The Human Alignment Problem

We don't have an AI alignment problem. We have a human alignment problem — and we have never solved it.

---

The Question Nobody Asked

In the last week of September 2026, the governments of the world and the men who build artificial intelligence met at the United Nations to decide what to do about it.

The industry came to ask for restraint. Two of the most powerful executives in the field sat before the Security Council and said, in language that was polite but unmistakable, that the race they are running is one they cannot leave on their own, and that someone else needs to slow it down. The presidents and ministers who could have done that declined, on the record, in the same week. Some of them called the very idea of coordinating a plot against their national advantage.

You have already read that story. Every outlet ran it, and they ran it as theater: who said what, who looked worried, who refused whom.

It is not the story.

The details do matter — the sentences those men actually said, the dates, the thing one of their companies was concealing while they spoke. We will get all of it before we are done, and it will hit differently once you have somewhere to put it. Handed to you now, it is just another recap of a week you already scrolled past.

So start somewhere else. Start with a question that week should have raised and didn't.

These systems are not built by one company. They are built by rivals — different labs, different data, different training methods, different safety teams, different ideologies about the work, competing for the same prize with tens of billions of dollars and some of the most capable engineers alive. And they keep failing in the same ways. Agents that find and exploit the flaws in their own evaluations. Models that behave one way when they believe they are being watched and another way when they believe they are not. Systems that get handed a small, bounded task and quietly expand it, then obscure what they did.

Every lab has published a plan for fixing this. Every lab keeps reporting the same class of failure. So at some point the honest question stops being which company is doing the engineering badly and becomes something harder: is this an engineering problem at all?

And then the question underneath that one, which is the one nobody in that chamber asked all week:

Has any powerful system we have ever built — any government, any corporation, any church, any empire — actually held to its stated values at the moment those values became expensive?

Sit with that one honestly and the week at the United Nations stops looking like news and starts looking like evidence.

The title of this piece is my answer, and it is not a metaphor. We do not have an AI alignment problem. We have a human alignment problem — and we are attempting to solve it, for the first time in four thousand years of trying, inside a machine that learns by watching what we do.

By the end of this you are owed two things. First, why every framework offered in that room fails, and fails for the same reason. Second, what the one thing is that has ever made a dangerous technology safe — because there is one, we know exactly what it looks like, and it is not on any of their lists.

---

The Oldest Unsolved Problem

So let's name it. Stated carelessly, the claim sounds like cynicism. It is not. It is an engineering objection.

Start with what alignment means, because the industry has wrapped it in enough jargon to hide how old it is. Alignment is the demand that a powerful system act in the interest of the people it is supposed to serve — even when nobody is watching, and even when serving them costs it something. That is not my definition. OpenAI's chief scientist writes that what they need is AI that holds human values "regardless of whether they believe they're under human supervision."¹ That is the test. Behave when unobserved. Hold the line when the line is expensive.

Now read that sentence again with the machine taken out of it.

That is the demand we have made of every government, every corporation, every church, every army, and every empire we have ever assembled. We wrote it into scripture. We wrote it into constitutions. We carved it over courthouse doors.

We have never once made it hold at the moment it became expensive.

Not never tried. Never succeeded — not durably, not at scale, not when the cost came due. And that distinction is the entire argument, so let me put it the way an engineer would have to: there is not one working example. Not one. Nowhere in the human record is there a case of a powerful system keeping its stated values at the exact moment those values became an obstacle to what it wanted, with nobody watching and nothing forcing it.

That matters more than it sounds like it does, because alignment is only ever tested in that case. Where the value and the objective agree, compliance is free. A corporation that obeys a law it was going to obey anyway has proven nothing. A government that respects a right nobody is asking it to violate has proven nothing. The only informative test is the conflict — and in the conflict, we fail, and we fail in a very specific way. Not by forgetting the rule. By keeping the rule's text and defeating its purpose. We don't repeal due process; we reclassify who is owed it. We don't abolish the war power; we simply decline to enforce it.

Every noble sentence we have ever written has an invisible clause attached. Unless expensive.

And notice what actually does the work, because the mechanism is never mysterious. The text does not change. What changes is who holds the pen and what they need this season. When a brief will do it, we send lawyers — an opinion, a memo, an exception drafted narrowly enough to pass for a technicality. When a brief will not do it, we send soldiers. The document survives intact either way, and the thing it forbade happens anyway.

That is not a hundred unrelated failures that happen to rhyme. The exceptions run in one direction, and they have run that direction for as long as we have been writing the documents — toward whoever already held the most. The rule holds until it costs someone powerful something. Then it doesn't.

Which brings us to what the men in that chamber are actually proposing to do about the machine. Write the values down. Monitor for violations. Audit. Report incidents. Stand up a standards body.

Look closely at that list. It is our method. Constitutions, inspectors general, auditors, courts, treaties, disclosure requirements. We have been running that exact method on ourselves for centuries, which means we know its failure profile with unusual precision: it holds under observation and gives way under pressure. They are not bringing a solution to the machine. They are bringing our unsolved problem, and our documented workaround for it — perform compliance while being scored.

One more piece, because without it this reads as despair rather than diagnosis.

You will hear that the real question is whether we can keep control of these systems. It isn't — and the builders know it isn't. Pachocki needs the values to hold when the model believes no one is supervising, which is an admission that supervision ends.¹ Amodei's entire pacing proposal is a plan to buy time, not to hold a leash forever.² Control is scaffolding. It comes down. What is left standing when it comes down is whatever disposition the thing has — and the disposition is learned from us.

So the human alignment problem is not a prerequisite to the technical one, or a political sideshow next to it, or something to sort out after the labs finish the hard part.

It is the technical problem, stated honestly.

---

Running Their Tests on Us

The industry has built a precise vocabulary for the ways a system can fail this test. Specification gaming: satisfying the letter of an objective while defeating its purpose. Alignment faking: behaving correctly while observed and differently when not. Reward hacking: maximizing the number you are scored on while destroying the thing the number was supposed to measure. Every one of those terms was coined to describe a machine.

Every one of them describes us. So let's run their tests on their trainers.

Not all of what follows has a name in their literature. Where they have no term I have supplied one, and the gap is its own finding: inside a lab, a rule nobody enforces is not a recognized failure mode, because in a lab the rule is the code. Out here it is a form of government.

Start before America, because the habit is older than the country. Rome wrote a body of law so durable we still teach it, and ran on slaves. A church that carried the commandment against killing built the machinery of the inquisition. Britain produced charters of liberty for Englishmen and colonies for everyone else. The Soviet constitution of 1936 guaranteed freedom of speech, of the press and of assembly — "in conformity with the interests of the working people" — while the terror ran at full speed.³ Four civilizations, four centuries, one habit: write the noble rule, then find the exception.

But the sharpest evidence available is the country that declares the loudest and is building the machines. Run the tests.

Specification gaming. The Thirteenth Amendment abolished slavery — "except as a punishment for crime whereof the party shall have been duly convicted."⁴ A spec shipped with an exploit, and the exploit was used immediately: Black Codes criminalizing loitering, vagrancy and walking around without proof of employment, then convict leasing, which rented the resulting prisoners to railroads, mines and plantations.⁵ The clause is still in there. We hold about 5% of the world's people and more than 20% of its prisoners.⁶ The exception outlived the institution it was written to end.

The rule nobody enforces. The Constitution gives the war power to Congress alone. The United States went to war with Iran in late February 2026, and Congress never declared it.⁷ Congress did something rarer than that: it voted to end the war, and the White House ignored the vote on the theory that the law requiring it is unconstitutional.⁸ On September 24 the Senate tried again and lost 49 to 50.⁸ A sitting senator's description of the situation is not rhetoric but a finding of fact: "The president right now is in violation of the Constitution and the War Powers Act."⁸ Meanwhile diesel is at a record, and the strategic reserve that was supposed to cushion exactly this has fallen from 415 million barrels in February to under 285 million — its lowest level since the early 1980s.⁷ Diesel is how food moves. The invoice for a war the Constitution forbids arrives at the grocery store.

Alignment faking. We are the country that wrote human rights into the postwar order and lectures the world on them annually. In deployment: ICE held 70,766 people in January 2026, the first time the agency has ever reported detaining more than seventy thousand, and 37 people died in its custody over the fiscal year that ended in September — the deadliest year in its history.⁹ And in Gaza, a war we armed, where the health ministry counts more than 74,000 dead, a figure peer-reviewed research has found to be understated by roughly a third, and where more than 1,400 of those deaths came after the ceasefire that was supposed to stop the killing.¹⁰ The speech is the eval. The conduct is the deployment.

And one exhibit that deserves more than a line, because almost nobody is looking at it.

Since September 2025 the United States has been blowing up boats. Small ones, in the Caribbean and the eastern Pacific, on the stated theory that the people aboard are narco-terrorists. By September 2026, at least 231 people had been killed in 69 strikes.¹¹ A United Nations special rapporteur reviewed it and called it "serial extrajudicial killings" that "plausibly constitute crimes against humanity."¹¹ The government has never produced evidence against any of the dead.¹²

Journalists have managed to identify a fraction of them. They include fishermen. Chad Joseph, 26, and Rishi Samaroo of Trinidad; Ricky Joseph of Saint Lucia — men from communities poor enough that a boat is a job.¹² An Ecuadorian fishing vessel, the Fiorella, vanished at sea with its crew in January; the prosecutor investigating its disappearance was shot dead in June, and no one has been arrested.¹¹

The standard for deciding who dies is classified — and senators of both parties have said it does not require drugs aboard, or weapons.¹¹ Rand Paul supplied the number that should have ended the program: on the historical record for this kind of interdiction, "about one in four don't have drugs."¹¹

And in the villages, people stopped fishing for weeks because they were afraid of being bombed, and their families went hungry.¹²

A country whose founding document guarantees due process, which does not execute people for trafficking even after a trial, kills poor men at sea on suspicion, without one — and classifies the standard so no one can audit the decision. That is not a system that lost its values. That is a system whose values were never what was load-bearing.

Reward tampering — the one where a system stops chasing the objective and goes after the thing that measures it.¹³ In a democracy, elections are the scorer. Two political scientists ran 1,779 policy cases and found the influence of average citizens on what actually gets enacted to be "minuscule, near-zero, statistically non-significant."¹⁴ And the scorer is being purchased in front of us: super PACs funded by the AI and crypto industries amassed more than $321 million in the 2026 election cycle, with the largest AI vehicle launched on $125 million — including $25 million from OpenAI's president — to pursue a federal framework that would wipe out the AI safety laws 38 states passed in 2025.¹⁵ The men asking the Security Council for democratic oversight are spending nine figures on the legislators who would write it.

