening of federal cyber defenses, a clearinghouse to coordinate vulnerability scanning and patching with industry, the use of AI to find and fix flaws in the infrastructure everyone depends on. That is defense as therapy — bounded, aimed at a baseline called resilience, complete when the systems are patched and the defenders are stronger. It honors human agency, operates under law, and has a definition of done. The line the framework mandates runs exactly there: a security state that uses AI to restore and defend the shared infrastructure of ordinary life stands on one mountain; a security state that reaches for autonomous violence, population-scale surveillance, and a discretionary global kill switch is climbing the other one — Babel with a security clearance, reaching for a heaven of total control and trampling the allies, the researchers, and eventually the citizens in the climbing.

The deepest irony of the AGI Manhattan Project, as I argued when I first examined it, is that it is failing on its own terms. The labs are suing the state that claims to be protecting them. The allies are legislating their way out of dependence on it. The adversaries are accelerating their own programs in response to it. Optimization without limit does not even deliver the maximum it promised. It never does. That was the encyclical’s oldest point, and it is as true of towers built from export controls as of towers built from brick.

The Two Lines Worth Holding

It is worth dwelling briefly on why Anthropic’s two red lines — no autonomous kill decisions, no mass domestic surveillance — are not a narrow contract dispute but a preview of the two ways state optimization does the most damage once it is unbounded.

Autonomous weapons systems fail the filter differently depending on where the human sits, and the difference matters. A system with a human genuinely in the loop, reviewing and authorizing each lethal decision, keeps the act of killing inside human moral responsibility; it can still be misused, but the responsibility has somewhere to land. A system without a human in the loop — one where the targeting, the decision, and the trigger are all machine functions — removes that landing place entirely. No person decided; a process ran. Militaries adopting such systems face an incentive structure that only tightens once an adversary fields one: decision speed becomes a battlefield advantage, which pressures every side to shorten or remove the human review it retains, which compounds into a world where lethal force is authorized at machine speed by processes no one is individually accountable for. That is the unbounded optimization in its purest form — the maximum being pursued is response time, and response time has no ceiling short of removing the human altogether.

Mass surveillance runs the same logic on a population rather than a battlefield. A state that acquires the capacity to monitor its own citizens at scale faces no natural stopping point either, because the case for expanding it is always available: one more data source closes one more gap, one more inference model catches one more threat. What such systems remove is not a single decision but the ordinary condition of civic life — the assumption that one is not, by default, a monitored subject. A citizenry that must assume constant observation does not merely lose privacy; it loses the practical space in which dissent, unpopular belief, and simple error are survivable, which is the same space that makes self-government possible in the first place. And the capability, once built for one purpose, tends not to stay confined to it; a surveillance architecture assembled to catch foreign threats is the same architecture that can be turned, with a change of political weather, on domestic dissenters.

Both failures share the structure this essay has traced at every scale: a system pursuing a maximum — total security, total information — with no baseline at which it is satisfied and stops. That is exactly why the two guardrails are not incidental terms in a technology contract. They are the boundary between a security state that defends its people and a security state that has begun optimizing its people out of the loop entirely.

The Interconnected Threat

Step back from the three circles and notice that they are not three problems. They are one problem, compounding across scales, and the compounding is the real danger.

The dystopia worth fearing is not the science-fiction one. No robot rebellion is required, no rogue superintelligence, no dramatic morning when the machines take over. The system fails quietly, through the ordinary operation of three unbounded optimizations feeding each other. It begins with individuals whose judgment has been allowed to atrophy — professionals who can operate the machine but can no longer evaluate it, a generation that arrived at the architect’s chair without ever building the architect’s foundations. Those individuals staff corporations that have automated their own governance, firms restructuring at machine speed toward metrics no human fully interrogates, extracting from their host communities until the communities revolt. And those corporations operate under a state that has learned it holds the switch — a state whose answer to every uncertainty is another gate, another classified benchmark, another act of discretion without published rules. Each layer disables the check that the other layers needed. Atrophied minds cannot govern autopoietic firms. Extractive firms cannot summon the public trust to resist an overreaching state. And an opaque state, holding a kill switch it may use on the strength of a political grudge, teaches every mind and every firm beneath it that agency is conditional and judgment is someone else’s department.

Follow the trajectory to its end and the two terminal failures come into view. Left unfiltered, corporate optimization yields a stratified order in which advantage compounds for those who own the systems — a technical aristocracy as rigid as the biological one the medical essay warned against, inherited not through germline edits but through capital, compute, and access. Left unfiltered, state optimization yields the permanent enclave: intelligence as a gated national asset, innovation proceeding at the pleasure of an unaccountable review, and a fragmenting world in which every bloc builds its own walled tower and calls it sovereignty. The two failures are not rivals. They are partners. A corporate sector that treats people as friction will happily accept a state that treats oversight as treason, so long as the contracts flow — and a state that wants total alignment will find its most useful allies among firms that have already dissolved their own human governance. The Babel of the twenty-first century is a joint venture.

The Mandate

Against that trajectory, the healing mean issues a mandate, and it can be stated plainly at each scale because the earlier essay already built its foundation.

