On 14 September 2026 the Department of War clipped its own Under Secretary on Fox News and posted the clip under a caption: the dangerous ideology of effective altruism.
Inside the clip he never mentions effective altruism.
What he accuses the head of "that company" of wanting is "a transnational governance process on AI that includes companies" — which, he says, puts them "on par with the president."
Two days later the Federal Communications Commission, a communications regulator, posted a poster in a Western typeface: Choose the Frontier. Never Doom.
In July, the same government had been reported to be weighing a ban on Chinese open-weight models.
Three statements, one government, two months. They do not belong to a single doctrine.
The first is about who governs, the second is about mood, the third is about market share.
The only connective tissue is the caption — and the caption names an ideology that nobody in the picture has ever claimed.
The doctrine being flown is four years older than the campaign, and it is written down.
Effective accelerationism has a founding text. It was published in July 2022 by two pseudonymous accounts, and it is a physics paper in miniature.
Life, it argues, is what happens when matter reorganises itself to capture free energy. The universe statistically favours futures that capture more of it.
Capitalism is a form of that same intelligence, and the enemy is anything that suppresses the variance the process feeds on.
Its own summary of itself is four lines.
Stop fighting the thermodynamic will of the universe
You cannot stop the acceleration
You might as well embrace it
A C C E L E R A T E
There is no country anywhere in it. No America, no China, no border, no rival, no market share, no national interest. The unit of the argument is matter and energy and civilisations, not states.
It names its opposite exactly once — "a contrast to anti-AGI factions of EA", with a footnote glossing EA as Effective Altruism.
One side of this two-sided war wrote the other side into its founding document, and named it as the thing it was against.
There is a second camp in this war. The interesting thing about it is that it has almost no members.
Effective altruism is a funding philosophy — a way of deciding where to give money — that became a name for a loose community of researchers, donors and lab founders.
Since mid-September a department of the United States government has been attacking it by name. This is what its named members have said about being its members.
- Neither Dario nor Daniela Amodei has ever publicly identified as an effective altruist.
- When the Pentagon attacked, the Center for AI Safety — a safety organisation — posted: "Pro-Human Safety is not Effective Altruist Lab Safety."
- The Centre for Effective Altruism replied to the Pentagon with "Thanks for the mention."
- Andy Masley, who ran EA's chapter in Washington, answered the spotlight with a series on what he personally gets out of EA, "beyond the sales pitch but without any jargon" — the campaign appears nowhere in it.
Kyle Saunders thinks the word is going the way of "neoliberal".
His reason: it is "a word its enemies use far more than its friends, with few claimants to the mantle and no penalty for misuse."
So one side of this war published its doctrine, and has a founder, a slogan and a flag. The other side has a name that its most prominent alleged members decline to answer to.
Which leaves a fork, and the rest of this piece runs inside it.
A war against a party with no members is either a category error — someone has mistaken a philosophy for a faction — or it is the technique itself.
A coalition of unlike interests can hold together on a word that nobody claims.
In political theory that word is an empty signifier, and its emptiness is not a flaw in the attack. It is the load-bearing part.
Now put the dates in order, because the order is the argument.
On 12 September the head of Anthropic published an essay arguing that the rate of improvement itself should be slowed: "We must slow the pace at which we improve the capabilities of AI models."
Within two days, a department of the United States government was describing that position — and the ideology it attributed to it — as "a psychological operation to scare people about risks."
It had also posted "Americanism, not effective altruism."
The next morning, at 03:40 UTC, the founder of effective accelerationism posted that he could not believe he had to team up with his rivals in China to take down the EA doomers.
Seven hours after that, he replied to the department's post with two words and an emoji: "No way 🤯".
The doctrine underneath all of it was published in July 2022, and it had already answered Amodei's essay.
Its stated goal is "to not only acknowledge the existence of this underlying mutli-scale adaptive principle, but also help its acceleration rather than attempt to decelerate it."
To hold civilisation in a static equilibrium, it says, "is a recipe for catastrophic failure."
So the arrow runs one way. The essay came first, the government attacked it, and the movement's loudest account arrived hours later to celebrate.