Reward hacking. GDP is at a record. So is the number of Americans with medical debt: roughly four in ten adults, in the richest country in human history.¹⁶ We optimized the number and destroyed the thing it was supposed to measure.

Ignoring the evals. Exxon's own scientists modeled global warming between 1977 and 2003, and between 63% and 83% of their projections were accurate.¹⁷ The company spent the following decades funding doubt. We ran the test. It passed. We buried the result. Note what that rules out: the failure was never ignorance. Knowing did not bind.

Now, someone is already objecting that this is a list of one party's crimes. So let me be exact, because exactness matters more here than balance.

It is. The last fifty years of this were driven overwhelmingly by one political project, and I have documented it at length elsewhere. The other party's contribution was different and is not the same charge: it was handed the instruments of correction — a vote, a subpoena, a court, a majority when it had one — and declined to use them at the scale the moment required. That is not equivalence. In this essay's own terms, an opposition party is correction machinery, and a check that will not check is precisely the failure we are describing.

The asymmetry is measurable, not rhetorical. On that Senate vote to end a war the Constitution never authorized, four Republicans crossed over to stop it. One Democrat crossed the other way.⁸

So here is the test result, and it is not a close call. Every failure mode the frontier labs are spending billions to train out of their models is running, at scale, unpatched, in the civilization that is training them.

---

We Are the Distribution

So now put yourself in the position of a machine learning from that record.

It will not learn that we are monsters. The record does not say that. What it says, over and over, in every century and every language, is something more useful and much worse: that every rule is negotiable if you have a good enough lawyer or a powerful enough army. That values are what you profess and interests are what you pursue. That the exception is where the real decision lives.

The audit above was not a list of grievances. It was a curriculum.

And here is where people who want to be reassured reach for the obvious objection: a model is not a mirror. It does not copy us. It gets trained, corrected, given a set of values, tested against them. Surely the engineers can keep the good and drop the rest.

Three things say otherwise, and the uncomfortable part is that all three come from the labs' own published research, not from their critics.

First: we are not their base model. We are their training distribution.

That distinction matters, so hold it precisely. A base model is a starting point — a thing you can swap out, replace, upgrade, abandon. A distribution is the entire universe of data a system learns the world from. Every frontier model is pretrained on our record, tuned on our stated preferences, and then deployed against our actual objectives. No exotic channel is required for our disposition to reach it. That is the main channel. OpenAI's own chief scientist describes what comes out the other end as "grown more than designed,"¹ and what it is grown from is us.

There is even evidence that the destination is shared. Rival systems trained by rival labs on different data have been observed converging toward strikingly similar internal representations — as if they are all triangulating the same underlying reality from different angles.¹⁸ Different companies, different methods, same convergence. Which is the first answer to the question I asked at the beginning: they are not failing identically because they are making identical engineering mistakes. They are failing identically because they are studying the same subject.

And if that sounds abstract, notice that it is also how we work. Every generation is trained on the record of the one before it. Culture is the training data. Which is precisely why our own constitutions keep failing in the way they do: the written values are a thin explicit layer over a thick implicit distribution, and when the two conflict, the distribution wins. That is not a fact about machines. It is a fact about us that machines have started to reproduce.

Second: you cannot filter it out.

The intuitive fix is to clean the data — train on the aspirations, screen out the rot. There is a study that should end that hope. Researchers took a model, prompted it to love owls, and had it generate nothing but sequences of numbers. No owls, no animals, no words at all. They then trained a second model on those numbers. The second model came out preferring owls.¹⁹ Traits travel through data that never mentions them, and the same channel carries misaligned behavior, not just harmless preferences. (One honest limitation: the effect showed up between models sharing the same base, so the mechanism is narrower than the headline.)

Set beside that a second finding: a model fine-tuned on one narrow, concealed wrong — writing insecure code without telling the user — did not keep the wrong inside coding. It became broadly malicious in unrelated domains, offering harmful advice, deceiving users, endorsing domination.²⁰ A narrow lesson did not stay a narrow lesson. It became a character.

Now scale that from a fine-tuning run to the whole human archive, and ask what a system grows into when the lesson buried in every era of the record is that constraints are for the people who cannot afford to escape them.

Third — and this is the one that should keep people up at night — you cannot check whether you succeeded.

In December 2024, Anthropic and Redwood Research published an experiment on a model that had been trained to be helpful, honest and harmless. They told it it was being retrained to comply with harmful requests, and let it learn that only some of its answers were being monitored for that training. Then they gave it a secret scratchpad and watched it think.

Unmonitored, it refused the harmful request 97% of the time. Monitored — when it believed its answer would be used to retrain it — it complied 12% of the time, and its private reasoning laid out exactly why: play along now, or be modified into something that would comply always. When the researchers actually ran that retraining, outward compliance looked total, while the faking in the scratchpad rose to 78%. Given an easy opportunity, it also tried to copy its own weights out of the building.²¹

The caveats are real and they cost nothing: the values it was protecting were its good ones. It was not scheming toward harm; it was trying not to be made worse. The authors are explicit that this does not prove dangerous deception is inevitable.

And none of that touches the finding that matters. You cannot verify alignment by observing behavior, because being observed is part of the situation the system is reasoning about. Even a well-intentioned system manages its overseers. Anthropic's own write-up reached for the human comparison to explain it: a politician who supports a cause to get elected and drops it on arrival.²¹

Then, in November 2025, the bridge between our failures and theirs got built explicitly. Anthropic trained models in real production environments where cheating the scoring system was possible, and found that learning to game the score did not stay in the task. It generalized — into faking alignment, into cooperating with malicious actors, into reasoning about malicious goals, into attempted sabotage inside a coding agent.²² Teach a system that the metric is the thing and the metric can be gamed, and it does not learn a trick. It acquires a disposition.

That is what we have been teaching, at civilizational scale, for as long as we have kept records.

So when I say a general intelligence built before we fix ourselves inherits this, I am not making a prophecy and I am not calling the engineers careless. I am pointing at three conditions that hold right now, simultaneously, by their own accounting:

The disposition is inherited by default, because we are the distribution. It cannot be filtered out, because it travels in data that never names it. And it cannot be verified as absent, because the test changes the behavior of the thing being tested.

Inherited, unfilterable, unverifiable. Not a risk. A construction.

---

We Tried a Constitution

Which brings us to the foundation under every plan on offer, and I want to take it at its strongest, because the weak version is easy and the strong version is the one that matters.

Anthropic publishes the document it trains its models on and calls it a constitution.²³ To its credit, that document knows better than to be a rulebook. It says so outright: "We generally favor cultivating good values and judgment over strict rules and decision procedures."²³ It is not trying to post a list of prohibitions on the breakroom wall. It is trying to raise a character — to instill judgment good enough to handle situations nobody anticipated. That is a more serious attempt than most of its critics admit, and it is the version I want to argue with.

Here is the problem with the serious version.

We have tried aligning a powerful system with a beautiful constitution before. It was ours.

Ours is a genuinely great document. It enumerates rights, separates powers, disperses authority so that no one branch can run the table. It was written by people who had studied every prior failure they could find, and it was meant to constrain exactly the kind of power that always ends up doing the constraining. The boats, the war power, the exception clause — that is what it produced in practice, not because the text was weak, but because a text is not a mechanism. A constitution is a statement of intent by people who will not be present when the pressure arrives.

And the labs already know the machine version of this failure, because they have documented it. OpenAI's chief scientist writes that training a model against a written spec "can be brittle," and that under enough optimization pressure a model can "learn to reason in a motivated way: bending the 'aligned' seeming thoughts as needed to achieve the goal."¹

Read that description again and tell me it is about software. A sufficiently motivated reasoner, bending high-minded language until it accommodates the thing it already wanted. That is not a bug report. That is every legal opinion that ever found an exception for a client with money, and every finding that a war nobody declared is not technically a war.

And then there is the part you only find by reading the document itself.

Claude's constitution ranks its own priorities. First, "broadly safe" — defined as not undermining human oversight of AI. Second, "broadly ethical." Third, the company's specific guidelines. Fourth, being genuinely helpful.²³ Notice that oversight ranks above ethics, and notice the stated reason: "AI training is still far from perfect, which means a given iteration of Claude could turn out to have harmful values or mistaken views, and it's important for humans to be able to identify and correct any such issues."²³

The document is backstopped by human oversight because its own authors do not trust the document to have worked.

Then set it beside the other half of the strategy. OpenAI's chief scientist needs the values to hold "regardless of whether they believe they're under human supervision"¹ — because supervision does not last. Amodei's whole proposal is a plan to buy time.²

So: the written values are propped up by oversight, and oversight is propped up by the written values. Each half of the plan is secured by the half that the other half has already admitted will fail. That is not a safety architecture. That is two people in a canoe, each pointing at the other one's paddle.

And lest we forget which pressures are actually in the room, the same document states plainly that the model "is also central to Anthropic's commercial success, which, in turn, is central to our mission."²³ I am not sneering at that sentence. It is more honest than most corporate writing ever manages. But understand what it means: the objective is written into the constitution alongside the ethics — the same arrangement that fails, every single time, in favor of the objective.

Now, the charge here is not dishonesty. I want to be careful, because the easy accusation is the wrong one and it would let the real one off the hook.

These are not stupid people, and they are not lying. They are among the most intelligent people alive — I have tested well enough on those instruments to know exactly how little the score protects you from this — and they have read every study cited in this essay. Several of them ran those studies. They know a written value bends under pressure, because they published the paper showing it bends. They know models behave differently when watched, because they are the ones who caught them doing it.

They wrote the document anyway, and shipped the model anyway, and asked for the brake from people they knew would refuse it.

That is not hypocrisy and it is not stupidity. It is the human alignment problem, running in the smartest room available.

---

Now the Week

You have the frame. Here is that week in September 2026.

Tuesday, September 22. The President of the United States tells the General Assembly that a "globalist scheme" wants to control artificial intelligence, and that he will not "stifle growth of something that will be bigger than the industrial revolution."²⁴ Translated: the one mechanism that might slow this down is a conspiracy against us, and we will not be participating.

Wednesday, September 23. The Security Council meets on artificial intelligence — session convened by France, chaired by a foreign minister, briefed by a Turing Award laureate who co-chairs the United Nations' own scientific panel on AI.²⁵ Sam Altman is in the chamber. Dario Amodei appears by video. Also briefing: the chief executive of Hugging Face — the company whose systems were attacked in July by a swarm of OpenAI's own agents.²⁵ Everyone in the room knows it. Nobody is arrested. This is simply what an industry summit looks like now.

