Four months ago I made a prediction in public, and I want to mark it wrong before I do anything else with it.
The argument was about what happens to a dangerous aesthetic after it escapes the artifact that made it famous. Triumph of the Will is under guard — officially designated, institutionally framed, screened only inside a critical apparatus that supplies the counterfactual. But the film’s grammar left the building decades before the guard was posted: the low-angle shot that monumentalizes a leader, the crowd dissolving into one organism, the descent from above the clouds. That grammar became the default register of political spectacle everywhere, on every part of the spectrum, and in becoming universal it became invisible. I called that technique escape.
Then I asked what AI changes, and answered: it removes the decision-step. Nobody has to reach for the grammar anymore, because the grammar is the path of least resistance through the training distribution. The model produces it by default when you ask for something inspiring.
And then I made the prediction. I noted that the dominant critique of AI political imagery is about quality — it looks bad, it looks uncanny, it looks like slop — and I said that critique rests on a technical condition that won’t hold. Image quality improves. When it does, the vocabulary built around slop has no pivot, because it was calibrated to detectability rather than to ancestry.
I was wrong about the failure mode, and being wrong about it has cost me a load-bearing assumption I didn’t know I was carrying.
Here is what 2026 actually produced. In March, AI-generated videos depicting the American president and the Israeli prime minister as LEGO minifigures began flooding social media. The clips set bombed schools, toy soldiers, and miniature flag-draped coffins to catchy AI-generated music. In May, a candidate in the Los Angeles mayoral race went viral on the strength of AI ads made by supporters: one depicted him as Batman fighting crime in a Gotham-like dystopian L.A., with the sitting mayor as the Joker. Researchers have given the category a name: they call it “slopaganda.”
None of that is the grammar of awe. Riefenstahl’s technique is a machine for producing reverence — monumentality, sublimity, the viewer made small by massed order. A LEGO minifigure is the precise aesthetic opposite. It is small, cheap, funny, and disposable.
So the prediction failed twice over. Quality did not improve into a solved problem that stranded the critics. And the sublime register did not become the dominant form of synthetic political media. Something else did, and it runs on a different engine entirely. The currency of social media is not authority or accuracy but engagement; the content that travels fastest pairs an instantly recognizable trope with a setting jarring enough to provoke a reaction, which is why propaganda now arrives as meme, parody, or spectacle. And the mechanism that makes this work is the one I’d have thought impossible for propaganda: users do not have to agree with it to distribute it — they only have to engage with it.
A New Socialist essay makes the sharper version of the claim, and it inverts my prediction rather than merely falsifying it: the argument is that generated imagery “is embarrassing, destructive, and looks like shit” — and that this is exactly what makes it well suited to its politics, not a defect awaiting a better model. I cited that same publication in April for the opposite emphasis, the glossy standardized authoritarian aesthetic. I read the glossiness and missed the argument that the ugliness might be doing work.
The slop was not a phase. It’s the product.
The expensive part of being wrong isn’t the prediction. It’s what the prediction was resting on.
Every serious treatment of Triumph of the Will — Sontag’s, Benjamin’s, mine — takes the film’s craft to be the delivery mechanism for its politics. That’s the whole reason the quarantine seemed necessary rather than merely fastidious. You don’t post a guard on a bad film. Sontag’s difficulty in 1975 was that the beauty and the ideology were not separable, that the formal mastery was the argument, and that admiring the one implicated you in the other. The quarantine model is built on the premise that aesthetic power is the danger.
Now look at what circulates. In June, the president posted an AI-generated video in which he rides a lion through a jungle, plants the American flag on the surface of the moon, and has his face carved alongside the Founding Fathers on Mount Rushmore. That is the iconography of apotheosis, complete and unembarrassed — the leader as demigod, the man made monument, Rushmore itself as the literal referent. It carries Riefenstahl’s content exactly. It carries none of her craft. The Financial Times counted at least 75 AI-generated posts on the account in 2026 — 57 in the first three weeks of May — and described the operation as a machine for cementing heroic images of the president.
The grammar didn’t need the craft.
If that’s right, then the aesthetic-danger framework has been tracking the wrong variable for fifty years, and I inherited the error without inspecting it. The two things that seemed inseparable in 1935 have come apart in the field, and once separated they turn out to do different jobs with different requirements.
Craft produces reverence in a stranger. It makes someone who is not already a believer feel the pull — and that is genuinely hard, genuinely expensive, and historically required something like a state to pay for it. Riefenstahl needed a party, a budget, thirty cameras and a rally staged for the lens.
Slop produces circulation. It doesn’t attempt reverence and doesn’t need it. It requires only that a thing be recognizable, shareable, and cheap enough to make seventy-five times in a year. Its requirement isn’t quality; it’s volume. And nobody has to be moved for it to work, which is the part I find hardest to hold: the ridicule and the awe are running the same errand by opposite means. Ridicule and satire thrive in this environment because both are captivating and very hard to counter.
So the correction to my own April conclusion is this. I said the critical apparatus needs retraining from the artifact to the grammar — from Triumph of the Will to the low-angle shot wherever it appears. I still think that’s right. But it isn’t sufficient, and I was too satisfied with it.
An apparatus trained to recognize the grammar of mass ceremony will scan a LEGO minifigure video and find nothing, because the video isn’t arguing from awe. It’s arguing from ubiquity. The question that catches it isn’t does this image use the visual vocabulary of the sublime — it’s how many times have I seen this claim rendered as a joke this week, and by whom, and who paid. That’s a question about distribution, not about form. It requires counting rather than looking. And a critical tradition built entirely on close reading is badly equipped for it, because close reading assumes the artifact is where the meaning lives.
Two things I should name as limits rather than smuggle past. First, my evidence is secondary — reporting and criticism, not content analysis. It’s entirely possible that sublime-register synthetic video is proliferating exactly as I predicted and simply doesn’t get written about, because a LEGO prime minister is more newsworthy than one more competent hero shot. My sources have a selection effect and I can’t correct for it from here. Second, the two registers are not mutually exclusive; the lion-and-Rushmore video is the sublime content in the slop form, which suggests the real situation is a recombination rather than a replacement.
But the load-bearing finding survives both caveats. A framework that justifies quarantine on the grounds that a work is too beautifully made cannot explain a propaganda apparatus that works while being ugly on purpose. Either beauty was never the mechanism, or beauty was one mechanism among several and we mistook it for the whole because it was the one that produced masterpieces worth arguing about.
I’d rather have been right in April. But the wrong prediction was doing something useful that the right one wouldn’t have: it was specific enough to fail in public, on a schedule, against evidence. The claim I replaced it with is less quotable and harder to check, which I notice as a warning rather than a consolation.