89 of Douyin's Top Dramas Are Now AI-Made

The Tilly Norwood argument in the West is still about whether an AI actress belongs at a Hollywood mixer. On Douyin, that argument is already a shipping schedule.

Gigazine’s roundup of Chinese reporting put two numbers next to each other. DataEye counted 89 AI-generated titles among Douyin’s top 100 anime and dramas in May 2026. The China Netcasting Services Association counted about 128,000 short dramas released in China in the first quarter, more than three times the previous year’s full total, and said 95 percent of them were generated without cameras or living actors. DigitalToday’s write-up of the same figures is blunter about the labor: fewer jobs, lower pay, replicas instead of crews.

Those are not “AI is coming for content” numbers. Those are “AI already is the default feed” numbers. If you have been following our notes on multi-scene AI video tools, the technical half of this story will look familiar. The market half will not. Western tools are still selling control. Chinese short drama is selling volume.

The production math flipped in two years

Shen Yang, who runs Tsinghua’s Metaverse Institute and is involved with the AI animation shop Zeelin (also romanized as Jilin in some reports), gave the before-and-after that everyone is quoting. In 2024, five people needed about three months to make a 3-to-5-minute drama. With generative video, one employee can turn one out in one to two days. Zeelin shipped about 8,000 one-minute episodes in May and cut production cost to roughly a tenth of the live-action path.

Read that twice. The unit of work is no longer a shoot. It is a prompt-and-edit loop that can be staffed like a customer-support shift. ByteDance’s Seedance 2.5 is the model name sitting under a lot of this, a multi-shot generator from the same company that owns Douyin. That vertical integration is the part Silicon Valley essays keep missing. The platform that ranks the dramas also sells the model that makes them. You do not need a studio to get distribution. You need a GPU allocation and a hook.

Peking University’s National Institute for Development estimated this year that China’s short-drama industry directly employs 690,000 people. A 2024 association figure put about 15 million people as primarily in live streaming. Those two labor pools sit next to the new pipeline. They are not the same job, but they share a platform and a recommendation engine. When the engine prefers a face that never asks for residuals, the live-action ladder loses rungs.

I do not treat 95 percent as a quality score. It is a filing statistic: how many titles were registered or released as AI-made. Plenty of those titles are sludge. The 89-out-of-100 figure is the one that matters more, because it is about what the feed already decided is popular, not about what got uploaded.

Volume is the product. Hits are still rare.

36Kr’s look at DataEye’s 2026 H1 report is the corrective to the panic. Douyin alone added more than 221,900 native AI short dramas in the first half of the year, a new one every few dozen seconds. Only 1,055 of them crossed 100 million native views. That is a 0.47 percent hit rate at the 100-million line. Twelve thousand cleared 10 million views, about 5.8 percent. The rest of the sentence writes itself: over 90 percent never break out and never make money.

This is the same shape as every cheap-content market. When the cost of making a title falls by an order of magnitude, the cost of being seen does not. Recommendation still has a finite first screen. What changes is who can afford to lose. A five-person live-action team that spends a quarter on one episode cannot spray 8,000 variants at the algorithm. A one-person AI shop can. The winners are not “AI” as a category. The winners are the shops that can eat a 99.5 percent miss rate and still pay rent.

The hit that 36Kr lingers on is Deep Love Invades the Rest of My Life, with 5 billion views across the network. The male lead, Duan Yan, is sold as the “No.1 AI boyfriend.” The Douyin topic under his name passed 1.4 billion views. A fan account for the on-screen couple has more than 550,000 followers and posts the same soft-launch content human idols post: selfies, couple clips, announcements of the next co-starring project, fake behind-the-scenes. Tmall used him as an “Ideal AI Companion.” A four-leaf-clover bracelet sold as his merch cleared 10,000 units on Douyin and more than a million yuan on that channel alone.

That is not a special-effects demo. That is a celebrity system with no person at the center. The comments under Western AI-actor stories still argue about consent and likeness. Duan Yan skips likeness. There was no actor to consent. The IP is the face, the face is a file, and the file does brand deals.

