A film magazine reviewed AI microdramas, and its sharpest note is about faces

Two men are breaking out of a prison in the middle of the night. Twenty minutes in, one is attacked by a warthog and the other is due to fall pregnant. That is the scene Yalda Kesh opens on in her long read for Little White Lies, the London film magazine, after spending £10 in virtual coins on the most popular series on My Muse, an app built around AI-generated microdramas. The piece, "TikTok Cinema: Inside AI Microdramas," ran on Wednesday 4 November, and it is a rare thing: a cinephile magazine taking AI microdramas seriously enough to pick them apart.

It is an essay rather than reporting. There are no interviews and no new data, and the evidence for its central complaint is Kesh's own viewing. She says she watched "a frankly concerning number of AI microdramas" and kept meeting "almost identical scenes with little more than a change of setting." Her argument runs in four steps: microdramas work like a slot machine, AI turns them into a faster content machine, AI casting keeps landing on the same faces, and the platforms keep the power over who gets seen. We checked the claims she links to, and most of them hold. Two of her numbers don't, and her best point is one the Chinese market has already started acting on.

What the essay gets right

The reported claims in the piece mostly trace back to named outlets. The Guardian reported on Sunday 25 October that RoseBerry Media is recutting the Australian soap Neighbours into AI-assisted microdramas built around Margot Robbie's old storylines as Donna Freedman, using a tool called RedSnapper. Her former co-star Matthew Werkmeister told the paper he objects to his own image being reused this way: "Not from any money-making point of view, but from my image being used through AI and a platform I did not consent to."

The Hollywood Reporter found in September that Los Angeles microdrama work, a lifeline for many actors after the 2023 strikes, has become scarce as apps such as DramaBox, DramaWave and FlareFlow add synthetic actors. Candace Mizga, an actor and co-founder of Chera TV, told THR that people who had worked in the space full time "are going back to their day jobs, going home, and that's cast, crew, producers."

The China figures come from the regulator. Wang Xiaoliang of the National Radio and Television Administration said in September that about 430,000 microdramas were made in the first eight months of 2026 and "more than 90% of them were produced using AI technologies," with 68,000 titles removed this year. Business Insider did report that some estimate US microdrama startups are directing up to 75% of their marketing budgets to Meta, though that figure is an estimate and covers marketing, not the whole spend. Her strongest outside evidence on formula also holds: a 2025 study in Open Research Europe had a small text model write 11,800 stories set in 236 countries, and it produced essentially one plot. That test used a text model, not a video pipeline, but the pattern is the one she describes.

Where her numbers slip

The weakest part of the essay is the data behind the slot machine argument, and nearly all of it comes from one vendor. Kesh writes that 93% of viewers were still watching an hour after subscribing, citing HolyWater, the company behind the AI app My Muse and the live-action app My Drama. HolyWater's public release from July says something narrower: on My Muse's flagship title, nine in ten paying subscribers were still watching after 100 minutes of cumulative viewing. The full report sits behind a sign-up form and we have not read it, but no public text we found contains the 93% figure. Her figure of 13 hours a week for heavy viewers is real (13.1 hours, up from 9.5), but HolyWater measured it on My Drama, the live-action app, so it says nothing about AI series.

The claim that an AI series can be made in two weeks against 12 for live action is not in any public HolyWater text either. The release only says, in the words of co-CEO Anatolii Kasianov, that testing a new genre "used to take months, and on My Muse it takes weeks." The AI drama with 120 million TikTok views that Kesh mentions has no title or link, so nobody can check it.

The Kyle Chayka quote also needs its context back. Kesh quotes the New Yorker writer saying that "what AI produces is fundamentally kitsch." He did say it, in a February 2024 Dazed interview about his book Filterworld, a conversation about algorithmic culture rather than microdramas. The essay leaves out his next sentence: "That's totally different from an artist actively using AI tools in their process, but I worry more about the consumer." In a piece about the people making these shows, that distinction matters.

The face problem already has a price in China

Kesh's best line is about casting. AI microdramas, she writes, favor "a limited gene pool of doe-eyed women and vaguely Mediterranean-looking men," and she sums the problem up as "a technology capable of conjuring almost anyone, determined to reproduce the same handful of faces." She has no measurement for Western apps, only her own viewing. In China, the same complaint has already moved from critics to the platforms.

Hongguo, ByteDance's short-drama app, published a notice on Wednesday 22 July on how AI drama characters should be created. It asks that a series' main characters be clearly distinguishable, tells creators to stop reusing highly similar looks across series, and announced a clean-up of what it called high-frequency AI faces. Xinhua reported on Monday 12 October that discomfort with the "AI face" keeps trending on Chinese social platforms, and it put a number on the glut: more than 220,000 new AI microdramas went up on Douyin in the first half of 2026, and only about 0.47% passed 100 million plays. Its explanation was blunt: viewers have to watch for a while to know whether a plot is formulaic, but a face is judged at first glance: the same face again, so they swipe away (our translation).

