Jensen Huang dancing like Michael Jackson: when a hit video needs no truth

In early October a genre of AI edits flooded Chinese platforms: users replaced the performer and the audience at a Michael Jackson concert with various famous people. The most widely shared one put Nvidia CEO Jensen Huang on stage dancing, titled “Jensen Huang world tour”, about 86 seconds long, and it travelled overseas.

Musk joined in with a quip to the effect of “found the reason memory prices are going up”.

Why this one is worth writing about

If the story were only “AI video looks convincing”, that would be news from three years ago. What is different here is the shape of the spread.

It is not a carefully produced piece. It is a template. The MJ concert is a ready-made stage, the person is the variable, and anyone can fill it in. Once the template exists, output speed stops being limited by creative ability.

What makes convincing fakes dangerous is not generation quality. It is generation speed.

Two technical lines of defence

The industry is betting on two approaches, each with a clear boundary.

One: provenance. A camera or generation tool signs the file, recording who produced the content, on what device, at what time. Its strength is cryptographic — signatures cannot be forged. Its boundary is coverage: only content from signing devices carries it, and the vast majority of content online does not.

Two: implicit watermarking. Embed a mark during generation that is invisible to the eye but readable by a detector. Its strength is that it needs no cooperation from the capture side. Its boundary is that it only works on generators willing to cooperate — a malicious generator will not embed a watermark, and may actively strip someone else’s.

Together, the two cover “mistaken identity in good-faith scenarios”. They do not cover malicious forgery.

A small consequence for a personal site

My site has no video, so the practical consequence for me is about images.

The illustration here is the original asset the user supplied, and the OG image is generated from it at build time. In a provenance world, “assets I own” are the cleanest kind — the source is me.

The scenario to actually watch for is the reverse: do not pass off AI-generated images as real photographs. That is not a legal issue, it is a trust issue. The most expensive asset a personal site has is “I believe what is written here”, and one exposed staged image is enough to destroy it.

My take

I am not worried that AI video looks too real. I am worried that default trust gets spent.

Once “is this video real” becomes a question you must ask before every video, the cost is not borne by the forgers. It is borne by everyone publishing honestly, who must add a provenance step just to be believed.

That cost is stupid, but there is no better option yet. What you can do: leave verifiable traces on your own content, and check for traces on other people’s before you believe it.