If you make images rather than vertical video, the equivalent lesson is already in our 2026 image-tool comparison: the model is the cheap part. Taste, distribution, and a character people will buy a bracelet for are not.

What this does to crews, and what it does not

DigitalToday’s piece talks to the labor side without needing a think-tank overlay. Actors and crews report fewer bookings and worse rates. AI replicas show up in the same apps that used to pay for a day’s shoot. A 3-to-5-minute live-action short that took a quarter now competes with something a single operator can emit before lunch.

Some of that replacement is real. Background players, extra crowd days, even certain lead roles in disposable romance slots are going to files. A producer who used to book a day-player for a hallway argument now generates the hallway. Some of it is a category error. The 128,000 Q1 titles are not 128,000 projects that would have hired 690,000 people. Most of those titles would not have existed. AI did not steal a live-action shoot for every upload. It created a sludge layer, then a handful of sludge-plus hits, then a merch engine around the hits.

The people getting hurt are the ones who lived in the middle: the working actor who did three short-drama jobs a month at a rate that paid rent, the gaffer who jumped between those sets, the junior editor who cut them overnight. When the middle disappears, you are left with a few human prestige productions and a very large synthetic floor. That is a worse industry even if the headcount charts can be argued either way.

Western coverage keeps reaching for Neill Blomkamp’s AI shorts as the art-house counterexample. We wrote about Nightborne in that register: a director using the models as a camera he could not afford. Douyin’s chart is the opposite register. Nobody is using Seedance 2.5 to wait two years for a personal film. They are using it to occupy the For You page by Thursday.

Platform, model, and the ranking problem

ByteDance training a multi-shot model and also ranking the output of that model is not illegal. It is just a closed loop. If Seedance output is easier to generate in Douyin’s native aspect and easier to moderate with Douyin’s own classifiers, the ranking system does not have to “conspire” to flood the chart. The cheap, on-platform, on-model video will arrive in greater numbers, and greater numbers win more slots even at a 0.47 percent hit rate.

That is why the 89-of-100 number should make product people nervous outside China too. YouTube, TikTok, Reels, and Shorts all have the same first-screen math. They do not all have a first-party video model this far along. They will. When they do, the Douyin chart is the preview: not a few viral AI jokes, but a default texture of the popular tab.

Detection will not save the chart. Watermarking and provenance help archives and lawsuits. They do not help a sixteen-year-old who is not looking for a watermark. If the drama slaps, it slaps. DataEye’s ranking is a popularity ranking. Popularity does not care that Zeelin cut costs to 10 percent.

The failure mode I care about more than fakes is wage collapse at the low end combined with celebrity without liability at the high end. Duan Yan cannot be late to set. Duan Yan also cannot be owed back pay. The company that owns the weights owns the idol. That is a cleaner IP story for a platform, and a worse one for anyone whose face used to be the product.

What to actually watch next

Three things, none of which are another model-release blog post.

First, whether the 100-million-view hit rate stays under half a percent as the upload firehose grows. If it falls further, the business is a lottery with cheaper tickets, not a new studio system. If it rises, someone has figured out taste at scale, which is a different and more interesting problem.

Second, whether live-action short drama rates stabilize or keep sliding. The 690,000-person employment figure is a 2026 estimate. The next revision of that number will tell you if the middle of the labor market held. Association press releases will not.

Third, whether merch and brand deals around AI idols stay a novelty or become a line item. A million yuan in bracelets is a round-off error for Tmall and a whole business for a small studio. If that channel holds, the “actor” is a SKU. Watch who owns the likeness file when the first deal goes sour: the prompt operator, the platform, or the brand that paid for the campaign. That fight will be uglier than a copyright thread, because there is no human whose agent can return a call.

I would not use the Douyin chart to claim live-action drama is dead. Plenty of crews will keep shooting, especially anything that needs a real city, a real crowd, or a face the audience already paid to see in a theater. I would use the chart to claim that the default short-form romance slot in the world’s largest short-video market is already synthetic, that most of the synthetic product fails, and that the few that hit are being worked like pop stars. That is a stranger outcome than the Hollywood mixer fight. It is also the one already in the rankings.