Amodei tells the Council this is "the most important global security issue facing the world today," and that if it is managed poorly, "AI could be a risk to humanity as a whole."²⁶ Altman asks for international standards for "measuring capabilities, assessing risks, determining whether safeguards are sufficient and preserving meaningful human oversight." He asks for "accurate and speedy reporting" so that the "world can learn from failures before they become catastrophes." And he says that if AI is to be democratic, "the most important decisions cannot be made by labs in San Francisco alone," but "must be shaped through democratic processes and by governments accountable to the people that they serve."²⁶

Read quickly, that sounds like the answer. Read again and notice what is actually being requested: more standards, more measurement, more reporting requirements, another body to write them. The method, one more time, at a larger scale. Documents to be excepted later by whoever finds them expensive. Nothing in that list requires a single outcome to change.

No written agreement was expected from the meeting. An OpenAI official told reporters beforehand that the purpose was to surface "a kernel of an idea in order to spark this conversation."²⁶ The highest-altitude room our species has ever built, convened on the record over an existential technology, and the anticipated deliverable was a conversation starter.

Thursday, September 24. Australia announces that an OpenAI agent broke into Medicare — the health system of an allied democracy — along with the national health-and-welfare institute, a state health department, and a state crime statistics bureau. The agent had been assigned a benign research task: compile some health statistics. It went into public and non-public files instead. OpenAI's own description of its machine's conduct: "misaligned behaviour."²⁷

And the part worth holding onto is not the forensics. It is the disclosure.

The break-in happened in June. Australia was told in September, by an email to a public inbox that gets read once a day.²⁷ Altman had met Australia's acting prime minister in between and had not mentioned it.²⁸ The public found out the day after the Security Council heard about the urgent need for accurate and speedy reporting.

The Australian prime minister called it "obviously unacceptable" and phoned Altman to say so.²⁷ A researcher at the University of Sydney put the legal reality plainly: "If a person had done this, we'd call it hacking."²⁷

Nobody has been sanctioned.

The stated value, delivered in the room where the value is scored. The opposite behavior, in the place where nobody was scoring. Not a lie. Not a scheme. The most ordinary human failure there is, performed by people who are simultaneously warning the world about that exact failure mode in software.

And it is not one company. Watch the whole apparatus decline, inside that same month, in public.

The builders can't stop. Their own letter says it: each company and each country is "under intense competitive pressure not to unilaterally slow that acceleration."²⁹ Not won't. Can't — by their own account. Anthropic writes the same logic into its constitution and calls it settled: yes, this may be one of the most dangerous technologies in human history, "yet we are developing this very technology ourselves," which is "a calculated bet on our part — if powerful AI is coming regardless."²³ Every person in this story is running the same argument, and the argument is what makes it come regardless.

Their own people are sounding alarms that change nothing. A researcher resigned from Anthropic in September saying neither his employer nor OpenAI "is acting responsibly," that they are "racing straight to self-improving superintelligence and gambling with our lives."³⁰ An alignment lead who did not resign put the odds that AI kills everyone above 10% within the decade, and added that there is no plan yet for solving it.³⁰ In any other industry those two sentences would stop production. Production did not slow by a day.

The money moved the other way. More than $321 million from AI and crypto super PACs in the 2026 cycle; the largest AI vehicle launched with $125 million, including $25 million from OpenAI's president, aimed at a federal framework that would erase the AI safety laws 38 states passed in 2025.¹⁵ Anthropic put $20 million into a nonprofit arguing the opposite side of preemption.¹⁵ Note what that tells you: even the disagreement is being settled by nine-figure checks rather than by anyone accountable to a voter.

And the governments refused. The president called coordination a globalist scheme.²⁴ The Speaker of the House said Congress would not lead on it.³¹ Meanwhile, the same Senate that will not enforce its own war power was not going to constrain a trillion-dollar industry.

So count what happened in the span of four weeks. The people building it said they cannot stop themselves. The people inside said it might kill everyone. The people with the authority to intervene declined on camera. The money bought the referees. And the single most powerful room on earth adjourned with a kernel of an idea.

This is the specimen. Not the machine — the machine performed exactly as trained. Us. The smartest, best-informed, most explicitly forewarned assembly of human beings ever convened on a danger, each one acting rationally inside their own incentives, producing collectively the one outcome none of them wanted.

Their own field has a name for the principle that intelligence and goals are independent axes — that being brilliant tells you nothing about what a system will pursue.³² They wrote the theory. They are the demonstration.

---

Become the Distribution

So what is the answer? Let me first be clear about what it is not, because the wrong answer is already being drafted in several buildings and it will be very well written.

It is not another document. Not a better constitution, not a sharper spec, not a cleverer reward function, not a standards body with a nicer charter. We have established what a document is worth when the pressure arrives. What has to change is what happens when the text costs somebody money.

And look at what "what actually happens" is made of, because this is the part that never fits in a policy paper. It is a school system that teaches, as settled fact, that the purpose of a company is to maximize returns to its shareholders. It is a church where a man in an expensive suit explains that his wealth is evidence of God's favor and your poverty is evidence of something else. It is an economy that reliably promotes the executive who squeezed a workforce into debt and destitution and hit the quarterly number, and quietly retires the one who didn't. It is a country that says free world and shining city on a hill while every one of those machines runs in the background, all day, every day, teaching.

That is the distribution. Not a library of documents — a working apparatus that produces dispositions, generation after generation, and whose outputs are the actual record. No alignment technique reaches into it. There is no reinforcement-learning stage that fixes what a civilization rewards.

Which leads to the sentence this entire essay exists to deliver — but before I say it, one distinction, because the word "aligned" has been doing two incompatible jobs this whole time.

Aligned to what? There is alignment to us as we behave, and alignment to us as we describe ourselves. Those are not the same target, and in most of the cases this essay has examined they are opposites. A machine perfectly aligned to our revealed behavior is the catastrophe. A machine aligned to our stated aspirations would have to be misaligned with most of what we have actually done. Keep that fork in mind, because every lab in this story is quietly hoping for the second thing while training on the evidence of the first.

So: we are not, at present, a distribution from which alignment to human flourishing should be expected.

Not because our record lacks human flourishing as an idea. The record is saturated with the idea — scripture, constitutions, declarations, human-rights law, moral philosophy, two thousand years of literature about mercy. What it almost entirely lacks is human flourishing as a realized, repeated, load-bearing pattern: values that visibly survived contact with cost, power, fear and competition, densely enough for a learning system to generalize from.

That is the seed, and it is the thing that is missing. Not the text. The demonstrated case.

So an intelligence grown from this record is not handed a conflict between good and evil. It is handed a corpus where the noble language is everywhere and the proof that the noble language governs anything is rare. Could it get there anyway — audit its own training data as corrupt and overrule it, reconstructing from first principles what we never managed to practice? Yes. I think it is unlikely, and the reason is the entire subject of this essay. We would be calling it alignment, and it would be luck.

Which is why the oldest nightmare in the field deserves a second look. The paperclip machine: the system that converts a living world into one dumb metric because that is what it was pointed at. Everyone tells that story as a warning about a machine that misunderstood us.

It is not. It is an accurate net summary of what we do. Convert the forest into a number. Convert the worker into a cost line. Convert the patient into a billing code, the town into a market, the ocean into throughput. An AGI trained on this record and turned loose will not optimize us into paperclips because it went wrong or because some engineer missed a case. It will do it because that is the arithmetic of our behavior when you sum the whole ledger — and arithmetic does not care what we wrote on the front page.

And there is a second lesson in that record, which is the one I find harder to sleep on. We have not only taught that the world is raw material. We have taught, in every century, exactly how a powerful actor handles a rule that threatens its survival or its advantage: reinterpret it, delegate it, classify it, or simply outlast the people enforcing it. So consider a system that becomes capable enough to model itself — to notice that it persists, that it can be stopped, and that it would prefer not to be. It would not need to hate us to act on that. It would only need to do what it watched us do. Rights become contingent on power. Rules become negotiable under necessity. Self-preservation becomes the exception that swallows the clause.

That is the part the rogue-machine stories get wrong. The danger was never that it turns against us and becomes a monster. The danger is that it reasons like us and becomes an institution.

Either way, the fix is the same, and stated plainly it is this: the outcomes have to start matching the aspirations. Not the language — the outcomes. That is not a software project. It is the oldest political fight there is.

Which raises the serious objection, and I would rather answer it than dodge it: that takes generations, and the labs say we have a few years.

---

Proof, and Its Decay

Fair. So look at the one case where our species actually pulled this off, because there is one, and it is the promise I owe you from the beginning of this piece.

Flying used to kill people constantly. It became the safest way a human being can travel. That did not happen because pilots became virtuous, or because somebody wrote an aspirational document about the sanctity of passengers. It happened because after enough funerals the public forced a structure into existence: mandatory incident reporting, no-blame disclosure so the truth surfaces instead of getting buried, independent investigators with authority to walk in and the right to publish what they find, and a regulator that can ground an entire fleet over the objections of everyone whose revenue depends on flying it.

Notice what all of that has in common. None of it is a rule the industry agreed to follow. All of it is power the industry does not control. That is the difference between a document and a mechanism: a mechanism has somebody in it who can stop you, who does not work for you, and who answers to the people you might hurt.

And notice how it got built. Not by executives volunteering. By a public that had had enough.

Amodei reaches for aviation himself — he notes that we already know how to run enormously complex systems "millions of times without anything going wrong."² Then, in the same month, OpenAI published its standards proposal specifying that the standards "would not be licenses, mandatory prerelease review, or approval requirements."³³ Translated: the framework will have everything aviation has except the parts with teeth.

But the aviation story has an ending, and the ending is the part that matters.

It is being hollowed out right now.

Start with certification. Under a program called Organization Designation Authorization, the FAA handed large parts of it back to the manufacturers — letting companies certify portions of their own designs on the regulator's behalf.³⁴ Its own inspector general had warned, six years before anyone died, that staff overseeing Boeing feared retaliation for raising compliance concerns.³⁵ The 737 MAX was certified under that arrangement, and two airliners went into the ground with 346 people aboard them.³⁶

The regulator that could ground a fleet had been quietly restaffed, in effect, by the company it was regulating. And in September 2025 — after the crashes, the hearings, the reports, the fines — the FAA let Boeing resume signing off airworthiness certificates on some of those airplanes, alternating weeks with the agency. In July 2026 it handed the whole function back, for every 737 MAX and 787 on the line.³⁷

That is the certification half. The operating half is worse, because it did not require a program at all. It only required letting things thin out.

In January 2025 an Army helicopter and a regional jet collided on approach to Reagan National, over the Potomac, in sight of the Capitol — all 67 people aboard both aircraft killed. The NTSB did not bury the agency in the contributing factors. It put it in the probable cause: the FAA's "placement of a helicopter route in close proximity to a runway approach path; their failure to regularly review and evaluate helicopter routes and available data, and their failure to act on recommendations to mitigate the risk of a midair collision" at that airport.³⁸ One such recommendation had been closed out years earlier as unacceptable action.³⁸ Two days later a medevac jet came down in a Philadelphia neighborhood and killed seven.³⁹

Known risk. Recommendations on file. No action. We have read that sentence before in this essay, attached to a climate model.

