What I do is minimize the remote video and move it to the corner of the screen, as close to the camera as possible. If the other person does the same, we are then close to making eye contact while gazing at each other's images.
Of course this would be avoidable through careful implementation: only redirect the gaze if it's already directed at the screen, or even specifically at the correct part of the screen. This might require a calibration step, or at the very least, it would require the software to be aware of the relative configuration between the camera and the screen.
Still, personally I'm not really comfortable with any software systematically editing the direction of my gaze.