China's Double Ninth AI Film Got 320 Million Views. The Science in the Pitch Doesn't Hold Up.

On October 9, the day China marks the Double Ninth Festival — the annual occasion for honoring the elderly — a short AI-generated film called With You (《和你在一起》) went out. It runs under ninety seconds. According to the platforms promoting it, it pulled 320 million plays in a single day, put the hashtag for AI videos about the festival into Douyin's top three, and accumulated 1.87 billion reads on Weibo. Those are the numbers the campaign published; I have not seen them independently audited, and single-day play counts on Chinese short-video platforms are notoriously generous.

Set the numbers aside. The format is the interesting thing, because it's about to become a product category.

What the film does

The construction is simple and effective. You supply three inputs: a faded old photograph, a fragment of a childhood recording, and one sentence spoken in a family dialect. The system produces paired imagery — on one side, a small child tying a red scarf around a grandmother's neck; on the other, the same person at twenty-eight, kneeling to tie the same scarf around a thinner, shaking neck. A father releasing a bicycle seat; the grown child steadying a rehabilitation walker.

The label attached to this is "time-twin imagery." It is not a deepfake in the sense of impersonation. It's closer to illustrated memoir: a real photograph as the seed, and generated frames around it to represent moments that were never photographed.

The demographic logic is not subtle

China had 312 million people aged 60 or over as of the third quarter of 2026, about 22.3 percent of the population. That is roughly one in five. Meanwhile, a state-backed digital-literacy push has put "silver classrooms" in 276 prefecture-level cities, and the campaign's own material claims 41 million younger users have been teaching parents to spot scam links, enlarge fonts, and record medication reminders by voice.

That's the actual market. Not people who want AI art. People in their thirties and forties, with parents who are aging, a phone full of photographs that were never taken, and a specific kind of guilt about distance.

Every Chinese New Year produces a wave of content aimed at this guilt. What changed this year is that the content can now be personalized. A brand film makes you feel something generic. A generated film starring your grandmother's actual face makes you feel something you have to sit down for.

Where the pitch gets sloppy

Here's the part that bothers me, and it's worth flagging because this exact paragraph is going to be copied into a hundred product decks over the next year.

The promotional copy argues that the technology is restoring not images but "the original code of love," and supports this with a scientific claim: that the hippocampus secretes a special neuropeptide during strong familial bonding, and that this biological imprint is three to five times more durable than verbal memory. Therefore, even when Alzheimer's blurs "who I am," it cannot erase "holding this hand feels safe."

I went looking for the study. I couldn't find one. "Special neuropeptide," "three to five times more durable," and the claim that motor memory of affection survives Alzheimer's — these are not findings I can trace to a paper. They read like science-shaped writing, which is a different thing.

The real neuroscience is more modest and more interesting: emotional memories do tend to be consolidated more strongly than neutral ones, and procedural memory (how to do something) is often preserved later than episodic memory (what happened). But "your grandmother will recognize the feeling of your hand" is a hope, not a mechanism, and selling it as a mechanism is where care turns into marketing.

The ethics arrive faster than the rules

Three questions are going to be asked of every company in this space, and none of them have good answers yet.

Consent. A photograph of a grandparent is usually supplied by someone else. The subject may be alive and unable to understand what a generative model is, or dead and unable to object. Neither situation has a settled norm. Chinese law protects likeness and portrait rights, and Chinese courts have been expanding "identifiability" as the test rather than literal copying — which is exactly the standard a generated family film would meet.

Accuracy. If the system invents a scene your mother never lived, and you show it at a memorial, what have you made? Not a lie exactly. A composite. But the emotional weight of a photograph comes partly from the belief that it records something. Move that belief and you change what the object is, even if the pixels look the same.

Grief. The hardest version of this product is not "here is a nice video of your family." It's "here is a video of your father saying he's proud of you," generated after he died. That product exists in several markets already. The clinical literature on persistent grief suggests that immersive continued bonds can help some people and prolong distress in others, and there is no screening step in a checkout flow.

What a good version looks like

I don't think this category is inherently exploitative. There is a legitimate, valuable product here, and it looks like this:

It tells you what it invented. Every generated frame carries a marker, and the output includes a plain statement of what came from your photograph and what came from the model. It asks the living subject for consent, in language they can actually evaluate. It refuses to generate speech in the voice of someone who has died. It prices per artifact rather than per subscription, because the point is one film you keep, not a feed you scroll.

The version that will actually ship at scale, I suspect, will do none of these things, because they all cost conversion.

The business that's forming around this

Independent of that one campaign, a service category has assembled itself in about six months. It looks like this: photo restoration and colorization as the entry product, at a few yuan per image; a "memory film" tier with music, voiceover, and generated connective scenes; and a premium tier that includes a short loop of the subject moving, which is where the pricing gets uncomfortable and the legal exposure starts.

The distribution channel is not the app store. It's the family group chat. One person in a family buys the film, posts it, and three cousins ask which service was used. That's a referral loop with essentially zero acquisition cost, which is why the category can survive on low ticket sizes — and why it will scale faster than any review process can keep up with.

There's also a quieter B2B version: funeral homes, memorial halls, and cemeteries in several Chinese provinces have started offering digital memorial packages alongside the traditional services. That's the version I'd watch most closely, because it arrives at the moment of maximum vulnerability, from a counterparty the family already trusts, with no time to read terms.

The thing that's genuinely new

For most of the history of photography, families had a fixed supply of images of their elders: whatever was taken. The constraint was technical and it shaped memory — you remember the version of your grandmother that was photographed, in the clothes she wore that day.

That constraint is gone. The supply is now unlimited and generated on demand. Which means the question stops being "do we have a picture of her" and becomes "which version of her do we choose to keep."

That's a bigger change than a festival campaign, and it arrived, as these things do, disguised as a nice video with a sad piano track.