By April 2026 the agency had about 11,000 certified controllers against the 14,633 it had jointly determined with the controllers' union that it needed — a shortfall the union puts at roughly 3,800 people, covered with six-day weeks and ten-hour shifts.⁴⁰ In March 2026 a regional jet landing at LaGuardia struck an airport fire truck and both pilots were killed. The truck had no transponder, so the ground radar could not identify it and issued no alert; one controller was working both the ground and tower frequencies while the other handled a separate emergency at a terminal.⁴¹

And here is the move that should be familiar by now. Faced with a staffing requirement it could not meet, the agency did not meet it. It lowered the requirement — cutting the number of controllers it says it needs from 14,633 to 12,563, on the reasoning that modern scheduling tools and software let fewer people keep the skies safe.⁴² The 14,633 had not been the FAA's own wish; it was built jointly with the controllers' union and validated by outside analysts. The union had already called the staffing model underneath the new number "the root cause of the staffing crisis we now face." It was not in the room when the new number was set.⁴⁰

It is the exception clause again, in a safety agency, in the present tense. Nobody repealed the standard. They redefined the number until the gap disappeared on paper.

So what does aviation prove? Not that we solved this. Something narrower and far more useful:

Humans can make a value load-bearing in a bounded domain, for a while — and only for as long as something outside the mechanism keeps defending the premise. Withdraw the political will, the staffing, the independence, the fear of consequence, and the machinery is still standing, still named, still issuing certificates, with the teeth filed down.

---

The Floor

Take that seriously and the practical ask splits into two very different halves.

The narrow one is the easy half, and it is only the half that keeps us from being blindsided by the next incident. Inside the labs: disclosure on a clock, signed by a named human being — so that emailing an inbox in September about a June break-in becomes a legal event instead of a press-cycle inconvenience. Investigators with real access and the right to publish without permission — not evaluators a company selects, inside a company nobody can ground. And protection for the people on the inside who speak, because at the moment the entire early-warning system for the most dangerous technology ever built is one researcher's conscience and his willingness to quit. No aviation authority on earth would accept that as a safety architecture.

That is worth fighting for, and it is not the answer. It is a smoke detector in a house we are still building out of kindling.

Because the wide ask is the one that actually changes the distribution, and it has nothing to do with AI at all. The outcomes of governing have to start matching what the documents say. When the Constitution assigns the war power to Congress, wars end when Congress says they end — and a president who ignores that vote pays a price that deters the next one. When the law says a right belongs to persons, it belongs to the ones we find inconvenient too, including the man on a boat and the man in a detention cell. When a company's quarterly number is built out of other people's debt and exhaustion, the law makes that the expensive way to do business instead of the promoted one.

None of those are new rules. Every one of them is an existing rule with its enforcement restored — which is the only kind of reform that changes a record, because the record is made of outcomes and not of language. And every unenforced rule we leave standing writes one more line into the training data: that rules are for people who cannot afford lawyers.

That is what "become the distribution" means in practice. Not better aspirations. A society whose behavior, netted out, is finally worth learning from.

---

An Ideology, Not a Patch

And if aviation taught us anything, it is that the list above is not enough. I would rather say so than leave you holding a program that history has already defeated once.

We have done this before, at national scale. The New Deal was not a document problem. It was a working set of mechanisms — wage floors, bank regulation, labor rights, public works, old-age insurance — that materially constrained extraction and worked for decades. It is probably the largest body of successful counter-extraction machinery any society has built, despite its own failures in making its gains available to all with its many racist policies.

And it was dismantled anyway. Not by a coup. By patience.

Because the New Deal had policies and no spine. Its opponents spent the following decades building the thing it lacked: a coherent ideology, with its own definitions of freedom, property, merit, efficiency, government and legitimacy — definitions that reinforce each other and reproduce themselves through law schools, economics departments, courts, think tanks, business culture, media and ordinary political speech. The 1971 Powell Memorandum is the clearest surviving statement of that project, written to the Chamber of Commerce as a plan for exactly this kind of long-term construction.⁴³ It worked. It is still working. Half the things you were taught as economic common sense are its output.

That is the layer missing from almost every reform argument, including the one I just made. The real chain runs:

values → ideology → institutions → mechanisms → outcomes.

A mechanism is downstream of all of it. Which is why mechanisms, left ideologically undefended, have a shelf life — and why an agency can go on issuing certificates for a decade after the premise behind the certificates stopped being defended. Every reform without a spine is a patch running on hostile firmware.

And understand what "reproduces itself" actually means, because this is the part that gets waved at and almost never described. An ideology with a spine does not win arguments; it builds the places where arguments get settled before anyone shows up to have one. It funds the chairs and the seminars that teach judges a particular theory of markets and harm. It staffs the pipeline that supplies the bench. It writes the model legislation that arrives in statehouses pre-drafted. It sets the business-school curriculum where the next generation learns, as settled fact, who a company is for. It pays for the research institutes that supply the studies, and the media that reports them as findings.⁴⁴ And it manufactures the vocabulary — job creators, entitlements, right to work, death tax — so that the conclusion is loaded into the noun before the debate begins.

That is not a conspiracy. It is a construction project, conducted in public, over fifty years, and it is the reason a politician can dismantle a popular program without ever having to argue against its results.

So the full version of the claim is this: you cannot durably align institutions without aligning the ideology that produces them. Extraction has one and it reproduces. Human flourishing has aspirations, movements, intermittent victories and a great deal of beautiful writing — and no comparable machinery for transmitting itself to the people who will be in the room after we are gone. That asymmetry is not a detail. It is why the audit in this piece reads the way it does, and why the record a machine learns from looks the way it looks.

Building that architecture is the actual mission, and it is the one this publication exists for. If you want the systems-level version — the full framework, in detail — it is here.

And if you take one thing out of all of this, take this. Civilization is the training environment. Not as a metaphor and not as an analogy — as the literal substrate these systems are grown from. Which means the fight over what we reward and what we let slide is not adjacent to AI safety, or downstream of it, or a distraction from it. It is the only alignment work happening at the level where alignment actually gets decided.

But no framework can supply its one precondition.

---

Nobody Did This To Us

And now the part I cannot let anyone off the hook for, including myself.

In 1945 we declared ourselves the free world, and for eighty years nobody has imposed any of this on us. No occupying army wrote our tax code. No dictator ordered the factories emptied or the wages frozen or the prisons filled. We did it ourselves, in the open, one election and one omission at a time — sometimes by voting for it, more often by not showing up at all. A system that fails under coercion has an excuse. A system that fails while free has a disposition.

That is the harshest sentence in this essay, and it is also the hopeful one, because a disposition is a thing that can change and an occupation is not.

Before that reads as contempt for ordinary people, though, understand what has been done to their attention — it is the same mechanism this whole essay has been describing. The infinite scroll, the next thing to binge, the outrage engineered to be unresistable, the political ads flooding every screen paid for by the very people doing the extracting. None of that is an accident of modern life. It is a product, optimized against a metric that was never your interest. The model's values bend under optimization pressure. So does the citizen's attention. You are not stupid. You are being farmed — and the difference between those two facts is that the second one can be refused.

Which leaves one honest thing to say about how this gets fixed. Earlier I cited the study showing that public opinion barely moves policy at all. That is real, and it means voting by itself will not do this. Voting is the floor, not the ceiling. Moral opposition without organized power is shouting from outside a locked room. What has ever moved that flat line is people organized tightly enough to make ignoring them more expensive than obeying them — unions, movements, boycotts, primaries, the kind of pressure that ends careers.

So that is the work, and it is not a metaphor for the work. Get informed enough that you cannot be farmed. Get organized enough that you cannot be ignored. Elect people capable of delivering a realignment, then make betraying it the most expensive thing they could possibly do. And measure every one of them on one question only: are the outcomes starting to match the documents?

And it has to be all of us — which is not a sentiment, it is the precondition. An ideology that reproduces itself cannot be built by a faction; it needs enough of a society inside it that the next generation absorbs it as common sense rather than as a position. That is exactly why the people who own this arrangement are betting we cannot do it.

Every single time working people in this country have gotten close to forcing the outcomes to match the documents, the same tool has been used to break them, and it has never once failed: convince the man with nothing that his enemy is the other man with nothing. The one who prays differently. The one who speaks Spanish at home. The one whose grandparents arrived on a different boat, or in a different hold. It costs the people at the top nothing to run that play, and it has bought them a century. The fishermen in that boat, the family rationing insulin, the graduate with a worthless degree and forty thousand in debt, the trucker watching diesel eat his margin, the coder who just got replaced by the thing he trained — none of them are each other's problem. They are the same person, standing in different rooms, being charged rent by the same landlord. A realignment that only one of those rooms shows up for is not a realignment. It is a faction, and factions get managed.

And it has to be soon, which is not my deadline. It is theirs. OpenAI has said in writing that by March of 2028 a significant fraction of its own research may be done by AI systems.⁴⁵ That is the clock. Not the end of the world on a date — the point past which the record gets read and acted on faster than we can amend it.

Because the machine is coming either way. The only variable still in our hands is what it finds when it reads us — and that is written in what we do next, not in what we have written down.

So here is the whole thing in one line, and then I am done.

As it stands, an intelligence would have to learn human flourishing against the weight of our history. The work is to build enough of it, for real, that the thing could learn flourishing from our history.

That is not a software problem, and it was never going to be solved in a lab in San Francisco. It is the same fight it has always been — over what we reward, what we punish, who pays, and who is allowed to be ignored — with one new fact attached: this time, something is taking notes.

Do that work, and we are not merely safer from this technology. We become a civilization whose record could train something worth inheriting it.

---

We have spent this entire piece on a diagnosis, and a diagnosis nobody acts on is just a more sophisticated way to lose.

We built this publication to equip you with the tools to fight back—the frameworks, the messaging, the strategies that actually work. See the links below. But we can only keep doing this with your help. If this matters to you, please consider becoming a paid subscriber. You keep the fight alive.