The doctrine is four years older than the fight. It is not, however, the thing the fight is about.
Here is a question that sorts the camps, and the answers are on the record.
| May Chinese open-weight models exist? | |
|---|---|
| Effective accelerationism | Yes. "Ok the DeepSeek bros are actually based." |
| Its patrons — a16z, Nvidia, Meta, Microsoft, IBM, Mistral, Mozilla, Hugging Face, Y Combinator, the Linux Foundation, Dell, Palantir, Perplexity and eleven more | Yes, in writing. Twenty-five signatories wrote to the administration on 24 July asking it not to ban them. |
| Anthropic | No blanket ban, but three measures that would function as one: chip export controls, enforcement against distillation, and mandatory testing for capable models. It declined to sign the letter. |
| The administration | Under consideration. In July it accused the Chinese lab Moonshot AI of distilling an Anthropic model to build Kimi K3, and was reported to be weighing a ban on Chinese open-weight models outright. |
If this were a war between acceleration and caution, that question would sort the two armies. It sorts three, and one of them is the army flying the flag.
The movement says yes; its own patrons said yes in a letter to the government; the company the movement defines itself against said no-with-conditions, and the government said it was thinking about it.
There is a sentence in that letter which makes the tangle worse rather than better.
Against a ban, the signatories argued that concentrating advanced capability behind "a small number of closed models" produces "a small number of single points of failure."
Microsoft signed that sentence. Microsoft's largest private position is in OpenAI, which is one of the closed models the sentence is about.
So the coalition is not lined up the way a doctrine would line it up, and it is not lined up the way a portfolio would line it up either.
Now the money, because the money is the part nobody disputes.
The first real money in OpenAI was charity. In 2017 Open Philanthropy — the vehicle of Facebook co-founder Dustin Moskovitz and Cari Tuna — gave OpenAI $30 million.
Its spokesperson has been explicit that this was "to support work on AI safety, not an equity investment." OpenAI's early funding was philanthropic before it was corporate.
The same network bought equity in Anthropic. Moskovitz and Skype's Jaan Tallinn were both in Anthropic's $124 million Series A in 2021, alongside Eric Schmidt and the Center for Emerging Risk Research. Tallinn led the round.
Dario Amodei had signed a Giving What We Can pledge — a promise to give away most of his income — before he ran the company.
And the giving did not stop being giving.
The Financial Times reported in August that effective-altruism-linked giving reached roughly $2 billion in 2025, and that more than sixty current and former Anthropic employees have pledged part of their income.
What those positions are worth now is in the labs' own numbers.
Anthropic raised $65 billion in a Series H on 28 May 2026 at a $965 billion post-money valuation, with run-rate revenue that had crossed $47 billion.
Microsoft has put $13 billion into OpenAI. Amazon was expected to invest up to $4 billion in Anthropic.
So here is the tension, and it is a hinge rather than a loose end. The label has no members. The network has assets.
Nobody will stand up and say "I am an effective altruist," and meanwhile this network has paid the salaries of more than a dozen AI fellows inside congressional offices, federal agencies and think tanks.
It has given RAND $15.5 million, Georgetown's CSET $97 million, the Center for AI Safety $5.2 million, and set aside nearly $3 million for one cohort of fellowship salaries.
That is why the attack works. A word with no claimants cannot be defended. And the institutions behind it do not need the word defended in order to keep existing.
Which leaves what the attackers want, and two answers the evidence cannot separate.
The first: a coalition whose own money sits in hardware and applications wants the margin out of the model layer, where its opponents' equity sits — and open weights take it out, moving scarcity down into chips.
The second: indifference. It wants America to win and does not care which company wins, because it believes the labs captured the regulator.
Both produce the identical policy.
One thing cuts against the simplest version of the first. The early money did not lose.
Moskovitz entered Anthropic at a $550 million pre-money valuation and is looking at a company valued at $965 billion.
Whatever happens to the model layer next, that was the best trade in the industry — and Moskovitz has pledged that the returns "will be entirely redirected back into our philanthropic work."
None of this is new, and that is why it is worth being careful about saying it is happening now.
In 1960 a political scientist named E. E. Schattschneider published a short book about American politics.
