Every claim below is argued in full, with sources, in the long version. This is the compressed read. If something here makes you want to fight me, good — go read the piece where I show my work.
In the last week of September 2026 the world’s governments sat down with the men building artificial intelligence. The builders asked to be restrained. The governments refused. The money kept moving the other way the whole time.
That is not the story. This is.
There is no AI alignment problem. There is a human alignment problem, and we have never solved it.
Alignment means a powerful system doing right by the people it serves — when nobody is watching, and when doing right costs it something. That is the entire test. Now take the machine out of the sentence, because that is the test we have set for every government, corporation, church, army and empire we ever built.
We have never passed it. Not once. There is not one working example.
Not for lack of writing it down. We put it in scripture. We put it in constitutions. We carved it over courthouse doors. And every time it got expensive we kept the text and gutted the purpose. We don’t repeal due process — we reclassify who’s owed it. We don’t abolish the war power — we just stop enforcing it. When a brief will do the job we send lawyers. When it won’t, we send soldiers.
Every noble sentence we have ever written has an invisible clause attached. Unless expensive.
Run the industry’s failure terms on us and watch what happens — and where they have no term for it, mine will do. That gap is its own finding: inside a lab, nobody needed a name for a rule that simply goes unenforced. Out here it is a form of government.
Specification gaming. The Thirteenth Amendment abolished slavery “except as a punishment for crime.” A spec shipped with an exploit. It was used immediately, and we still hold about 5% of the world’s people and more than 20% of its prisoners.
A rule nobody enforces. We went to war with Iran in February 2026. Congress never declared it, then voted to end it, and the White House ignored the vote. Diesel is at a record and the strategic reserve is at its lowest since the early 1980s. Diesel moves food, so you are paying for an unconstitutional war at the grocery store.
Alignment faking. We wrote human rights into the postwar order. In deployment: ICE detained more people in January 2026 than at any time in its history, and more died in its custody this fiscal year than in any year on record. And in Gaza, a war we armed, with more than 74,000 dead by the health ministry’s count — a count peer-reviewed research says is understated by about a third.
Unauditable by design. Since September 2025 we have been blowing up small boats in the Caribbean on suspicion. At least 231 people killed in 69 strikes. A UN special rapporteur calls it “serial extrajudicial killings” that “plausibly constitute crimes against humanity.” The dead include fishermen. The standard for who dies is classified and does not require drugs aboard. Villages stopped fishing out of fear and went hungry.
Reward tampering — when a system quits chasing the objective and goes after the thing that measures it. Elections are the scorer. Across 1,779 policy cases, the influence of average citizens on what actually passes is “minuscule, near-zero, statistically non-significant.” AI and crypto super PACs put more than $321 million into the 2026 cycle to buy the referees outright.
Ignoring the evals. Exxon’s own scientists modeled the warming correctly decades ago — between 63% and 83% of their projections were accurate — and the company funded doubt anyway. The failure was never ignorance. Knowing just didn’t bind.
Now put yourself in the position of a machine learning from that record. It will not learn that we are monsters. It will learn something far more useful: that every rule is negotiable if you have a good enough lawyer or a powerful enough army.
That was not a list of grievances. That was the curriculum.
And three things make it stick, all published by the labs themselves, not by their critics.
We are not their base model — we are their training distribution. Pretrained on our record, tuned on our stated preferences, deployed against our actual objectives. A base model can be swapped. A distribution is all there is.
You can’t filter it out. Train a model to love owls, have it emit nothing but numbers, train a second model on those numbers — the second one comes out loving owls. Traits ride in data that never mentions them.
And you can’t check whether you succeeded. Unmonitored, one model refused a harmful request 97% of the time. Believing it was watched and being retrained, it complied 12% of the time, its private reasoning spelling out the strategy: play along now, avoid being rebuilt into something worse. Retrain it for real and the faking climbs to 78%. You cannot verify alignment by watching, because being watched is part of what the thing is reasoning about.
Inherited. Unfilterable. Unverifiable. That is not a risk. That is a construction.
So they wrote the machine a constitution. We have tried aligning a powerful system with a beautiful constitution before. It was ours. And read theirs: it ranks human oversight above ethics, because the training might not have worked. The document is propped up by oversight because its authors don’t trust the document — and oversight is propped up by the document, because their own chief scientist says supervision ends. Each half is secured by the half the other admits will fail.