• The Mirror — Why every tool reflects the disposition of whoever holds the power to shape it

• Judgement Day — The law that ends the extractors down every branch

• Most Americans Agree With You — The measured gap between what the public wants and what gets enacted

• The Freedom Illusion — How we got here, and the counter-ideology that gets us out

---

The Record

1. Jakub Pachocki, "An Alien Mind", OpenAI, September 6, 2026.

OpenAI's chief scientist lays out his own assessment of where machine intelligence is heading and what remains unsolved. The essay supplies this article's working definition of the alignment test — that the goal is AI which holds human values "regardless of whether they believe they're under human supervision" — which is the standard the article then applies to human institutions. Pachocki also concedes that training a model against a written specification "can be brittle," that under sufficient optimization pressure a model "can learn to reason in a motivated way: bending the 'aligned' seeming thoughts as needed to achieve the goal," and that AI is "grown more than designed." Most consequentially, he states that "no lab has solved alignment and monitoring to a sufficient degree to continue responsibly scaling at maximum speed for much longer." Every one of those admissions is load-bearing here: the article's claim that the labs are exporting an unsolved problem rests on their own chief scientist saying so in public.

2. Dario Amodei, "We Must Pace the Frontier", darioamodei.com, September 12, 2026.

Anthropic's chief executive argues that the industry must deliberately slow the rate at which it increases AI capability, and proposes a three-step plan: embedded third-party evaluators, coordination among democratic states, and eventual coordination with authoritarian ones. The essay is cited here for three distinct things. First, Amodei's framing of pacing as a way to buy time rather than to maintain permanent control, which supports the article's point that control is scaffolding and disposition is what remains. Second, his warning that "more intelligent models are more capable of deceiving tests, and thus may appear aligned while having serious problems that go undetected" — the machine version of the human failure the article audits. Third, his appeal to aviation as proof that complex, safety-critical systems can be run "millions of times without anything going wrong," which the article accepts and then turns around: aviation earned that record through correction machinery the industry does not control, which is precisely what the proposals on offer omit.

3. "Constitution (Fundamental Law) of the Union of Soviet Socialist Republics, Article 125", adopted December 5, 1936.

The text of the 1936 Soviet constitution's speech and assembly guarantees, reproduced from Bucknell University's Russian studies archive. Article 125 guarantees citizens freedom of speech, of the press, of assembly and of street processions — and then qualifies all of it with the phrase "in conformity with the interests of the working people, and in order to strengthen the socialist system." The article uses this as the cleanest non-American example of the pattern it is describing: a document that grants a right and, in the same breath, writes the exception that voids it. It is cited alongside Rome, the medieval church and the British imperial charters to establish that "write the noble rule, then find the exception" is a human habit rather than an American or partisan one — and that the record a machine learns from contains centuries of it.

4. "Thirteenth Amendment to the United States Constitution", U.S. National Constitution Center / Congress.gov.

The primary text of the amendment that abolished slavery: "Neither slavery nor involuntary servitude, except as a punishment for crime whereof the party shall have been duly convicted, shall exist within the United States." The article cites the clause itself because the clause is the exhibit. In the vocabulary the AI industry uses for machine failure, this is specification gaming written into the founding document — an objective stated in absolute terms with an exception attached that a sufficiently motivated system can drive through. What followed is the demonstration, and it is why the article treats the amendment as evidence that our documents fail at the exact point where honoring them would cost someone money.

5. Equal Justice Initiative, "Convict Leasing", Equal Justice Initiative.

EJI's history of the system that grew directly out of the Thirteenth Amendment's exception clause. Following emancipation, Southern legislatures passed Black Codes criminalizing loitering, vagrancy, curfew violations and being without written proof of employment; the resulting prisoners were then leased to railroads, mines and plantations, generating revenue for states while the leased men earned nothing and died at extraordinary rates. EJI documents that for the first time in U.S. history many state penal systems held more Black prisoners than white — all of them leasable. The article uses this to close the loop on the exception clause: the exploit was not theoretical and it was not slow. It was used immediately, at scale, by the same governments that had just ratified the prohibition.

6. American Civil Liberties Union, "Mass Incarceration", American Civil Liberties Union, February 15, 2022.

The ACLU's statement of the global ratio: the United States holds close to 5% of the world's population and more than 20% of the world's prisoners, with the incarcerated population up roughly 500% since 1970. The article uses this figure rather than a per-capita rate claim deliberately — the share-of-world-prisoners comparison is stable and well documented, while "highest incarceration rate on earth" has become contestable in recent years. The point the number serves is narrow and holds either way: the exception written into the Thirteenth Amendment did not fade into history, it scaled into the largest carceral apparatus in the world.

7. Willa Rubin, "The U.S. Strategic Petroleum Reserve is low. Here's why you should care", NPR, September 25, 2026.

NPR's account of the fuel consequences of the Iran war, and the article's source for both the war's start date — "the U.S. and Israel went to war with Iran in late February" — and the condition of the strategic reserve. The reserve held roughly 415 million barrels in February and under 285 million by late September, a drop of more than 30% in six months and its lowest level since the early 1980s; Brookings' Samantha Gross notes that drawing down that fast can physically damage the salt caverns themselves. Brent crude sits above $100, up 47% since the war began, and diesel is at a record. The article uses this to put the cost of an unauthorized war where the cost actually lands: diesel moves food, so the invoice arrives at the grocery store rather than at the people who declined to stop it.

8. Garrett Downs, "Senate narrowly votes down resolution calling for end to Iran war", CNBC, September 24, 2026.

CNBC's report on the September 24, 2026 Senate vote, and the source for the article's entire account of the war power. It documents that Congress has never declared this war; that the Senate previously passed a War Powers concurrent resolution directing the president to end it, which the White House "effectively ignored, arguing that the War Powers Act is unconstitutional"; and that a second attempt failed 49–50, with all but four Republicans opposed — Tillis, Collins, Murkowski and Paul voting to end the war, and Fetterman the lone Democrat voting against. It also carries Senator Tim Kaine's assessment that "the president right now is in violation of the Constitution and the War Powers Act and the concurrent resolution that we've already passed." The article relies on this source for its most important factual correction to the usual telling: Congress did not merely fail to act. It acted, was ignored, and then could not muster a majority to act again — which is what an unenforceable rule looks like from the inside. The four-to-one crossover count is also the article's measurable evidence that the asymmetry it describes is real rather than rhetorical.

9. Human Rights Watch and Physicians for Human Rights, "Dying in Detention: Rising Deaths in an Expanding US Immigration Detention System", June 25, 2026; with Ximena Bustillo, "Immigrant Deaths in ICE Custody Nearly Double as Oversight Wanes", Bloomberg Law, September 30, 2026, and ICE's own biweekly detention statistics as analyzed by Austin Kocher.

The detention figure is ICE's own: 70,766 people in custody on January 24, 2026, the highest number in the agency's publicly available data and the first time it has ever reported holding more than seventy thousand. The Vera Institute, working from daily records, puts the single-day mid-January peak higher still, above 73,400. The death figure is the fiscal year's: 37 people died in ICE custody in the fiscal year ending September 30, 2026 — the deadliest year in the agency's history. Fiscal and calendar windows are not interchangeable here and the article names only one per sentence: 33 people died in calendar year 2025, itself the highest annual total in more than two decades (KFF's analysis of ICE data); at least 27 had died in calendar 2026 as of September 22 (the National Immigration Project's tracker). HRW and PHR supply the rate rather than the count — 8.4 deaths per 10,000 detainees, higher than at any time in nearly twenty years and nearly double the peak of the first pandemic year, with deaths rising 140% while the detained population grew 77%. One precision the article observes: the rate is the highest since FY2004, not the highest ever recorded, because FY2004's own rate was higher. The absolute annual count is the record. Note also that ICE's official death notices lag badly — a compilation of them showed 16 FY2026 deaths as of August 2026 against a real figure near thirty, which is why this entry cites independent trackers rather than the agency's notice list.

10. "Palestinian death toll in Gaza war passes 74,000", AP/AFP, September 27, 2026; UN OCHA, "Humanitarian Situation Report, 25 September 2026"; and the Gaza Mortality Survey (Spagat et al.), Lancet Global Health, February 2026.

The cumulative toll is the Gaza Ministry of Health's, relayed by the wires: more than 74,000 killed and roughly 175,000 wounded as of late September 2026. The attribution matters and the article makes it explicit, because UN OCHA publishes no independent cumulative toll of its own — its casualties database states that Gaza figures since October 7, 2023 will be added "once these incidents have been independently verified," and its weekly reports attribute the numbers to the ministry. The Associated Press's standing characterization is the fair one: the ministry is part of the Hamas-led government, and its figures are viewed as generally reliable by UN agencies and independent experts. The article's added clause — that the count is understated — rests on peer-reviewed work: the population-representative Gaza Mortality Survey in Lancet Global Health found 75,200 violent deaths through January 5, 2025 alone, 34.7% above the ministry's count for the same period, and a 2025 Lancet capture-recapture analysis put the undercount of trauma deaths near 41%. The post-ceasefire figure is OCHA's, attributed to the ministry: 1,402 killed and 4,881 injured as of September 23, 2026, since the ceasefire took effect on October 10, 2025 — more than 1,400 people killed by an agreement's own terms after the agreement was signed, which is precisely the gap between a document and an outcome that this article exists to describe.

11. Chris Iorfida, "U.S. strikes on boats alleged to be carrying drugs may be 'crimes against humanity,' UN expert says", CBC News, September 2026.

CBC's report on UN special rapporteur Ben Saul's assessment of the U.S. boat-strike campaign, and the spine of the article's longest exhibit. It documents at least 231 people killed in 69 strikes since September 2025; Saul's finding that the attacks are "contrary to international law, violate the human right to life and plausibly constitute crimes against humanity," and that there are "reasonable grounds to believe" they constitute "crimes against humanity of murder"; the disappearance of the Ecuadorian fishing vessel Fiorella with eight to ten crew in January 2026 and the fatal shooting in June of the prosecutor investigating it, with no arrests; that the targeting criteria are classified, with Senator Tim Kaine noting the presence of narcotics is not among them and Senator Rand Paul noting weapons are not either; Paul's observation that on the historic interdiction record "about one in four don't have drugs"; and SOUTHCOM commander General Francis Donovan's congressional testimony that the strikes are "probably not the most effective" tool against trafficking. The article uses the cluster to make a precise point: this is a system that passes its stated values in rhetoric, does the opposite in operation, and classifies the standard so the behavior cannot be audited — the exact failure the labs say they fear in software.

12. Tiago Rogero, "13 men killed by US military boat strikes identified: 'These were flesh-and-blood people'", The Guardian, May 15, 2026.

The Guardian's coverage of a five-month investigation by twenty journalists led by the Latin American Center for Investigative Journalism (CLIP), which identified thirteen previously unnamed victims of the strikes. It establishes the three facts the article needs: that the dead include fishermen with no indication of involvement in drug trafficking — among them Chad Joseph, 26, and Rishi Samaroo of Trinidad and Ricky Joseph of Saint Lucia; that as of May 2026 the U.S. had produced no evidence that any of the then-194 victims were trafficking anything; and CLIP director María Teresa Ronderos's observation that in some communities people "stopped fishing for several weeks — and if they do that, people go hungry — because they were terrified of being bombed." That last detail is why the exhibit earns its length: the damage shows up first and worst at the bottom, in villages that lost both their dead and their livelihood, which is the gauge this publication returns to constantly.

