In early August, a 34-second AI video called “Chinese Heaven” went viral: white jade colonnades, cloud-sea palaces, robes drifting through fog. No plot, no dialogue — just 5 million plays in four days, 48,000 likes and 9,700 shares across overseas platforms. The same week, “aesthetic sense is the best tool in the AI era” hit the top of Weibo’s hot list.
The real signal is not the video. It is what the video’s ecosystem is doing to business. The loudest reaction to AI-generated aesthetics has come not from critics but from the brands paying for content. In 2026, Unilever said it will no longer use AI-generated faces or skin to promote its products, and Dove said it will not use AI to alter women’s images. L’Oréal banned AI-generated template faces and bodies. Aerie’s “no AI, no retouching, 100% real” post drew more than 44,000 likes. LEGO quietly keeps refusing AI.
This is not Luddism. It is a new line item. The same companies are raising marketing budgets, not cutting them: Unilever’s marketing investment rose from 13.1% of revenue in 2022 to 15.9%, and it plans to push social-advertising from 30% to 50% of total ad spend (from $3.1 billion to $5.1 billion). L’Oréal spent €14.2 billion on advertising and promotion in 2025 — 32.2% of its €44 billion in sales. What they are actually paying for now is quality control on taste.
Why the Slop Happened: The Cost Collapse
The root cause is arithmetic. A commercial short video used to cost hundreds or thousands of yuan and take more than ten days. AI now produces one for a few yuan in thirty minutes — a cost cut of more than 90%. When marginal cost approaches zero, supply becomes infinite, and infinite supply without a filter is slop. Merriam-Webster made “slop” its 2025 Word of the Year: AI-generated content with no soul, produced at scale. And the flood is now machine-driven: for the first time, bots outnumber humans online, which means most of what is generated is not even seen by a human before it spreads.
The flood has measurable costs. Stanford HAI puts AI error rates at 26–31% in home-furnishing, maternal-infant and outdoor categories — exactly the categories where visual trust drives purchase. Gartner finds 53% of consumers lack basic trust in AI-generated content. McKinsey reports 80% of enterprises have seen no substantial positive profit impact from AI. The Journal of Business Research shows AI-created marketing material actually lowers positive word-of-mouth and purchase intent.
The result is a growing list of brand backlashes in 2026 alone — Aupres, LAN, Gucci, babysheep, Colgate, Supor — all caught running gimmicky AI ads that damaged trust rather than built it. When a 26–31% error rate lands on a living room or a baby product, the “cheap content” stops being cheap.
The Moat Moved: From “Can You Generate” to “What You Choose”
Here is the structural shift. For three years, competitive advantage in generative AI lived in the generator: faster, bigger, higher resolution. That advantage is now collapsing. The gap between Midjourney, DALL-E and Firefly has narrowed to the point where output quality alone no longer differentiates. Stanford’s 2026 AI Index puts Claude Opus 4.6 at an Elo of 1503, with China’s top model just 2.7% behind. When everyone can generate anything, the moat moves to the selection layer — taste. “Generate well” is table stakes; “choose well” is the moat.
You can see the selection layer in the flood of imitations. “Chinese Heaven” was replicated within days across Douyin, Weibo, Bilibili and Kuaishou — every clone using the same template of ancient architecture, atmospheric light and cultural symbols. One widely shared analysis noted that the AI does not understand the culture behind the celestial palace; it has learned pixel-level statistical patterns of “clouds + palace + gold tiles.” A style any model can copy in hours is not a moat. A taste a model cannot copy — judgment about what should exist — is.
Taste Is Being Industrialized
Because taste is now economically valuable, it is being turned into infrastructure and IP.
The infrastructure: Peking University’s “ZhiJing” project converts abstract Chinese aesthetic categories — artistic conception, spirit, vitality — into measurable indicators that AI can be tested and iterated against. Shanghai AI Lab and the China Academy of Art built “ArtiMuse,” a professional aesthetic-understanding model. Overseas, Design Arena raised $7.9 million to train taste into models, using human evaluations from 5.3 million users to make outputs feel less mechanical.
The IP: In June, Volcengine licensed Stephen Chow’s classic films — King of Comedy, God of Cookery, and CJ7 — for AI video creation. Disney and OpenAI signed a three-year deal giving Sora access to more than 200 Marvel and Pixar characters and scenes for fan videos. Decades of validated aesthetic systems are now being scaled lawfully through AI, which means the most proven taste systems become licensable assets rather than copyable styles.
The Survivors and the Paradox
The clearest proof is Midjourney. Against Big Tech pressure, it holds 26.8% global market share and roughly $500 million in annualized revenue. And in July 2026 it acquired Co-Star — an astrology app — specifically for its aesthetic and tonality. The message: Midjourney is no longer selling image generation; it is selling taste as a platform.
Meitu tells the same story in China. A company often dismissed as a legacy tool grew total revenue 28.8% to 3.86 billion yuan in 2025, with imaging and design products up 41.6% to 2.95 billion yuan — 76.6% of the whole. Paid subscribers hit a record 18.44 million by mid-2026. A decade of accumulated aesthetic understanding turned out to be the asset that survived.
But here is the paradox: we hail aesthetics as the last moat of the AI era while simultaneously using the most efficient tools to fill that moat in ourselves. The same AI that makes taste valuable also makes it replicable. The moat only survives if the taste stays proprietary — your own data, your own judgment, your own curated selection — rather than something a model can absorb in an afternoon. This is why provenance is becoming the trust layer of the whole system: as AI content becomes impossible to verify by eye, text watermarking is turning into the verification standard.
What to Do Now
- If you are a brand: budget for aesthetic quality control as a separate line item. Unilever and L’Oréal are not abandoning AI; they are refusing cheap AI. The winning move is human-in-the-loop taste: use models for draft volume, keep humans for final selection.
- If you are a creator: stop competing on style, which is now copyable. Compete on judgment and narrative. The viral winners of 2026 — human warmth, intangible heritage, nature-healing content — win on curation, not generation.
- If you are building a product: treat taste as the moat. Feed your own proprietary taste data, build evaluation systems like ZhiJing or ArtiMuse, and remember that the moment your “aesthetic” can be replicated by any model, it stops being a moat.
The AI slop era is not the end of taste. It is the proof that taste — proprietary, curated, judgment-based — is the only thing left that scales without being copied. In an era where everything can be generated, the filter is the product.
