There's also always the "analog loophole". Display the AI-generated video on a sufficiently high-resolution / color gamut display and record it on whatever device has convenient specs for making the recording, then do some light post-processing to fix moire/color/geometry. This would likely be detectable, but could shift the burden of (dis-)proof to the defendant, who might not have the money for the expert witnesses required to properly argue the technical merits of their case.

More likely, the signing would have to use compression-resistant steganography, otherwise it's pretty easy to just remux/re-encode the video to strip the metadata.