It's one thing to use existing footage of someone (e.g. the president of the U.S.) talking and combine it with a different audio stream to create the impression

It's one thing to use existing footage of someone (e.g. the president of the U.S.) talking and combine it with a different audio stream to create the impression that that person said something which they in fact didn't. However, an approach based on generating images with neural networks can go far beyond that. In theory, you could generate footage of arbitrary actions performed by any person - the president beating up a child, the president shooting a gun into a group of protesting homosexuals, the president in an X-rated home movie, you name it. There are a lot of things special effect teams could do in old Hollywood movies, but newer movies have the potential to do the same kind of effects much more convincingly. Likewise, newer technology like this one will become better and better at creating fake footage. Another research direction that is related is replacing faces in videos with that of other actors, or morphing parts of a real video to match a second input source, both discussed previously on HN a couple of times.