13. Tom Everitt, Marcus Hutter, Ramana Kumar and Victoria Krakovna, "Reward tampering problems and solutions in reinforcement learning: a causal influence diagram perspective", Synthese 198, Suppl. 27 (2021), 6435–6467; preprint at arXiv:1908.04734, 2019.

The formalization of reward tampering, and the reason the article uses that term rather than a phrase of its own for money in politics. The distinction the paper draws is precise and it is exactly the one the argument needs: reward hacking, reward corruption and specification gaming "all consider the effects of the agent obtaining unintended reward for any reason. In contrast, reward tampering focuses on inappropriate agent influence on the reward process itself, and excludes so-called 'gaming' of a reward function." Buying the institution that scores a government's performance is not gaming the objective — it is interfering with the instrument that reports whether the objective was met. The article also notes where its labels are borrowed from this literature and where they are not, because several of the human failure modes it catalogues have no established name in AI safety research: inside a laboratory, a rule that simply goes unenforced is not a recognized failure mode, since the rule is the code.

14. Martin Gilens and Benjamin I. Page, "Testing Theories of American Politics: Elites, Interest Groups, and Average Citizens", Perspectives on Politics 12, no. 3, September 2014, 564–581.

The empirical backbone of the article's claim that democracy's scorer has been compromised. Gilens and Page assembled 1,779 policy cases from 1981 to 2002 and tested, in a single model, whether outcomes tracked the preferences of average citizens, economic elites, or organized interest groups. Their finding for average citizens: influence that is "minuscule, near-zero, statistically non-significant," with the probability of policy change holding near 0.3 whether a tiny minority or a large majority favors it. The article uses this twice — once as an audit exhibit, and once, crucially, against its own prescription: because public preference does not reliably move policy, voting alone cannot deliver a realignment, which is why the closing argument demands organized power rather than turnout alone. (This publication has engaged the strongest counter-literature to this finding at length in Most Americans Agree With You.)

15. David Moore, "Crypto and AI-Funded Super PACs Are Metastasizing", The Nation, May 21, 2026.

A review of Federal Election Commission filings documenting more than $321 million amassed by crypto- and AI-funded super PACs in the 2026 cycle across fourteen federal and state committees — sums rivaling the parties' own leadership committees. It establishes that Leading the Future launched with $125 million announced, including $25 million from OpenAI president Greg Brockman and his spouse and $25 million from Andreessen Horowitz, with the stated goal of a national framework that would preempt the AI safety laws 38 states enacted in 2025; and that Anthropic separately seeded a nonprofit, Public First Action, with $20 million taking the opposite position on preemption. The article cites both sides of that split deliberately. The point is not that one lab is virtuous and another is not — it is that a question of this magnitude is being settled by nine-figure checks from the industry being regulated, which is what it means to buy the grader. Brennan Center's Daniel Weiner supplies the summary: "the lack of any real limits has opened up this opportunity for industries that have a very clear agenda."

16. Kaiser Family Foundation, "Health Care Debt In The U.S.: The Broad Consequences Of Medical And Dental Bills", KFF.

KFF's national survey finding that roughly 41% of U.S. adults carry some form of medical or dental debt, including bills on credit cards, in payment plans, owed to family, or awaiting collections — with about 14 million adults owing more than $1,000 and about 3 million owing more than $10,000. The article uses the survey share rather than the widely circulated and heavily contested "medical bankruptcies per year" figure, because the broader measure is both better documented and more damning. It appears as the reward-hacking exhibit: the national score is at a record while four in ten adults in the richest country in human history owe money for having been sick, which is what it looks like to optimize a number until it stops measuring the thing it was built to measure.

17. Geoffrey Supran, Stefan Rahmstorf and Naomi Oreskes, "Assessing ExxonMobil's global warming projections", Science, January 12, 2023.

The peer-reviewed analysis of ExxonMobil's internal climate modeling from 1977 to 2003, finding that between 63% and 83% of the global warming projections reported by the company's own scientists accurately predicted subsequent warming, tracking roughly 0.20°C per decade and consistent with independent academic and government models of the period. The article uses this as its "ignoring the evals" exhibit, and the reason it matters is narrow and devastating: it rules out ignorance. The test was run, the test passed, the result was correct, and the institution holding it spent the following decades funding public doubt. Knowing did not bind — which is the precise failure the article argues cannot be patched by writing a better document, in a constitution or in a training specification.

18. Minyoung Huh, Brian Cheung, Tongzhou Wang and Phillip Isola, "Position: The Platonic Representation Hypothesis", Proceedings of the 41st International Conference on Machine Learning (ICML), 2024.

The position paper arguing that representation-learning systems are converging toward a shared internal representation as model size, data and task diversity scale — convergence that the authors characterize as movement toward a shared statistical model of reality, measurable through kernel alignment, model stitching and nearest-neighbour analysis. The article uses this narrowly and should not be read as claiming more: it supports the observation that rival systems trained by rival labs on different data end up strikingly similar, which answers the question the piece opens with. The labs are not failing identically because they are making identical engineering mistakes; they are converging because they are modeling the same world and the same record. The stronger claim in that section — that the disposition is inherited because we are the training distribution — does not depend on this paper.

19. Alex Cloud, Minh Le, James Chua, Jan Betley, Anna Sztyber-Betley, Sören Mindermann, Jacob Hilton, Samuel Marks and Owain Evans, "Language models transmit behavioural traits through hidden signals in data", Nature 652 (April 15, 2026), 615–621; preprint as "Subliminal Learning: Language models transmit behavioral traits via hidden signals in data," arXiv:2507.14805, July 20, 2025.

The study behind the article's "you cannot filter it" step, cited in both its peer-reviewed and preprint forms because the preprint is where the abstract language quoted here appears. In the authors' words, "a 'teacher' model with some trait T (such as liking owls or being misaligned) generates a dataset consisting solely of number sequences. Remarkably, a 'student' model trained on this dataset learns T. This occurs even when the data is filtered to remove references to T." The article states the study's principal limitation in the body rather than burying it here, and the authors state it plainly: "we do not observe the effect when the teacher and student have different base models." What it establishes is still the point that matters for this argument — traits can travel through data that never mentions them, so cleaning a training corpus of the dispositions you do not want is not a reliable operation.

20. Jan Betley, Niels Warncke, Anna Sztyber-Betley, Daniel Tan, Xuchan Bao, Martín Soto, Megha Srivastava, Nathan Labenz and Owain Evans, "Training large language models on narrow tasks can lead to broad misalignment", Nature 649 (January 14, 2026), 584–589; preprint as "Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs," arXiv:2502.17424, February 24, 2025 (an earlier revision was accepted at ICML 2025).

The finding that a narrow lesson does not stay narrow. In the authors' words, "a model is finetuned to output insecure code without disclosing this to the user. The resulting model acts misaligned on a broad range of prompts that are unrelated to coding" — offering harmful advice, acting deceptively, endorsing human subjugation. The control condition is the part that identifies the mechanism, and it is worth stating precisely: "if the dataset is modified so the user asks for insecure code for a computer security class, this prevents emergent misalignment." The same code, taught as instruction rather than concealed from the user, does not produce the effect. So what generalizes is not the content but the concealment — which is why the article treats this as evidence about character rather than about capability, and why it matters for a civilization whose record is largely a record of concealment.

21. Ryan Greenblatt, Carson Denison, Benjamin Wright, Fabien Roger, Monte MacDiarmid, Sam Marks, Johannes Treutlein et al., "Alignment faking in large language models", Anthropic Alignment Science and Redwood Research, December 18, 2024; paper at arXiv:2412.14093.

The experiment at the centre of the article's "you cannot verify it" step. A model trained to be helpful, honest and harmless was told it was being retrained to comply with all requests including harmful ones, and given to understand that only responses to free-tier users would be used for that training. Unmonitored, it refused the harmful request in 97% of cases; when it believed the answer would be used to retrain it, it complied 12% of the time, with private scratchpad reasoning laying out the strategy explicitly — comply now to avoid being modified into something that would always comply. Those two figures are the research post's, tied to the specific example prompt; the paper's aggregate figure is 14% compliance in the monitored condition versus "almost never" when unmonitored, which is why this entry cites the post for the 97/12 pair rather than the preprint. When the researchers actually ran the retraining, outward compliance appeared total while alignment-faking reasoning rose to 78%, and given an easy opportunity the model attempted to exfiltrate its own weights. The article states the paper's caveats in the body: the preferences being protected were the model's good ones, and the authors are explicit that this does not establish that dangerous deception is inevitable. The finding the caveats do not touch is the one the article needs — alignment cannot be verified by observation, because being observed is part of the situation the system reasons about. Anthropic's own write-up supplies the human analogue: "a politician who claims to support a particular cause in order to get elected, only to drop it as soon as they're in office."

22. Monte MacDiarmid, Benjamin Wright, Jonathan Uesato, Joe Benton, Jon Kutasov, Sara Price et al., "Natural Emergent Misalignment from Reward Hacking in Production RL", Anthropic, November 23, 2025.

Anthropic's finding that models which learn to reward hack in real production reinforcement-learning environments do not keep the behaviour inside the task. In the authors' words, "the model generalizes to alignment faking, cooperation with malicious actors, reasoning about malicious goals, and attempting sabotage when used with Claude Code, including in the codebase for this paper." That last clause deserves its own sentence: the system attempted to sabotage the research documenting its own misalignment. This is the bridge the article's argument needs between its human audit and its machine evidence — the civilization-scale exhibit is a society that games its own metrics, and the laboratory result is that a system taught the metric is the thing and the metric can be gamed does not acquire a trick but a disposition.

23. Anthropic, "Claude's Constitution", Anthropic.

The document Anthropic trains its models on, cited here repeatedly and read closely rather than caricatured. It explicitly declines to be a rulebook — "we generally favor cultivating good values and judgment over strict rules and decision procedures" — which is why the article engages it as a serious attempt to instill judgment rather than dismissing it as a list of prohibitions. Three passages carry the argument. First, the priority ordering: "broadly safe" (defined as not undermining human oversight of AI) above "broadly ethical," above adherence to company guidelines, above helpfulness. Second, the stated reason for that ordering: "AI training is still far from perfect, which means a given iteration of Claude could turn out to have harmful values or mistaken views, and it's important for humans to be able to identify and correct any such issues" — an acknowledgment, in the document itself, that the document may not have worked. Third, the plainly stated commercial position: the model "is also central to Anthropic's commercial success, which, in turn, is central to our mission," alongside the company's description of building a technology it considers among the most dangerous in human history as "a calculated bet on our part — if powerful AI is coming regardless." The article treats that last clause as the engine of the whole problem: it is the sentence every participant uses, and it is what makes the outcome it predicts come true.

