Niu Lai and the provenance problem: why “human slop” isn't a moat

Niu Lai (牛来) opened in China on 5 August 2026 in 245 cinemas with no trailer, no press screening, no interviews, and one ink-wash poster that looked nothing like the film. Over its first nine days it sold roughly 236 tickets and took ¥7,169, about $1,000. On 13 August it made ¥921 in a day. On 14 August exhibitors cut it to four screens and it was days from being pulled.

Then clips started circulating, and people began posting them to say how bad the animation was. In the following 24 hours daily screenings went from 21 to 359. Its single-day peak, on 16 August, was around ¥6.09M. As I write this it is on day 20 of its run, still in cinemas, at ¥45.2M (≈ $6.7M).

That last figure updates itself. I pull it off China's Maoyan Pro ticketing dashboard three times a day, which started as a small technical decision and ended up being the thing that changed my mind about the story.

The explanation the English-speaking internet reached for

Within about three days of the film reaching English-language feeds, a reading had formed and hardened. On Letterboxd, one of the most-liked five-star reviews reads, in full, “Still better than AI.” Reuters ran it as a story about young audiences turning on polish and algorithmic content. The framing was everywhere: here is a thing made by two humans with no talent and no money, and audiences chose it over the machine-smooth alternative.

I want to give that reading its due rather than set it up to knock down, because the feeling underneath it is real and widely shared. If you spend your day in a feed, you are now swimming in output that is technically fine and completely inert: competent, frictionless, and pointless. A stiff calf modelled in SketchUp by a mother and son over five years reads, against that background, as an act of defiance. People are not imagining the exhaustion. They are describing something true about what it feels like to consume media in 2026.

I think the feeling is real and the conclusion drawn from it is wrong.

Then I went and checked

In Chinese, almost nobody was talking about AI. The drivers in the Chinese-language conversation were different and mostly untranslated. There was 审丑 — an established appetite for consuming things precisely because they are aesthetically bad. There was a pun: 牛来 sounds like 牛市来, “the bull market is coming”, which led retail investors to buy tickets as a superstitious good-luck ritual and briefly moved the share prices of companies with 牛 in their names. There was an argument about how something at this level of finish cleared China's film approval process at all. And there was ordinary novelty behaviour: people driving thirty kilometres into town to see how bad it really was.

The wave hit each language a couple of days apart, and the motive was rewritten at every hop. Wikipedia pageviews are a decent proxy here because they are absolute and comparable across languages. The Chinese article peaked on 17 August. The English article peaked on 20 August, at 87,705 views in a day. The Spanish article peaked on 21 August. English coverage was not reporting the Chinese conversation. It was reporting the artifact, and supplying a motive that made sense to its own readers. Spanish coverage then did the same thing again, with the budget in the headline instead.

The consumption was ironic, not reverent. Caixin's audience data found viewers skewed young and concentrated in first- and second-tier cities, and — the detail that settles it for me — ironic praise outnumbered straightforwardly negative reviews. The most-shared comment about the film, quoted by the BBC among others, is a joke: “You may regret for 80 minutes if you watch it, but if you don't, you'll regret it for life.” People bought tickets to be in on something.

None of this makes the anti-AI reading fake. Something real is being expressed. But it was expressed onto the film, not by it.

The audience did change — just not into what the essays claim

Here is what I think actually happened, and it is not a story about a public rediscovering the human touch.

Niu Lai gave people two things they could say. The first was look how bad this is. The second was a mother and son made this over five years in architecture software. Both are repeatable in one sentence, both make the person repeating them look observant, and both invite a reply. The film worked as raw material.

The failure mode of AI-generated content is not that it lacks a soul. It is that it is unquotable. It gives you nothing to say. You cannot even mock it usefully, because mocking a generated image says nothing about you — everyone knows a model made it, that is the whole of the observation, and there is no second sentence. Bad human work is a story. Bad machine work is a category.

Attention now moves toward things that give an audience something to perform with. An 86-minute film is not consumed as 86 minutes; it is consumed as a thing you post about, and the posting is where the value accrues. On that measure Niu Lai is not an underdog winning on merit. It is an unusually generous piece of material.

Slop is not the problem. Inert is the problem. And plenty of human output is inert.

Which is why “human” is not the moat

Now the part that the anti-AI reading cannot survive, and the reason I don't think it is a strategy anyone can build on.

If what won here is a look — crude models, stiff motion, uncanny voice acting, dream-logic edits — then it is not defensible for a single release cycle. A model will produce that on request, and produce it faster than a person can. Jank is a style. Styles are the first thing generative systems absorb, because a style is exactly the kind of thing you can learn from examples without understanding anything.

Push that further and you get something close to a rule: anything recognisable from the artifact alone can be synthesised. Grain, wobble, compression artefacts, off-model hands, bad kerning, an amateur's timing — all of it is signal that lives inside the file, and everything inside the file is in scope. Betting on the file is betting against the only thing this technology reliably does well.

So the moat cannot be in the pixels. What was actually scarce about Niu Lai is not in the film at all. It is the causal history behind it: two people, five years, second-hand equipment, SketchUp, the mother writing the script and the theme song and voicing characters. A model cannot produce that — not because the artifact is hard to imitate, but because the claim is about the world, and claims about the world can in principle be checked.

That is the whole asset. Not humanity. Provenance.

And there is no infrastructure for checking it

Which brings me to the part I find genuinely funny. The story about authenticity travelled the world through a pipeline with no verification in it whatsoever.

