Don't rush to scale up AI-generated content

Don't rush to scale up AI content.

Recently, I saw a discussion about Hongguo Short Dramas. What's truly scarce for content platforms is verifiable audience attention; generation capability is just the barrier to entry. The industry says AI can bring 530 billion in value-add in the future, reducing jobs by one-third or even half. Sounds impressive. But Hongguo has already disposed of over ten thousand low-quality AI dramas. This is more factual than claims like "90% of content is made by AI." Once the barrier drops, garbage floods out first, while masterpieces become rare.

I've done small visual generation evaluations with students, referencing the ImageNet paper's approach of defining a validation set first: fixed prompts, model versions, and random seeds. The result? Same sentence, different model or seed, and the visual quality varies significantly. Content production is the same. If platforms only ask "can we make it" and not "why is it worth watching," it ends up being involution-style mass production.

This logic applies to film and TV, and also to AI content governance. I suggest platforms don't treat AI as universal capacity. First, make the generation process auditable: record model versions, asset sources, human reviews, and retained data. Only with traceability can you filter for quality.