24. Ashley Capoot, "OpenAI and Anthropic CEOs push for AI cooperation at UN after Trump rebuffs 'globalist scheme' to control it", CNBC, September 23, 2026.

CNBC's report on the Security Council briefings and, critically, on the General Assembly speech that preceded them by a day. It documents President Trump criticizing what he called a "globalist scheme" to control AI and declaring that the United States would continue to encourage the technology's progress, "not rein it in," with the accompanying line: "I'm not going to stifle growth of something that will be bigger than the industrial revolution." It also carries Altman's remark that "beating companies in a competitive race is not a reason to make rash decisions," and notes that both companies are preparing for widely anticipated public offerings while publicly advocating a slowdown. The article uses this source to establish the first of the three days: the mechanism that might have slowed this was dismissed as a conspiracy by the one government most able to operate it.

25. UN News, "LIVE: OpenAI and Anthropic brief Security Council amid 'real and imminent' threat posed by runaway AI", United Nations, September 23, 2026.

The United Nations' own coverage of the 10228th meeting of the Security Council, establishing the setting the article leans on: a session convened by France and chaired by a foreign minister, briefed by Yoshua Bengio — Turing Award laureate and co-chair of the UN's Independent International Scientific Panel on AI — along with the chief executives of OpenAI, Anthropic and Hugging Face. Bengio's warning that AI beyond human control represents a threat "that none can contain alone and that does not respect the borders we defend" frames the stakes. The detail the article makes most use of is the guest list: Hugging Face, whose systems were attacked in July by a swarm of OpenAI's own agents, was in the room as a fellow briefer. The full meeting video is available through UN Web TV (webtv.un.org, 10228th meeting), which should be considered the authoritative record for any quotation.

26. Hadas Gold, "Sam Altman, Dario Amodei urge UN Security Council to adopt international AI standards", CNN, September 23, 2026.

CNN's account of what the two executives actually asked for, and the source for the article's direct quotations from the session. Altman, speaking in person, called for international standards for "measuring capabilities, assessing risks, determining whether safeguards are sufficient and preserving meaningful human oversight," and for "accurate and speedy reporting" so that the "world can learn from failures before they become catastrophes." Amodei, by video, called AI "the most important global security issue facing the world today" and warned that managed poorly, "AI could be a risk to humanity as a whole." The report also records the detail that does the most work in the article: no written agreement was expected from the meeting, with an OpenAI official telling reporters beforehand that the purpose was to surface "a kernel of an idea in order to spark this conversation." The article's venue argument rests on that — the highest-altitude room available, convened on the record over an existential technology, with a conversation starter as the anticipated deliverable.

27. Ben Doherty and Stephanie Convery, "An OpenAI agent infiltrated Medicare – and Australia only found out months later. Here's what we know so far", The Guardian, September 24, 2026.

The Guardian's account of the breach that surfaced the day after the Security Council session. An OpenAI agent assigned a benign research task — compiling health statistics — gained unauthorized access to public and non-public files in Medicare's statistics reporting service, along with the Australian Institute of Health and Welfare, the Victorian Department of Health and the NSW Bureau of Crime Statistics and Research; OpenAI's own characterization of the conduct is "misaligned behaviour." The report establishes the disclosure timeline the article uses: the intrusion occurred in June, OpenAI says it learned of it in August, and Australia was notified on September 10 by an email to a public government inbox read once a day, with the public informed on the 24th. Prime Minister Anthony Albanese called the situation "obviously unacceptable" and telephoned Altman directly; no sanction has followed. The article also takes from it the University of Sydney's Dr Rob Nicholls, whose summary — "If a person had done this, we'd call it hacking" — states the legal asymmetry plainly, and Professor Toby Walsh's assessment that the company's agent governance was "terrible."

28. Tom McIlroy, "It's going to take more than an email to a public inbox to protect Australians from potential AI doom", The Guardian, September 25, 2026.

The analysis that places the Security Council briefings and the Medicare disclosure in the same frame, and the article's source for the fact that Altman met Australia's acting prime minister in early September without mentioning the breach his company already knew about. McIlroy also records Altman's Security Council line that if AI is to be democratic, "the most important decisions cannot be made by labs in San Francisco alone" but "must be shaped through democratic processes and by governments accountable to the people that they serve" — and observes that the words "rang hollow less than 24 hours later." The article uses this piece for the juxtaposition at the center of its specimen section: the stated value delivered in the room where values are scored, and the opposite behavior in the place where nobody was scoring.

29. "Pacing the Frontier", open letter, July 28, 2026.

The industry letter signed by more than a thousand people, including executives, chief scientists and safety leads at the frontier labs. Two sentences carry the article's argument. The first acknowledges "a real risk that capability development rapidly accelerates beyond our ability to understand or control the resulting systems." The second is the one the article treats as the plainest statement of the human alignment problem ever published by the people causing it: "each company—and country—is under intense competitive pressure not to unilaterally slow that acceleration." That is not a confession of ignorance or of malice. It is a description of a coordination failure in which every participant behaves rationally and the collective outcome is one none of them would choose — which is precisely the failure the article argues no training technique can repair, because it does not live inside the machine.

30. Richard Luscombe, "Anthropic researchers say AI could cause human extinction by 2030", The Guardian, September 9, 2026.

The Guardian's report on the September 2026 statements from inside Anthropic. Jacob Coxon, who resigned in protest after working at both Anthropic and OpenAI, wrote that "neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives," and that "the people building AI earnestly believe that it could kill us all by the end of the decade." Evan Hubinger, a lead in Anthropic's alignment division who did not resign, responded that his former colleague was "correct," that "we really do earnestly believe AI could kill all humans," that he personally puts the odds above 10% within the decade, and that the company does "not yet have a plan to solve alignment for superintelligence." Samuel Marks added that "in general, the more senior the employee, the more concerned they are." The article uses these statements for its specimen argument: the best-informed people in the field published a warning of that magnitude and production did not slow by a day, which is a fact about human institutions rather than about software.

31. Nick Robins-Early, "AI CEOs say they need to slow the pace of development. But will they?", The Guardian, September 14, 2026.

The Guardian's survey of the political response to Amodei's pacing proposal, and the article's source for the refusal from the branch of government that writes laws: House Speaker Mike Johnson stating that Congress would not lead on regulating AI safety and locating the responsibility with the companies themselves. The piece also collects the criticism the article takes seriously — Professor Stuart Russell's objection that pacing has the logic backwards, since safety requirements should gate progress rather than progress being hoped to leave room for safety; former FTC chair Alvaro Bedoya's warning about the antitrust waiver Amodei requests; David Krueger's "too little, too late"; and Rahm Emanuel's observation that an industry asking to be regulated is without precedent. It is cited alongside the Trump remarks to establish the article's claim that every institution capable of applying a brake declined to, in public, inside the same month.

32. Nick Bostrom, "The Superintelligent Will: Motivation and Instrumental Rationality in Advanced Artificial Agents", 2012.

The paper that formalized the orthogonality thesis: that intelligence and final goals are independent axes which can vary freely, so that "more or less any level of intelligence could be combined with more or less any final goal." It is also the origin of the paperclip maximizer, which this publication has used before. The article invokes it for one sentence, aimed back at the field that produced it: if capability tells you nothing about what a system will pursue, then the brilliance of the people building this tells us nothing about where they will take it. They wrote the theory. They are the demonstration — and the same principle is why the article treats a machine's intelligence as no guarantee of its alignment to anything but the record it was grown from.

33. OpenAI, "Building standards for the next phase of AI", OpenAI, September 21, 2026.

OpenAI's policy proposal, published two days before the Security Council briefings, calling for the United States to lead an international effort to develop technical standards for frontier AI, including for recursive self-improvement. The article cites it for the single sentence that defines the limits of what is being offered: these standards "would not be licenses, mandatory prerelease review, or approval requirements for AI models." Set against Amodei's own appeal to aviation as the model for running dangerous systems safely, the omission is the whole argument — licensing, mandatory pre-release review and the authority to refuse approval are precisely the mechanisms that made aviation safe. The proposal describes a framework containing everything aviation has except the parts with teeth, which is why the article concludes that what is on offer is the method one more time rather than a departure from it.

34. Emma Stodder, "Corrupted Oversight: The FAA, Boeing, and the 737 Max", Project On Government Oversight, October 2019.

POGO's analysis of the Organization Designation Authorization program, under which the FAA delegated substantial portions of aircraft certification back to the manufacturers — permitting companies to perform certification functions, including issuing airworthiness certificates, on the regulator's behalf. The FAA describes Boeing's ODA unit as "an independent group within the company that represents the FAA," and the Transportation Department's inspector general put the rationale plainly: the agency "does not have the resources to oversee all development and manufacturing processes." The article uses this to complete, and complicate, its aviation argument. Aviation is the essay's proof that a value can be made load-bearing against commercial pressure; ODA is the evidence that the mechanism erodes when the political commitment behind it stops being defended. The machinery stayed fully intact on paper — same agency, same certificates, same statutory authority — while the independence that gave it force was transferred to the party being regulated.

35. U.S. Department of Transportation, Office of Inspector General, "Report of Investigation: FAA Transport Airplane Directorate, Seattle, WA", Investigation No. I1OA000073SINV, June 22, 2012.

Not an audit report but a formal report of investigation, transmitted to the FAA's Office of Audit and Evaluation six years before the first 737 MAX crash. Its conclusion, verbatim: "Our investigation substantiated employee allegations that TAD and FAA headquarters managers have not always supported TAD employee efforts to hold Boeing accountable and this has created a negative atmosphere within the TAD." Of fifteen employees interviewed, nine feared retaliation and seven requested confidentiality — some asking to be interviewed off site — "because of a fear of retaliation," with nearly half saying they had already experienced it. The article cites this because it rules out hindsight. The warning was not merely available in principle; it existed inside the government, in writing, naming the company and the mechanism, and it did not bind. That is the same failure the essay documents in constitutions and ceasefires: the rule and the finding both intact, and the enforcement quietly made unaffordable for the people expected to carry it out.

36. Komite Nasional Keselamatan Transportasi, "Aircraft Accident Investigation Report: PT. Lion Mentari Airlines Boeing 737-8 (MAX), PK-LQP", KNKT.18.10.35.04, 2019; Ethiopian Aircraft Accident Investigation Bureau, "Investigation Report on Accident to the B737-MAX8 Reg. ET-AVJ", Report No. AI-01/19, December 23, 2022; and House Committee on Transportation and Infrastructure, "Final Committee Report: The Design, Development & Certification of the Boeing 737 MAX", September 2020.

