In 2026 the buzzword shifted from AIGC to AI super-creators—one person or a tiny team owning planning, scripting, visuals, generation, and post, with a stable style. MCNs chase matrices; Douyin and Bilibili compete with contests, cash, and compute. Xiaohongshu (RED) took a tighter product path.

A 36Kr reprint of “新剧观察” describes the dual track: fold strong AI work into mid/long video, short film, and next-gen creator programs—while aggressively policing AI-hosted accounts, fake personas, and batch low-quality notes. The teardown is how incentives and anti-fraud share one distribution rule—and what RED Skill turns a “creator” into commercially.

For builders, that matters more than prize size: the platform drew a public line with enforcement numbers—AI can shoot for you; it cannot be your persona.

Mechanism: encourage creation, limit stand-ins

The support side is not an AI-only contest. March 2026’s first RED Next-Gen Creation Contest: tens of billions of event traffic and ¥5M in funds; 50 selected creators get ¥100k each plus 12 months of boost. Entry bar: 2+ minute mid/long video or 1,000+ character longform; judging favors sincerity, viewpoint, persona, resonance. AI can enter short-film tracks; batch homogenized AI is explicitly opposed. Tech may enter; assembly-line mindset stays out.

From May, mid/long-video ecosystem funding scales: year-round tens of millions RMB, up to ~¥1M per project; promising newcomers can get 1:1 ops and support framed up to a ~¥1M “annual salary” equivalent; plus a ~50B-traffic video incentive with monthly traffic coupons up to ~80k. Ranking weights favorites, follows, and watch time; strong mid/long pieces can stay in recommendation ~90 days—coverage cites ~5× exposure share in half a year and >~30% YoY plays; Yicai is quoted at nearly 100M daily mid/long-video viewers. Not AI-specific policy—but friendly to AI films that need world-building. Short model flexes struggle to earn that long window.

Governance numbers are sharper—and staged. Around April 27 the platform published a full AI governance stance (Sina Finance / The Paper): encourage honest AI labeling; oppose hosted ops, fraud, IP theft, and low-quality homogenization; gradient actions include throttling, warnings, bans. Contemporary counts already cited 1M+ AI abuse cases YTD, ~800k+ hosted accounts, ~150k fake notes. By May 2026, the “新剧观察” cumulative figures read ~180,000 notes suspected of AI fraud and 1.2M+ AI-hosted accounts cleaned—the same line, still lengthening. Coverage citing Science and Technology Daily frames the test as authenticity and original value—not “whether AI was used.” One line: traffic for AI creators, not for AI assembly lines.

The dual track rewrites the creator production function. “Many accounts × one template × high frequency” used to be a growth play; homogenization fingerprints are now risk. Xiaohongshu did not brand AI as a must-chase category; it stuffed AI work into existing quality systems—mid/long video, short film, next-gen—and measured it with the same “human taste” ruler. Public materials also describe auto-detection of AI / deepfake traces plus account-history interception—governance as a distribution filter, not a slogan.

Growth and commerce: from account matrices to Skill trade

On-platform exemplars share a pattern: AI is the implement; taste, viewpoint, and persona stay human.

Example (as reported)	Form	Where “human taste” sits
思花	Visual poems; coverage cites 500k+ followers	Designer aesthetic + original text; little model flex
史蒂芬霍京	Low-pixel “霍京市” world	Updatable space/characters, not one-shot spectacle
山音	Workplace dark comedy	Drama-lit background + regional voice; brand deals via style path
Spongecake	Light female short drama	Growth/identity themes; coverage cites 100k+ likes
自然养生道	Knowledge personification	23s clip ~54k likes / 32k saves; followers ~3k→22k in two weeks

Newrank mid-year notes (as cited) put it cleanly: AI is not the theme; it is a new visual language. Douyin skews conflict spectacle; Bilibili skews tech-community play; Xiaohongshu screens for visual taste, female expression, emotion, and personal style—aligned with a trust-first commerce premise.

MCNs (Malt, Bee Swarm, OST, Moment Culture, and others) already staff teams and sign matrices—Bee Swarm’s AI matrix ~60 accounts with plans toward ~100; Malt roughly 60/40 incubate vs sign, with a ~20-person “director + editor” squad around Spring Festival. Yet hot-topic accounts often lack brand spend, classic IP hits copyright walls, and templated women’s content risks backlash—cheaper production ≠ easier competition. Drama, vlog, and virtual variety travel better precisely because they resist full batch generation.

For MCNs, Xiaohongshu turns “copyable templates” from assets into liabilities: bigger matrices and faster generation look more like homogenization and hosting. The rational shift is writer/director/visual services + rights/legal + biz development—not another prompt warehouse. Many top super-creators prefer biz deals over full lock-in—when one person already owns production, what is scarce is clients, negotiation, IP, and distribution.

In July 2026, Xiaohongshu launched RED Skill, soft-launched Vibe Coding, and upgraded Builder Hub: developers mount Skills or mini-tools inside notes so viewing becomes using. Public figures: 160,000+ active AI developers; RED Skill’s first ~month produced 7,300+ original Skills; related topic exposure ~600M+. WAIC-linked messaging in the same piece: future inventory may be tools and solutions, not only ads—content → trust → demand → transaction as one chain. For product people, supply expands from “distributable notes” to “tryable capabilities.”

Put governance and Skill on one product map: governance compresses “fake person / assembly line” supply; Skill opens monetization for real creators/developers. Less noise, denser trade per unit of supply—more like a platform product choice than a prize campaign.

Where intuition fails

Intuition 1: platforms racing on AI must run the biggest contest.
Xiaohongshu nested AI inside mid/long-video quality systems and drew lines with enforcement—not an “AI” traffic lottery.

Intuition 2: MCN edge is template replication.
Under these rules, mature templates are dangerous; non-copyable taste and persona are the asset. Malt’s “traffic without brand spend” tests already show: efficiency does not buy trust premium.

Intuition 3: AI content monetizes only as ads.
RED Skill turns notes into a tool shelf; high-intent, style-led paths may fit Xiaohongshu’s user mix better than pure scale—creators like 山音 report brand deals landing more easily from in-app style discovery.

Who should do what

Role	Math to run	Don’t mythologize
AI creators / super-creator candidates	Whether style/persona still reads after you strip the model name; whether work survives a ~90-day recommend window; labeling and originality boundaries	Trading 100 homogeneous notes/day for traffic
MCNs	People as assets vs matrices as liability; biz deals vs full lock-in; rights that can sell long; whether directing craft beats a prompt warehouse	Porting Douyin-scale playbooks wholesale
Platform-tool / Skill builders	In-note loop: content → trial → trade; governance fingerprints and disclosure; distribution inside a 160k-developer pool	Shipping generators while ignoring hosting detection and IP

Xiaohongshu’s hardest AI-content product choice is not prize size. It is answering—with enforcement numbers and a Skill surface—which AI creation deserves to be seen and sold. Content can be batch-made; creators who cannot be copied are the supply the platform wants to keep.

Sources

• Xiaohongshu doesn’t want to mass-produce “AI super-creators” (36Kr / 新剧观察)
• Xiaohongshu’s first public AI governance stance (Sina Finance)
• When AI mass-makes “fake people,” Xiaohongshu stops staying silent (The Paper)