Hey paede, I think it's working correctly too. I don't think you'll find any other AHI that behaves any better.
The reason its moving is because the plane _is_ moving. The AHI is overreacting because the plane is pointing straight up. When you're pointing straight up, even a few degrees of movement can have you pointing from north to south. I used to work on magnetic tracking systems in aerospace, and the problem is sometimes called 'gimbal lock.' The math represents the 3d orientation read from the accelerometers as Euler ( Oiler) angles, Azimuth, Elevation and Roll. When the plane is pointing up, your elevation reaches 90 degrees, and now your azimuth degree of freedom and your roll degree of freedom are nearly parallel. So any tiny change in orientation shows up as a change in roll in the AHI.
Even if the math in the AHI is done with quaternions, which represent orientation with four numbers, which are a hypercomplex number... you don't have any problems with math around the elevation approaches 90 orientation... the visual representation still just extracts the roll...
The day that I see a good AHI, that really follows the horizon as the horizon is represented in the camera (it would have to be configurable.) and you don't see the horizon in the OSD when you don't see it in the camera... then I'll buy one. I've only seen one OSD in a FPV ground recording ever that wouldn't just be a maddening distraction for me. (Of course my opinion will change next time I'm in fog... and I'll be thrilled to have any cue to keep me upright!)