For the individual, the mandate is the preservation of the thinking self. Use the machine, and use it ambitiously — but design the use so that the human is more capable of independent judgment after a year of it, not less. Assisted, never substituted. The mind’s effortful engagement is not inefficiency to be optimized away; it is the asset the entire system exists to serve, and the only check the higher scales will ever have.

For the corporation, the mandate is the restoration of the institution and the place. Automate the drudgery and keep the judgment human, at every level including governance. Build on the industrial scars rather than the farmland; return the waste heat, the fiber, and the tax base to the communities that host the machines; write the obligations into enforceable covenants rather than press releases. A company’s purpose is not capital’s self-generation. It is a human community coordinating toward a purpose — and the technology passes the filter only when it makes the community more capable of that coordination, not when it makes the community unnecessary.

For the state, the mandate is bounded defense under law. Protect the infrastructure, harden the grids, hunt the vulnerabilities, and share the intelligence — the therapeutic work of security, which has a definition of done. But refuse the unbounded temptations by name: no machine making the final decision to kill, no population-scale surveillance of a state’s own citizens, and no perpetual discretionary switch over the world’s intelligence infrastructure without published criteria, statutory basis, and courts empowered to say no. If the review gates are to exist, codify them; a rule that binds everyone, including the ruler, is the difference between a border and a whim.

And across all three, the mandate is the one that closed the medical essay, translated into new territory: measure the system by the floor, not the ceiling. The transhumanists asked how high the enhanced could climb; the corporate optimizers ask how much the automated firm can extract; the security state asks how absolute its advantage can become. The healing mean asks a different question at every scale, and it is always the same question: what can the ordinary person count on? Can they count on work that engages their judgment rather than harvesting their clicks? Can their town count on the industry next door to give back warmth, bandwidth, and an honest contract? Can their country count on a government whose emergency powers end somewhere, visibly, by law?

Our technical architecture becomes magnificent — to borrow the encyclical’s word one last time — not when it erects autonomous corporate monoliths or state-controlled enclaves that reach for their respective heavens. It becomes magnificent when its every layer is bounded to secure the baseline: the dignity, transparency, and agency of the ordinary citizen, who was never raw material for anyone’s maximum. The institutions we are building with this technology will be the walls the next century lives inside. The only question that matters is the old one: whether we are building the tower, or rebuilding the city — and in the city, every stone that the optimizers rejected has a place, and the wall exists so that everyone, the quick and the slow, the enhanced and the ordinary, is protected by it and belongs inside it.

Human institutions must remain fundamentally human. That is not nostalgia. It is the design specification.

Sources and Further Reading

This essay builds on five earlier pieces in this series, where fuller citations for the studies, polls, filings, and public documents referenced here can be found:

  • “Neither Gods Nor Martyrs” (May 2026)
  • “The Architect’s Inversion: How AI is Redefining the Value of Human Capital” (May 2026)
  • “The Third Pillar of Responsible AI Datacenters: Why the Future of AI Belongs on Brownfield Sites” (May 2026)
  • “The Fourth Pillar of Responsible Data Centers: Why Economic Reciprocity Is the Ultimate Social Contract” (May 2026)
  • “The AGI Manhattan Project: Treating Frontier AI Like a Weapon” (June 2026, updated July 2026)

Key primary and secondary sources cited directly in this piece:

  • Pope Leo XIV, encyclical Magnifica Humanitas: On Safeguarding the Human Person in the Time of Artificial Intelligence (May 15, 2026) — vatican.va
  • Goldman Sachs Global Investment Research, “The Potentially Large Effects of Artificial Intelligence on Economic Growth” (March 26, 2023) — gspublishing.com
  • McKinsey Global Institute, “Agents, Robots, and Us: Skill Partnerships in the Age of AI” (November 25, 2025) — mckinsey.com
  • World Economic Forum, Future of Jobs Report 2025 (January 2025) — weforum.org
  • Kosmyna, Hauptmann, Yuan, Situ, Liao, Beresnitzky, Braunstein, and Maes, “Your Brain on ChatGPT: Accumulation of Cognitive Debt when Using an AI Assistant for Essay Writing Task,” MIT Media Lab (June 2025) — media.mit.edu
  • U.S. Environmental Protection Agency, “Guidance on the Redevelopment of Superfund and Brownfield Sites as AI Data Centers” (January 2026) — epa.gov
  • City of Lancaster, Pennsylvania, Community Benefits Agreement Summary (November 2025) — cityoflancasterpa.gov
  • Brookings Institution, Nicol Turner Lee and Darrell West, “Why Community Benefit Agreements Are Necessary for Data Centers” (January 2026) — brookings.edu
  • Brookings Institution, Dany Bahar and Greg Wright, “New Evidence on Data Center Employment Effects” (May 4, 2026) — brookings.edu
  • The White House, “Launching the Genesis Mission” (November 24, 2025) — whitehouse.gov
  • The White House, “Promoting Advanced Artificial Intelligence Innovation and Security” (June 2, 2026) — whitehouse.gov
  • Anthropic, “Claude Fable 5 and Mythos 5 redeployed” (July 1, 2026) — anthropic.com

A note on sourcing and certainty: the distillation figures attributed to foreign AI labs in Part Three are allegations made by Anthropic in public statements and in correspondence to the U.S. Senate. They have not been independently verified, and the named companies dispute them. Readers are encouraged to consult the primary sources above directly.