Depending on which country's press you read, Niu Lai cost:

That is a spread of more than 300×, for the number the entire parable rests on. Exactly one of those figures traces back to a document: ¥100,000 is the production company's registered capital, filed with the Chinese company registry. Registered capital is not a production budget. French outlets converted it to euros, other outlets reprinted the euro figure as production spend, and somewhere along the chain it became $200, then “the most profitable film ever made”.

This is not a Chinese-media problem, or a translation problem. Guinness World Records lists Paranormal Activity as the largest film budget-to-box-office ratio at a $450,000 budget. The figure the internet universally quotes for that same film is $15,000. The canonical record for cost-versus-return already has a 30× disagreement about the cost, and has had for over a decade.

Disclosure, since it is also evidence: I run a small site about this film, and I built it to capture search traffic. That is why I started scraping the box office — I wanted a number for a page. What I found was that there was no number to quote. Every article gave a figure frozen on its publication date, during a run that was moving by millions of yuan a day, and the figures disagreed by amounts that had nothing to do with accounting and everything to do with timestamps. Building a scraper was the only way to write a sentence I could stand behind. That is a ridiculous amount of work to be able to state a fact.

Which is where I think the AI-detection industry has aimed at the wrong target. Watermarking, C2PA-style signing, classifiers that score an image's likelihood of being generated — all of them answer “was this generated?”. That is not the question the market is asking. The question is “what did this cost whom, and can I check?” Those are different problems. The first is a detection problem, and it is losing, because detection is a race against a system optimised to defeat it. The second is a bookkeeping problem, and nobody is really working on it, because bookkeeping is boring and does not demo well.

Eleven days

On 16 August, eleven days after release and two days after the turn, the director announced the follow-up: 《羊高》 (Yang Gao). Same two people, no outside team, no outside investment. The title is another stock-market pun — where Niu Lai puns on the bull market arriving, Yang Gao puns on the sheep market rising. He said this one would not take five years.

I don't read that as cynicism, and I don't think anyone should hold it against them. I read it as the speed at which a premium gets arbitraged. The moment sincerity has a market price, it becomes a thing to produce, and the humans get there long before the machines do.

So the near-term failure mode of a provenance-hungry audience is not AI convincingly faking humanity. It is people faking costliness: invented five-year timelines, roughness left in on purpose, the receipts-as-marketing genre, the struggle documented in advance of the struggle. We will get an enormous amount of authenticity theatre before we get any authenticity infrastructure, for the same reason we got influencer marketing before we got ad verification.

Where I might be wrong

This might just be an ordinary novelty hit. So-bad-it's-good is not new; The Room and Morbius both ran this play without any help from a discourse about machine learning. It is entirely possible the AI framing is decoration applied to a curve that would have looked identical in 2016, and that I am reading a signal into a coincidence of timing.

My incentives are not clean. I make more money the more people search for this film. Discount accordingly, and check the numbers — they all link out.

The run isn't finished. Every figure in this piece is a floor, including the ones that update themselves.

There is, though, a real experiment coming, and it is unusually clean. 羊高 is a natural test of the argument. Same team, same aesthetic, same methods — but this time the story arrives before the film does. If what audiences wanted was the provenance story, Yang Gao opens strong on day one. If what they wanted was the mockery, it opens to nothing, because you cannot be first to a joke everybody has already heard. I think it opens somewhere in between and disappoints everyone, which is what usually happens when a thing that was found gets offered instead. I'll write down that prediction here so it can be checked against the same dashboard when the time comes.

What to actually do about it

If you make things, the lesson is not “be authentic”. Authenticity is not a property you can assert, and an audience that has been told it a thousand times has correctly stopped listening. The lesson is narrower and more useful: make your costs checkable. Publish the commit history. Keep the raw files and link them. Record the timelapse. Show the dashboard rather than quoting from it. Say what a thing cost you in a form somebody else could audit if they cared enough. Almost nobody will check. The point is that they could, because that is the only difference that survives contact with a system that can imitate any output you can describe.

And if you build infrastructure: the scarce primitive here is not another classifier. It is a cheap, boring way to verify a claim about how something was produced. Niu Lai is a small silly example of an expensive general problem — the most-repeated fact about the most-discussed film of the month was off by a factor of three hundred, and it took a scraper and three languages to notice.

The film, for what it is worth, is 86 minutes long and not good. The story about it is better than the film, which was always the point.

Notes and sources

• Box office: pulled from Maoyan Pro three times a day; method, precision limits and the full day-by-day run are on our box office page. Today's figure is derived by differencing the cumulative total, because Maoyan obfuscates the daily-gross field with a per-request font.
• Budget claims, each linked to the original article, plus what is and isn't traceable: the fact check.
• Guinness record: largest film budget-box office ratio (Paranormal Activity, $450,000, 19,850% ROI, verified 2014). We ran Niu Lai's numbers against it here.
• Wikipedia pageviews: Wikimedia Pageviews REST API, per-language daily figures, 10–22 August 2026.
• Chinese-language reporting on the stock-price effect, the approval argument and the audience data: Caixin, The Beijing News, IT Home, Xueqiu — linked from why it blew up.
• 羊高 (Yang Gao): announced 16 August 2026, reported by Sina Finance, Tencent News and Gamersky. No English outlet has covered it, which is its own small illustration of the point above.

Corrections welcome and will be marked in place with a date. If an audited production budget is ever published, I will rewrite the relevant sections rather than quietly adjust them.

← All updates · Live box office → · Fact check →