The fix the market is reaching for is real people. For the horizontal AI series Baoqing Matou, the Hunan studio Qianzhen Huanjie scanned 37 locals from Shaoyang over three days, on a budget of more than 2 million yuan with nearly 1 million of it spent on compute. Face-licensing platforms such as ActID in Shenzhen and New Claw in Chengdu now pay ordinary people roughly 99 to 500 yuan per episode for their likeness, with the use and term stated up front. Even Shortical, the Tel Aviv-based app Kesh criticizes, now carries a License Your Face page in its site navigation.

Who keeps the audience

Kesh's other strong argument is about distribution, and it is harder for creators to engineer around. Shortical says on its creator page that its top ten creators earned more than $2.4 million last quarter on a 70% revenue share of paid unlocks, and that it onboards about 50 creators a month. Those are the company's own numbers. Kesh gives the opportunity its due in one generous sentence: "For a generation facing increasingly precarious creative careers and questioning whether their work will ever matter, the possibility of making a living this way is hardly insignificant."

Then she lands the point: "The tools might be available to everyone, but the audience certainly isn't." The apps decide which series get marketed, and if the Business Insider estimate is anywhere near right, much of that marketing is bought on Meta. HolyWater's own survey cuts the other way on taste, with 89% of My Muse users saying they want more AI stories and paid conversion level with live action (23.6% against 23.9%), but those are vendor numbers from a company that sells the format.

What it means for AI creators

The critique and the market data end up pointing in the same direction. A cheaper pipeline is no longer an edge when every studio has one, so what separates a series people finish from one they swipe past is the work Kesh says is missing.

• Character consistency now has two meanings. Keeping a face stable across episodes is the baseline, and keeping it distinct from every other series on the app is the newer test, one Hongguo has written into its rules.
• Licensed real faces are becoming a budget line. Chinese studios pay locals and face-licensing marketplaces per episode, and Western apps are starting to build the same pipeline, so consent paperwork is now part of the workflow.
• Formula is the first weakness a critic notices. Kesh's warthog-and-pregnancy plot reads as a joke because nothing in it is motivated, and fixing that is a writing job, not a rendering job.
• Distribution decides the payout, because a revenue share only pays on series an app chooses to market, so makers who bring their own audience have the stronger hand in any platform deal.

Our take: Kesh is right that sameness is the format's weak spot, and wrong to treat it as permanent. China's platforms already reward distinct faces and clean up copies, and the makers who cast like a casting director will be the ones still worth watching.

Timeline

• Date: Tue 20 Feb 2024. Event: Kyle Chayka's Filterworld interview, the source of the "fundamentally kitsch" line; Source: Dazed
• Date: Tue 29 Jul 2025. Event: Open Research Europe paper: 11,800 model-written stories, essentially one plot; Source: Open Research Europe
• Date: Thu 11 Jun 2026. Event: Blog citing Omdia's $14B global microdrama projection for 2026; Source: MinionArts
• Date: Wed 22 Jul 2026. Event: Hongguo issues its notice on AI drama characters and high-frequency AI faces; Source: Beijing News, 23 Jul
• Date: Thu 23 Jul 2026. Event: HolyWater and Owl & Co release The State of Microdrama 2026; Source: HolyWater
• Date: Thu 10 Sep 2026. Event: THR reports LA microdrama jobs drying up as apps add synthetic actors; Source: The Hollywood Reporter
• Date: Sun 20 Sep 2026. Event: NRTA official: about 430,000 microdramas in eight months, over 90% made with AI; Source: China.org.cn
• Date: Thu 8 Oct 2026. Event: Business Insider: up to 75% of microdrama marketing budgets go to Meta (estimate); Source: Business Insider
• Date: Mon 12 Oct 2026. Event: Xinhua on AI face fatigue, licensed faces and the 0.47% hit rate on Douyin; Source: Xinhua
• Date: Sun 25 Oct 2026. Event: Guardian reports the Neighbours AI recut around Margot Robbie's storylines; Source: The Guardian
• Date: Wed 4 Nov 2026. Event: Little White Lies publishes TikTok Cinema: Inside AI Microdramas; Source: Little White Lies

Sources and what is unconfirmed: Little White Lies (Yalda Kesh), 4 Nov 2026; HolyWater Tech release, 23 Jul 2026; Shortical creator page, read 5 Nov 2026; GamesBeat, 21 Oct 2026; The Guardian, 25 Oct 2026; The Hollywood Reporter, 10 Sep 2026; Business Insider, 8 Oct 2026; China.org.cn, 20 Sep 2026; Xinhua, 12 Oct 2026; Beijing News, 23 Jul 2026; Open Research Europe, 29 Jul 2025; Dazed, 20 Feb 2024; MinionArts, 11 Jun 2026. Unconfirmed: HolyWater's 93% and two-weeks figures (the full report is gated and we did not read it), the unnamed 120-million-view TikTok series, and the essay's claim that marketing can run to ten times the production budget (the paper it cites was not readable). Shortical's earnings figures are self-reported. We did not seek comment from Little White Lies, the author, HolyWater or Shortical.

Your Voice

Quoted from public posts on X and credited by handle. If one of these is yours and you want it removed, message @the__generated on X and it comes down.

Got something to say? Tag @the__generated on X