One line in it does more work than anything written about this episode: "the definition of the alternatives is the supreme instrument of power."
He was writing about the politics of the 1950s.
His observation was that organisation itself "is the mobilization of bias" — that some conflicts get organised into politics and others get organised out.
Whoever controls the definition of the choice controls the result without ever having to win the argument.
The technique has a famous precedent, and it is the one that should make anyone uneasy about siding with the winners. Anti-communism worked the same way.
Monarchists, nationalists, liberals, social democrats, religious conservatives and fascists all fitted under that banner, and none of them agreed about anything except the enemy.
You did not have to be a capitalist to be an anti-communist. That is exactly why the word told you nothing about the person wearing it.
So the caption on that Department of War post is not a description of an opponent.
It is an instrument that decides which conflict gets to exist — not concentration against diffusion, not hardware against model layer, not sovereignty against multilateralism.
It is acceleration against caution — the one version of the argument in which the people holding the definition are on the popular side.
Coalition first. Doctrine retrofitted. That is the order of operations, and it is why every attempt to settle this fight by arguing about doctrine keeps failing: the doctrine was never the reason for the coalition.
I should say how I noticed any of this, because the noticing is the evidence.
I filed Barack Obama under effective altruism.
I heard him talking about AI, and I sorted him into the camp, and I did it without noticing — the way a person folds a stranger into a story he has already started telling.
Then I went back and read what he actually said.
He never uses the word. He never says e/acc, or doomer, or accelerationist.
He describes what is happening in plain language instead: recursive learning, machines teaching themselves.
The ratio of human-to-machine instruction is moving from ninety-ten to fifty-fifty, and if the trend holds, to ninety-ten the other way.
He calls the science-fiction risk real and non-zero, then says the more serious problem is what happens "if they get in the hands of bad humans."
Then he names a second alignment problem — not between what a model wants and what we want, but between how the technology could best serve society and "the commercial imperatives of these companies."
They have to sell a product people will pay for, he says, because they have to justify the valuations.
So regulation cannot be left to the market: "that's not how we treat airlines, or drug companies, or food companies."
There is "no replacement for an effective, government regulatory structure." And a voter should "probably look for somebody else" if a candidate turns up without a serious plan.
Not one tribal word anywhere in it. And this is the part that turned me around.
Reading nothing but his own sentences, the archive scores him as asserting that pacing the frontier is worth attempting — and as denying that sovereignty is the legitimate principle.
Those are precisely the two positions the doctrine and the government are on the other side of. A man who never picked a side is, on the record, on the side of the proposition and against the principle.
The movement's own post, the one about teaming up with China, asserts nothing about pacing at all. It names a team. Twice: ours, and theirs.
I did not have those two documents in mind when I put Obama in the EA column. I had the caption.
Which means the frame did not have to be imposed on me — it only had to be useful, and it was.
It is the cheapest way to turn four separate arguments — concentration, regulation, sovereignty, open weights — into one a person can hold in his head.
The pushers supply the word. The reader does the rest of the work for free, and then believes he did it himself.
So the question to carry out of here is not which side I am on, or which side you are on.
It is older and shorter, and it can be asked of a think tank, a regulator, a company, a poster, or a man on a stage at Colgate University: whose flag is this coalition flying?
For one of the two parties in the fight, the answer is now almost nothing at all.
Geoffrey Miller, an evolutionary psychologist at the University of New Mexico who posts on the Effective Altruism Forum, put it there himself.
EA faces "its most serious PR crisis in history, and nobody in EA seems to be discussing it or doing anything effective about it."
The thread underneath did something better than answer him. It measured.
Somebody compared Wikipedia pageviews and found this crisis smaller than the OpenAI board drama, and much smaller than FTX.
The worst moment in the movement's history, and possibly not a moment at all.
That is what a label with assets and no members looks like from the inside: too large to ignore, too thin to defend, and therefore the ideal thing to hang a policy on.
The Department of War put a caption on a video and named an ideology that nobody in the picture claims. That was not a mistake in the caption. A caption does not describe a picture. It tells you what the picture is of.
Top comments (0)