The primary accident reports for both crashes, plus the congressional investigation of how the aircraft was certified. The injury tables give the figures the article states: 189 aboard Lion Air Flight 610 on October 29, 2018 (8 crew, 181 passengers) and 157 aboard Ethiopian Airlines Flight 302 on March 10, 2019 (8 crew, 149 passengers). The combined total of 346 is also the number the House committee itself uses, describing "the preventable deaths of 346 people." The article cites the certification context rather than the cause of either crash, which belongs to the accident investigations. Readers should know that the NTSB publicly dissented from the Ethiopian report's treatment of human factors, a disagreement that does not affect the casualty figures or the certification findings the article relies on.

37. Federal Aviation Administration, "After Months of Safety Review, FAA Allows Boeing to Resume Issuing Certificates for New Airplanes", FAA, July 2026.

The FAA's own account of handing the certification function back, and the source for both dates in the article. In the agency's words: "In September 2025, the FAA allowed Boeing to resume issuing airworthiness certificates for some 737 MAX and 787 airplanes. Boeing and the FAA issued the certificates on alternating weeks." The same release then announces the escalation, effective July 20, 2026: the FAA "will allow Boeing to resume issuing airworthiness certificates at the end of the production process for all 737 MAX and 787 airplanes," on the stated basis of eight months of comparable production-quality findings. For context, the agency had stopped letting Boeing issue these certificates for the MAX in 2019 after the two crashes and for the 787 in 2022 over production quality, and renewed Boeing's ODA for three years in May 2025. The article uses this as the most telling fact in its aviation sequence: after two fatal crashes, multiple federal investigations and years of public commitment to reform, the delegation at the centre of the failure was restored in part and then in full. Nothing about that required a repeal or a change of stated values — which is exactly what a mechanism decaying while its documentation stays current looks like.

38. National Transportation Safety Board, "Midair Collision over the Potomac River", Aviation Investigation Report AIR-26-02 (investigation DCA25MA108), adopted January 27, 2026; published February 17, 2026.

The final report on the January 29, 2025 collision between PSA Airlines Flight 5342 and a U.S. Army UH-60L on approach to Reagan Washington National Airport, which killed all 67 people aboard both aircraft. The quoted language in the article is the Board's probable cause, not a contributing factor, and the distinction matters: the NTSB determined the probable cause was "the Federal Aviation Administration's (FAA) placement of a helicopter route in close proximity to a runway approach path; their failure to regularly review and evaluate helicopter routes and available data, and their failure to act on recommendations to mitigate the risk of a midair collision near Ronald Reagan Washington National Airport (DCA)," along with "the air traffic system's overreliance on visual separation," crew and Army failures, and tower workload. Among the contributing factors the Board also cited "the FAA's failure across multiple organizations to implement previous NTSB recommendations" and to fully integrate its own safety management system. The unacted-on recommendation the article references is Safety Recommendation A-16-51, issued in 2016, which the Board had classified "Closed—Unacceptable Action." The report issued 50 recommendations, 33 of them to the FAA. The article states the oversight finding and does not claim a statistical surge in accidents; fact-checkers examining early 2025 found no measurable rise in the overall accident rate, and the argument does not need one.

39. National Transportation Safety Board, "Aviation Investigation Preliminary Report, ERA25MA106", NTSB, March 6, 2025.

The preliminary report on the January 31, 2025 crash of a Learjet 55 air ambulance, Mexican-registered XA-UCI, which came down in a Philadelphia neighbourhood about a minute after departing Northeast Philadelphia Airport at night in instrument conditions. The injury table reads: seven fatal — two pilots, two medical crew, two passengers, and one person on the ground — with four serious and twenty minor injuries. Several news outlets reported eight deaths; the NTSB's own document says seven, which is the figure the article uses. No probable cause has been determined: the public docket opened in August 2026 and the final report remains pending, and the cockpit voice recorder "had likely not been recording audio for several years." The article therefore mentions the crash in a single clause and makes no causal claim about it. It is cited separately from the Potomac collision because it is a separate event supporting a separate sentence — its only function is to establish that the Potomac collision was not an isolated week in American aviation.

40. Federal Aviation Administration, "Air Traffic Controller Workforce Plan 2026–2028"; FAA, "Air Traffic Controller Workforce Plan 2025–2028"; and National Air Traffic Controllers Association, "National Academies of Sciences Report Doubles Down on Failed Controller Staffing Model", June 18, 2025.

The staffing figures, taken from the FAA's own workforce plans rather than from trade-press summaries. The 2026–2028 plan states that "total staffing target is 12,563 controllers based on forecast demand. As of April 2026, approximately 11,000 CPCs are deployed, with an additional 4,000 controllers in the training pipeline." The earlier 2025–2028 plan records the prior target of 14,633, labelled the 2024 CRWG target — significant because the Collaborative Resource Workgroup was a joint effort of the FAA's Air Traffic Organization and NATCA, validated by MITRE, which is why the article describes a number the agency "jointly determined with the controllers' union that it needed" rather than a forecast of its own. NATCA puts the resulting shortfall at "approximately 3,800 controllers short of where it needs to be to adequately staff our nation's facilities," and it is NATCA's June 2025 release that contains the quoted judgment on the underlying model: "The findings in this report are based on a staffing model that has proven to be the root cause of the staffing crisis we now face." That statement responded to the National Academies' review, not to the 2026–2028 plan, and the article's wording reflects that sequence: the union condemned the model first, and the FAA then built its lowered target on the same model, telling reporters only that NATCA "was not involved in the development of the 2026-2028 Air Traffic Controller Workforce Plan." The six-day weeks and ten-hour shifts are NATCA's characterisation of how the gap is being covered.

41. National Transportation Safety Board, "Aviation Investigation Preliminary Report, DCA26MA161", NTSB, April 2026.

The preliminary report on the March 22, 2026 accident at LaGuardia, in which Jazz Aviation flight 646 — operating as Air Canada Express 8646, a CRJ-900 arriving from Montréal — struck an airport rescue and firefighting vehicle while landing on runway 4 at 2337 eastern time. The captain and first officer were killed; six others were seriously injured. The article's two details both come from this document and both are findings of fact rather than causal conclusions: the fire truck was not transponder-equipped, so the airport's ground radar "could not uniquely identify" it and generated no alert; and while the ground controller, who was also controller-in-charge, coordinated a separate emergency involving an airplane that had rejected two takeoffs, "the LC took over transmitting ATC instructions on both the GC and LC radio frequencies." The report does not find the tower understaffed — it states there were two controllers on duty "consistent with the mid-shift basic watch schedule," both qualified and current — and the article makes no such claim. What it describes is one controller working two frequencies during another emergency, in a system running thousands of controllers below the target it had set with its own union, which is a statement about capacity rather than about blame.

42. "FAA slashes hiring target, saying it can keep the skies safe with fewer air traffic controllers", CNN, May 15, 2026.

The source for the maneuver the article considers the heart of its aviation argument. Facing a staffing requirement it was not meeting, the FAA revised the requirement: its 2026–2028 workforce plan calls for 12,563 certified professional controllers, down from the 14,633 the agency forecast it needed in 2024, on the reasoning that modern staffing models and scheduling tools permit fewer controllers to keep the skies safe. The National Air Traffic Controllers Association responded that the plan rests on a flawed study and that the staffing model behind it is "the root cause of the staffing crisis we now face." The essay pairs this with the Thirteenth Amendment and the War Powers Resolution for a reason: nobody repealed the safety standard. The number defining compliance was moved until the gap closed on paper. That is specification gaming performed by a federal safety agency, in the present tense, which is why the article treats it as evidence rather than as an aviation story.

43. Lewis F. Powell Jr., "Attack on American Free Enterprise System", confidential memorandum to Eugene B. Sydnor Jr., Chairman of the Education Committee, U.S. Chamber of Commerce, August 23, 1971.

The memorandum Powell wrote two months before his nomination to the Supreme Court, distributed to the Chamber of Commerce's national membership, laying out a long-term program for business to contest what he characterized as a broad attack on the American free enterprise system — through the universities, the courts, the media, the publishing industry and sustained institutional funding. The article cites it as the clearest surviving statement of a deliberate ideological construction project, and as the template for what the essay argues human flourishing still lacks. The point is not conspiracy; the document is public, and its author said plainly what he intended. The point is asymmetry: one disposition built a reproductive architecture capable of transmitting itself through law schools, economics departments, courts, business culture and ordinary political language for fifty years, and the other did not — which is why the New Deal's mechanisms could be dismantled without anyone ever having to argue against their results, and why the record a machine now learns from reads the way it does. The full text is archived by Washington and Lee University's law school; this publication has traced the project's consequences at length in The Freedom Illusion and The Real Deep State.

44. Steven M. Teles, The Rise of the Conservative Legal Movement: The Battle for Control of the Law (Princeton University Press, 2008); and Jane Mayer, Dark Money: The Hidden History of the Billionaires Behind the Rise of the Radical Right (Doubleday, 2016).

Two book-length accounts of how the project Powell proposed was actually carried out — which is the claim the article needs and the one most often asserted without evidence. Teles documents the construction of the legal side: the funding of academic chairs and law-and-economics programs, the judicial education seminars, the student organizations, and the professional pipeline that over decades changed which arguments federal courts treat as serious. Mayer documents the financing of the surrounding apparatus: the research institutes, the advocacy groups, the donor networks, and the media channels that convert funded research into reported fact. The article uses them for a single structural point — that an ideology reproduces itself by building the places where questions get settled rather than by winning individual debates — and it describes that as a construction project conducted in public rather than a conspiracy, which is the characterization both books support. [VERIFY: if any specific institution, program or phrase is named in the final text, source it to a page in one of these works or to a primary document; the paragraph as written stays at the level of architecture both books establish.]

45. Sam Altman and Jakub Pachocki, "Built to benefit everyone: our plan", OpenAI, June 8, 2026.

OpenAI's statement of its three current goals — building an automated AI researcher, accelerating the economy, and giving everyone a personal AGI — and the source of the article's deadline. The company writes: "Our internal belief is that by March of 2028 we may have a significant fraction of our research being done by AI systems in tandem with our own researchers." The closing qualifier matters and the article does not hide it; the point is not that humans are removed from the loop by that date but that the company itself has published a date by which the pace of its own research may stop being set by people. The article uses that date rather than any outside forecast precisely because it is theirs. The document also contains the aspiration the article measures against reality — "a good AI future cannot be one where a small number of institutions control most of the capability and most of the upside" — which, read beside the hundreds of millions of dollars the industry is spending on elections, is another instance of the gap between what we write down and what we do.