They're already doing this
Doctored videos of Nancy Pelosi slowed down to make it sound like she's drunken slurring their words are making the rounds tonight. You don't need fancy AI tricks able to work off a single photo if you have actual video of people to mess with.
The downside of this trend is that even if you get a video of someone admitting to / committing a crime, they will be able to cry "fake video!" and it will be difficult if not impossible to prove whether or not it is legitimate. It used to be that audio and video evidence was the most convincing proof you could come up with in a trial, but before long it will be seen as on the level of something bearing a signature - easily forged and therefore easily dismissed as not being "proof beyond a reasonable doubt".
So the question is, could someone theoretically come up with a way to work around this? Maybe have video recording devices use a private key to encode their identity into the recording, so if you could present the phone that made the recording into evidence it could be proven that the video shown is as was originally recorded by the device and not modified?