AI dramas went prime time in China this year, and the volume question behind them is settled. 《桃花潭记》, a short drama made entirely with generative tools, began airing on Anhui Satellite TV in 20 episodes of roughly 10 minutes, with AI production and AIGC director labels in its opening credits [1]. On July 23, iQIYI carried 《奇谭:纸刃渡荒墟》, the first AIGC web feature to obtain the national web-drama distribution license, a feature running over 60 minutes [1]. Supply followed the milestones. Platforms released 367,000 micro-dramas in the first half of 2026, more than 1,500 AI titles arriving per day, which now make up over 74 percent of new supply [2]. A People's Daily report in late October put the eight-month total at 430,000 titles, with AI dramas above 90 percent [3].

For anyone building with generative video or audio tools, shipping models mass-produce supply, and they do not yet mass-produce polish. The remaining distance sits in three specific places, each with a measurable state as of November 2026. Long-form coherence is a preprint problem. Character consistency is a community-documented weakness that industrial pipelines claim to have locked down. Audio sync is where vendors are still shipping. Price your own pipeline against coherence, character consistency and audio sync before trusting any cinematic-grade claim.

What shipping models hold

Generation is the cheap part of the stack. Mainstream video models hold for roughly 5 to 10 seconds before characters deform and scenes jump, a failure mode the September coverage records for Sora and Kling alike [4]. A drama assembly line has to work around that ceiling.

The production data shows where its hours go. The team behind the hit series 《被裁掉的女孩》 spends two to three days per episode on average. Scenes, costumes, makeup and visual tone take most of that time, and generating the video itself takes less of it [2]. Its director said in an interview, "Knowing how to use AI does not mean knowing how to make content" (translated) [2]. Projects that once needed dozens or hundreds of people now run with core teams of a few [1]. One unnamed industry estimate cited in the same reporting puts AI short-drama production cost more than 90 percent below live-action shooting, with the fastest title finished in a week [2].

Whether that cheapness holds up shows in the retry numbers. A June 2026 compilation of more than 30 Reddit threads, mostly from 2025, gives the user side. Users reported that Runway replaced its unlimited plan with a credit system, with $95 per month buying roughly 19 fifteen-second generations at about $5 each. The same threads report a 90 percent miss rate at roughly $1.50 per wasted 5-second clip, and 10 to 20 attempts per keeper [5]. Local rendering is no shortcut, with one user reporting 15 minutes to produce a 5-second 720p clip on an RTX 4090 [5]. Between a still image and a usable clip the same threads count 5 to 10 steps and 3 to 4 tools [5]. The commodity is the seconds-long clip; the production cost lives in the attempts around it.

Google's coherence research arrives as preprints

The long-form coherence problem now has a research answer, published as three preprints. In September 2026 Google Research released a blog titled Automating coherent long-form video generation [6], and the underlying arXiv papers carry researcher self-reported benchmark numbers.

Co-Director, submitted April 27, 2026, formalizes video storytelling as a global optimization problem. A multi-armed bandit picks promising creative directions while a local multimodal self-refinement loop counters identity drift, and its GenAD-Bench evaluates 400 fictional-product scenarios for personalized advertising [7]. A²RD, submitted May 7, 2026, runs a Retrieve, Synthesize, Refine and Update cycle over a multimodal video memory, and across benchmarks spanning one- to ten-minute videos it reports gains of up to 30 percent in consistency and 20 percent in narrative coherence over state-of-the-art baselines [8]. CANVAS, submitted April 15, 2026, plans continuity explicitly through character continuity, persistent background anchors and location-aware scene planning, reporting benchmark improvements of 21.6 percent in background continuity, 9.6 percent in character consistency and 7.6 percent in props consistency [9].

Those percentages are benchmark comparisons on storyboards, advertising scenarios and one-to-ten-minute long-video benchmarks. A drama production line has yet to ship a system that carries them. A production-side practitioner quoted in the September coverage draws the same line. Demo environments and real production are different things, and complex lighting, micro-expressions and spatial logic still need a human backstop (translated) [4]. The marker that would change this picture is a product integration, and none has shipped.

Consistency and voices stay the human-carried axes

Character consistency is the axis users break first. The Reddit compilation quotes one tester calling it "the most broken part of AI video production" in 2025, with Kling the closest in that testing and Sora making it "basically impossible" [5]. The same threads judge the current workarounds incomplete. Runway's reference image is unstable, Kling Elements is the closest answer yet "not perfect", and ComfyUI plus LoRA is viable but complex enough to deter most users [5]. An aggregated Chinese industry review from May 2026 describes the other tier. Industrial platforms combining IP-Adapter methods and character asset libraries claim consistency above 95 percent, while stability for individual, non-industrial work remains a question mark [10]. Two contexts, two dated answers.

Viewers see the difference. Mainstream reporting records complaints about characters growing a third hand mid-scene and comment sections full of "why does everyone have the same face" [2]. A screenwriter interviewed there says current AI dramas lack "the feel of a living person" (translated) [2].

The aggregated review calls dubbing the decisive retention factor, with some industry views holding it above 50 percent of the viewing experience, and names audio-visual sync as the failure point that produces a "reading aloud from a textbook" flatness [10]. On the vendor side, ElevenLabs confirms Dubbing v2 is available through ElevenAPI [11] and promotes its v4 model as its most emotionally expressive yet, a vendor self-claim [11]. A secondary industry report describes Dubbing v2 upgrades in naturalness, timbre fidelity, lip-sync precision and long-content stability, making the point that a 90-minute film whose dubbing drifts at minute 80 loses everything before it [12]. Regulators have also flagged unauthorized face fusion and the cloning of voice actors' voices as infringement cases [3].

The audience still spends its attention on live action

Outcome data now exists at scale. Douyin added 221,900 AI drama and comic-drama titles in the first half of 2026 with an industry hit rate of 0.47 percent, over 90 percent of the output seeing weak traffic [2]. The aggregated review puts mainstream AI comic-drama completion below 15 percent and reports total playback for live-action drama at 25 times that of AI drama, on live-action output of roughly 5 percent of supply; those figures carry the weight of a single aggregated community piece [10].

Investment still points into AI production. DataEye projects the 2026 AI drama and comic-drama market above 40 billion yuan (RMB), up 138 percent year over year, an analyst projection [2]. Platforms are paying revenue-share incentives for productions that combine real performers with AI, and the Micro-drama Development Management Measures took effect in September with classification filing, AI content labeling and algorithm governance [3]. Regulators have separately ruled, at a symposium, that face and voice fusion require rights-holder authorization [3]. Enforcement this year has removed 68,000 noncompliant titles and handled more than 1,200 accounts [3].

The division of labor practitioners describe matches the tool evidence. Basic storyboards and plain dubbing can shift to AI, while aesthetic judgment and emotional expression stay with people [2], and the genres where AI's cost advantage bites hardest are science fiction and mythology, which need large worldbuilding and heavy effects [2].

Current tools carry a drama's supply side, its effects-heavy genres and a cost floor that took live-action pricing apart. Polish still depends on human work at three points, assembling coherence beyond seconds-long clips, keeping characters continuous outside industrial pipelines, and delivering emotional dubbing with true audio-visual sync. On the takeover question, the November evidence sets a measurable boundary and stops short of a verdict. AI holds above 90 percent of new supply and a 0.47 percent hit rate, while live action holds roughly 25 times the playback per the aggregated review [10]. What would move that boundary would show up in two places, the first product integration of the coherence research and the next half-year of audience outcome data.