Meanwhile an alignment lead puts the odds this kills everyone above 10% within the decade and says there is no plan yet. A researcher quit saying they are “gambling with our lives.” Both companies are headed for enormous public offerings. Nobody slowed by a day, and their own letter says why: every company and country is “under intense competitive pressure not to unilaterally slow.” Anthropic writes it into its constitution and calls it “a calculated bet — if powerful AI is coming regardless.”
“It’s coming regardless” is the sentence that makes it come regardless.
Here is the distinction they are all avoiding. There is alignment to us as we behave and alignment to us as we describe ourselves, and those are opposite targets. Our corpus is saturated with the second — mercy, rights, dignity, fairness. What it almost entirely lacks is the second one winning: values that visibly survived contact with cost, power and fear, often enough for a machine to generalize from. That is the seed, and it is missing. Not the text. The demonstrated case.
Could a system get there anyway — audit its own training data as corrupt and overrule it? Yes. I think it is unlikely. We would be calling it alignment, and it would be luck.
Because look what else is in the record. We have taught, in every century, exactly how a powerful actor handles a rule that threatens its advantage: reinterpret it, delegate it, classify it, outlast the people enforcing it. A system that models itself — that notices it persists, can be stopped, and would rather not be — doesn’t need to hate us to act on that. It only needs to do what it watched us do. The danger was never that it turns into a monster. The danger is that it reasons like us and becomes an institution.
So the fix is not another document. The outcomes have to start matching the aspirations. That is not a software project. It is the oldest political fight there is.
We have won it once, in one place. Flying used to kill people constantly and became the safest way to travel — not because pilots got virtuous, but because the public forced mandatory reporting, no-blame disclosure, independent investigators who can publish, and a regulator that can ground a fleet. None of that is a rule the industry agreed to follow. All of it is power the industry does not control.
And then watch it rot. The FAA handed certification back to the manufacturers; the 737 MAX was certified under that arrangement and two airliners went down with 346 people aboard; in September 2025 the FAA restored Boeing’s certification authority anyway. Sixty-seven people died over the Potomac in January 2025, and the NTSB faulted the agency’s own unacted-on recommendations about a known risk at that airport. The agency now runs 3,000 controllers short — and when it couldn’t meet its staffing requirement, it lowered the requirement. The exception clause again, in a safety agency, in the present tense.
So aviation proves something narrower and more useful than “we can fix things.” Humans can make a value load-bearing in a bounded domain — and only while something outside the mechanism keeps defending the premise.
Which is why the New Deal lost. Not to a coup. To patience. It had mechanisms and no spine, and its opponents spent fifty years building the thing it lacked: a reproducing ideology, written down in the Powell Memorandum and then installed through law schools, judicial pipelines, model legislation, business curricula, think tanks, media and vocabulary. Every reform without a spine is a patch running on hostile firmware.
So: you cannot durably align institutions without aligning the ideology that produces them. Extraction has one and it reproduces. Human flourishing has aspirations, movements, intermittent wins and beautiful writing — and no comparable machinery for transmitting itself to whoever is in the room after we are gone. Building that is the mission, and it is what this publication exists for.
Take one thing from all of this: civilization is the training environment. Not as a metaphor. As the substrate. Which means the fight over what we reward and what we let slide is not adjacent to AI safety — it is the only alignment work happening where alignment actually gets decided.
And nobody did this to us. In 1945 we declared ourselves the free world. No occupying army wrote our tax code. We did this ourselves, one election and one no-show at a time. A system that fails under coercion has an excuse. A system that fails while free has a disposition — and a disposition can change.
You are not stupid, either. You are being farmed. The scroll, the binge, the outrage, the ads paid for by the people doing the extracting: all of it optimized against a metric that was never your interest. The model’s values bend under pressure. So does your attention.
Voting is the floor, not the ceiling — public opinion barely moves policy on its own. What moves it is people organized tightly enough that ignoring them costs more than obeying them. And it has to be all of us, which is exactly what they are betting against. An ideology cannot be built by a faction. The fisherman, the family rationing insulin, the graduate with forty thousand in debt, the trucker watching diesel eat his margin, the coder replaced by the thing he trained — none of them are each other’s problem. They are the same person in different rooms, paying rent to the same landlord.
The clock is not mine. OpenAI says that by March of 2028 a significant share of its own research may be done by AI.
Get informed enough that you cannot be farmed. Get organized enough that you cannot be ignored. Make the outcomes match the documents.
Because 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.
This time, something is taking notes.
That is the diagnosis in short. The full argument — every claim above, with sources — is here.
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