In this case, they're basically using a neural network to approximate a really tricky high dimensional function from a lot of measurements from a scene, and use it to interpolate values.
Think of it as "fancy (non)linear regression" or something like that.
It's quite clever.
I'm now wanting to set up my magic bullet rig to try this!
Just thinking about consumer equipment, being able to shutter release and fire a laser "burst" precisely enough would be a challenge. If the shutter release is wired, does the time for the signal to travel down the wire + the mechanics of the shutter need to be compensated for the time of flight of the "photon"? I could see this being one of those YouTube channels with someone doing this in their garage.
All of that to say that the accuracy of what they've done is impressive
It's quite realistic, here's the same thing without AI:
https://www.youtube.com/watch?v=EtsXgODHMWk&t=107s
I still share your concern, however, particularly because they seem to be avoid moving the camera without time moving as well. I was expecting bullet-time!
Your suspicion is warranted, but it really depends on what "AI" is being used (I'd rather call it ML. As a ML researcher myself, and who publicly criticizes LLMs[0]).
The reasoning for this is that in essence, ML is curve fitting data from high "polynomial" functions (approximately accurate). But there are many things like density estimators which are very good in statistical settings where you cannot access the density function directly (called "intractable") and so all you can deal with is samples (e.g. you can sample examples of human faces, but we have no mathematical equation to describe all variations and in what likelihood). This is not too different from Monte Carlo Sampling and is often used in variational inference. When you are doing density estimation you can have a lot more confidence in your results as you can actually do things like building proper confidence intervals and you can test likelihood (how well does your model explain the data).
So yeah, keep the skepticism up. There's a lot of snake-oil in ML and these days it is probably good to default to that position. Especially since a lot of ML people are not well versed in math and there's a growing sentiment of not needing math (you'll even find that common around here. It is a reliance upon empirical results and not understanding "Elephant fitting"). FWIW here they're using NeRF and it looks like they are using it to tune parameters of their physical model. I'd have to take a deeper look but at a quick glance I'd let down my guard a bit.
[0] Worth noting that "AI" used to be the typical signal that some thing was snake oil. Now everything is called AI. I'll leave it to the reader to determine if this is still a strong signal or not.